DYNAMICS OF CARBON STORAGE IN THE WOODY BIOMASS OF NORTHERN FORESTS

Transcription

1 BOSTON UNIVERSITY GRADUATE SCHOOL OF ARTS AND SCIENCES Dissertation DYNAMICS OF CARBON STORAGE IN THE WOODY BIOMASS OF NORTHERN FORESTS by JIARUI DONG B.S., Beijing Institute of Meteorology, P. R. China, 1989 M.A., Chinese Academy of Sciences, P.R. China, 1992 Ph.D., Chinese Academy of Sciences, P. R. China, 1997 Submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy 2002

2

3 Acknowledgements I am deeply grateful to my major advisor, Dr. Ranga B. Myneni, who gave me the opportunity to pursue higher education in the Department of Geography, Boston University. His scientific rigor and rich knowledge have enriched my scientific pursuits. His optimism, selflessness and scientific supervision have fostered my enthusiasm for the research presented here. I would like to thank Dr. Robert Kaufmann for his active guidance, great help and valuable input to my dissertation research. I express my special thanks to Dr. Compton Tucker, for I benefited a lot from his encouragement. His insightful thoughts broadened my vision in the field of physical climate and global carbon research. Many thanks to Dr. Guido Salvucci, who guided me intensively during my first years at BU. It is due to his involvement I have a good basis in soil hydrology. Many thanks to Dr. Yuri Knyazikhin, whose expert knowledge of mathematics, physics of remote sensing, and generous personality set a good example for me. I am honored to have them as the members of my dissertation committee. I feel very fortunate to have conducted my Ph. D. program at Geography Department of Boston University. I would like to thank Dr. Alan Strahler and Dr. Curtis Woodcock for sharing their knowledge on remote sensing in classes, and also for their excellent management of our Department that fostered a friendly environment. I would also like to thank all the faculty and staff of the Geography Department, whose teachings enriched my knowledge. I am also pleased to have spent time with so many fellow graduate iii

4 students, because through them I could understand the American society better. I particularly thank my office mates: Yujie Wang and Wolfgang Buermann, because our close cooperation lead us all to succeed in our Ph. D. programs. I would also like to thank the international work team for their contributions? they are from Austria, Finland, Russia, and the USA and include Pekka Kauppi, Jari Liski, Vladislav Alexeyev, and Malcolm Hughes. I would like to dedicate this dissertation to my wife and my son for their infinite love and support, and to my parents for their selfless dedications. iv

5 DYNAMICS OF CARBON STORAGE IN THE WOODY BIOMASS OF NORTHERN FORESTS (Order No. ) JIARUI DONG Boston University Graduate School of Arts and Sciences, 2002 Major Professor: Ranga B. Myneni, Associate Professor of Geography ABSTRACT Part of the puzzle of greenhouse gases and climate change is determining where carbon dioxide (CO 2 ) is absorbed, and what causes a region to become a "carbon sink". Analyses of atmospheric CO 2 concentration changes indicate a carbon sink of about 1 to 2 billion tons on land in the northerly regions. Elsewhere the land is suggested to be neutral, which implies that emissions of another 1.5 billion tons of carbon a year from cutting and burning of tropical forests are nearly balanced by sinks of similar magnitude there. The geographical detail of the land carbon sink has, however, remained elusive. Forest greenness observations from sensors on National Oceanic and Atmospheric Administration satellites were combined with wood volume data from forest inventories to produce relatively high resolution maps of carbon stocks in about 15 million square kilometers of northern forests, roughly above the 30th parallel. Comparison of carbon stock maps from the late 1990s and early 1980s identifies where forests were storing v

6 carbon and where they were losing carbon. Results indicate that about 61 billion tons of carbon is contained in the wood of these northern forests. Further, the analysis indicates that forests in Europe, Russia and America have been storing nearly 700 million metric tons of carbon a year, or about 12% of annual global carbon emissions from industrial activities, during the 1980s and 1990s. American forests absorbed 120 million tons of carbon a year, which is about 11% of the USA s annual emissions. With the exception of some Canadian boreal forests, which were found to be losing carbon, most northern forests were storing carbon. Russia, the country with the most forests, accounted for almost 40 percent of the biomass carbon sink. This study has important scientific, economic and policy implications. The scientific implication is that it deconstructs the mystery of the land carbon sink by providing geographically detailed maps of forest carbon pools, sources and sinks. The economic implication is that the wood volume maps provide valuable information to the forest industry. This study also has political relevance as a potential tool for monitoring carbon sequestration in the future. vi

7 Table of Contents Acknowledgements Abstract Table of Contents List of Figures List of Tables List of Abbreviations iii v vii xi xvii xix 1 INTRODUCTION Global Carbon Budget Climate Change and Forests Terrestrial Carbon Processes Temporal Scales Statement of the Problem 12 vii

8 2 DATA Satellite Greenness Evaluation of GIMMS NDVI The Effect of Orbital Drift Evaluation for the Period June 1991 to January Evaluation of GIMMS NDVI Data Quality Forest Inventory Data Introduction Data Sources and Quality Northern Temperate and Boreal Forests Evaluation of Biomass from Inventory Data of Wood Volume Matching NDVI and Biomass Data 36 Appendix: List of Provinces/States Used in This Research 38 3 METHODS Methods for Estimating Terrestrial Carbon Processes The Biomass-NDVI Equation 59 viii

9 3.3 Evaluation of the Biomass-NDVI Equation Representativity Stability The Effect of Latitude on the Relation Between NDVI and Biomass Temporal and Spatial Relations Between NDVI and Biomass 72 4 RESULTS and DISCUSSION Patterns of Woody Biomass Sinks and Pools Comparison of Remote Sensing and Inventory Estimates Bias Analysis of the Estimates Uncertainty of the Carbon Sink Estimate Comparison of Estimates for Canada, China, Russia and the USA Comparison of Pooled vs. Country-Specific Relations Comparison of Pooled vs. Forest Type-Specific Relations Reasons for the Observed Changes Limitations of this Research Significance of this Research Analysis of Emissions and Sinks of Industrialized Nations 99 ix

10 Appendix: Lists of provinces, states and countries in Figure 4.5 and CONCLUSIONS 127 Bibliography 132 x

11 List of Figures 1.1 Illustration of the global carbon cycle ( Illustration of terrestrial carbon cycle. Arrows indicate fluxes, and boxes indicate pools. CWD represents coarse woody debris; GPP, gross primary production; NPP, net primary production; NEP, net ecosystem production; NBP, net biome production; R a, autotrophic respiration; and R h, heterotrophic respiration by soil organisms. Emissions of carbon dioxide on the right side are due to respiratory losses, and on the left due to nonrespiratory losses (Schulze et al., 2000) Hypothetical seasonal patterns of NDVI (thick line) and temperature (thin line) Interannual correlation between NDVI (light color) and near-surface air temperature anomaly (black color) in Eurasia and North America (Zhou et al., 2001) Centennial components of temperature variations. The high frequency is the reconstructed temperature. The low frequency line, resulted from component analysis and has about a 120-year cycle period (Shabalova and Weber, 1999) Zonal distribution of terrestrial and oceanic carbon fluxes. These were deduced from eight inverse models using different techniques and sets of xi

12 deduced from eight inverse models using different techniques and sets of atmospheric observations after accounting for fossil-fuel emissions. Positive numbers are fluxes to the atmosphere. Results are shown for the 1980s (plan bars) and for (hatched bars) (Heimann, 2001) Growing season NDVI totals (hatched area). Growing season NDVI total is determined by its two dimensions: growing season duration and the magnitude of the observations Time series of spatial averaged NDVI anomaly for different latitudinal bands. The vertical solid line denotes the time of Mt. Pinatubo eruption. GIMMS denotes results are from the original GIMMS NDVI data, and BU-GIMMS denotes results from NDVI data after corrections of satellite orbital drift and Mt. Pinatubo eruption and climate-predicted NDVI for the June 1991 to January 1995 period Time series of spatial averaged NDVI for different latitudinal bands. The vertical solid line denotes the time of Mt. Pinatubo eruption. GIMMS denotes results are from the original GIMMS NDVI data, and BU-GIMMS denotes results from NDVI data after corrections of satellite orbital drift and Mt. Pinatubo eruption and climate-predicted NDVI for the June 1991 to January 1995 period Distribution of provincial and state forest area in the six countries (Canada, USA, Russia, Sweden, Finland, and Norway). a) Number 1 on x-axis xii

13 USA, Russia, Sweden, Finland, and Norway). a) Number 1 on x-axis represents forest area less than 1 million ha, and 2 represents forest area between 1 and 2 millions hectare, and so on. b) The first bar is forest area less than 0.25 million ha, and the second bar is forest area between 0.25 and 0.5 million ha, and so on. The last bar represents forest area above 5 million ha a) Distribution of forest area by genus in the USA, Canada, and Russia. b) Distribution of forest area by genus in Sweden, Norway and Finland Age structure of forests in Canada, Russia and the USA Administrative map of Sweden with 24 provinces for which the inventory data are available. Sweden spans a latitude range of about 55 o N to 70 o N Land cover map of Sweden with 12 landcover types at a spatial resolution of 1x1 km (Hansen et al., 2000). Forests are defined as broad leaf and needle leaf forests, mixed forests and woody savannas Comparison of forest and land area between remote sensing estimates and inventory reports Plots of total woody biomass versus cumulative growing season NDVI for individual provinces in (a) Sweden and (b) Russia Zonal distribution of biome types in Eurasia and North America Ill-conditioning in the tracer inversion problem (Enting, 2000). 75 xiii

14 3.2 Plot of total (A) and above-stump (B) woody biomass versus cumulative growing season NDVI. The NDVI data are five year averages prior to the date of inventories. Outlier 1 is British Columbia (CAN) and outliers 2 are data from Washington, Oregon and (northern) California (USA). These represent 16% of North American forest area The relations (a) between total biomass and latitude at three levels of total growing season NDVI, and (b) between total biomass and total growing season NDVI at three latitudes (a) The relation between NDVI and biomass for samples defined by NDVI. (b) The stability of the relation between NDVI and biomass for samples defined by NDV Difference in growing season NDVI totals between two time periods, and , for all vegetated regions Map of forest fraction defined as the fraction of each quarter degree pixel area under forest land covers. Forests include broad leaf forests, needle leaf forests, mixed forests and woody savannas, land covers in Hansen et al. (2000) classification Spatial detail of changes in woody biomass carbon pool of northern temperate and boreal forests between late 1990s and early 1980s. Biomass estimates were converted to carbon by multiplying by 0.5, a standard factor xiv

15 for converting woody biomass to carbon (TBFRA-2000, 2000) Spatial detail of pool size in the northern temperate and boreal forests during late 1990s. Biomass estimates were converted to carbon by multiplying by 0.5, a standard factor for converting woody biomass to carbon (TBFRA-2000, 2000) Comparison of remote sensing and inventory estimates of the biomass carbon pool. We show estimates at the provincial, state and national level, rather than on per unit forest area basis, to include uncertainties associated with differences in respective estimates of forest area Comparison of remote sensing and inventory estimates of the rate of biomass carbon pool change. We show estimates at the provincial, state and national level, rather than on per unit forest area basis, to include uncertainties associated with differences in respective estimates of forest area Comparison of the pooled estimates (dash lines) and country-specific estimates (solid lines) for six countries (Canada, USA, Russia, Sweden, Norway, and Finland). The values of latitude are set to 35 o N for the USA, 55 o N for Canada, and 65 o N for the other countries Comparison of the pooled estimates and forest type-specific estimates at (a) the latitude of 40 o N, and (b) the latitude of 70 o N Detailed map of changes in the carbon pool for Canada. 120 xv

16 4.10 Detailed map of changes in the carbon pool for the USA Detailed map of changes in the carbon pool for Russia Detailed map of changes in the carbon pool for European countries Detailed map of changes in the carbon pool for China and Japan Ranking of Annex 1 countries by average 1980s and 1990s sink size. #Australia, Iceland, Ireland, Luxembourg, New Zealand are not included Cumulative global CO 2 emissions from fossil-fuel combustion and cement production. Half of the total 270 Gt C released is since Ranking of Annex 1 countries by their ratios of carbon sinks to emissions. #Australia, Iceland, Ireland, Luxembourg, New Zealand are not included. *Annual means emissions are from 1992 to The others are from 1982 to Ranking of Annex 1 countries by their ratios of carbon sinks to emissions per capita. #Australia, Iceland, Ireland, Luxembourg, New Zealand are not included. *Annual means emissions are from 1992 to The others are from 1982 to xvi

17 List of Tables 1.1 Contemporary carbon budget for the 1980s and 1990s (Schimel et al., 2001) Launch date, and ascending (northbound Equator crossing) and descending (southbound Equator crossing) node times in Local Solar Time for NOAA series of satellites Regression results between NDVI and climate in different latitudinal bands for the six biome types during July 1981 to May 1991 and February 1995 to December Regression results for NDVI-Climate equation in different latitudinal ands for the six biome types during July 1986 to May 1991 and February 1995 to December Regression results for the Biomass-NDVI equation (3.1) Regression results for the total biomass equation (3.1) with data from an individual nation and period Remote sensing estimates of biomass carbon pool (1995?99) and sinks in xvii

18 temperate and boreal forests of North America and Eurasia Remote sensing estimates of carbon pools (1995?99) and sinks in the above stump biomass of temperate and boreal forests in North America and Eurasia Country estimates of carbon pools and sinks in the woody biomass of temperate and boreal forests Monte Carlo simulation results to evaluate the change in carbon storage that is generated by the uncertainty in biomass-ndvi equation Comparison of estimates for Canada, China, Russia, and the USA Regression results for the total biomass equation (3.1) with data from individual country and for Sweden with two time periods Growing season NDVI total range and latitude for the six countries Comparison of carbon storage estimated by the pooled relation and the country-specific relations for the six countries Regression results for the total biomass equations (3.1) with data from different biome types (Needle forests, broad leaf forests, and total forests). 112 xviii

19 List of Abbreviations AVHRR BATS CWD DUM ENSO FAO FIA FOWL FPAR GIMMS GPP Gt C ha IPCC KPR LAI LSM LST LULUCF Mha Advanced Very High Resolution Radiometer Biosphere-Atmosphere Transfer Scheme Coarse Woody Debris Dummy Variable El Niño-Southern Oscillation Food and Agriculture Organization Forest Inventory and Analysis Forest and Other Wooded Land Fraction of Photosynthetically Active Radiation Absorbed by Vegetation Global Inventory Monitoring and Modeling System Gross Primary Production giga tons of carbon hectare Intergovernmental Panel on Climate Change Kyoto Protocol Report Leaf Area Index Land Surface Model Local Solar Time Land-Use, Land-Use Change and Forestry Million hectare xix

20 MISR MODIS NASA NBP NCAR NDVI NEP NOAA NPP ppmv SZA TBFRA UN-ECE UNFCCC USDA Multi-angle Imaging SpectroRadiometer Moderate Resolution Imaging Spectroradiometer National Aeronautics and Space Administration Net Biome Production National Center of Atmospheric Research Normalized Difference Vegetation Index Net Ecosystem Production National Oceanic and Atmospheric Administration Net Primary Production parts per million by volume Solar Zenith Angle Temperate and Boreal Forest Resources Assessment United Nation-Economic Commission for Europe United Nations Framework Convension on Climate Change United States Department of Agriculture xx

21 1 Chapter 1 INTRODUCTION 1.1 Global Carbon Budget The global annual mean temperature of the earth s surface has increased by 0.3 to 0.6 o C over the past 100 years according to the Intergovernmental Panel on Climate Change (IPCC, 2001). The largest increases in temperature are observed on land in the northern high latitudes, with lower increases nearer the tropics and in regions with a strong oceanic influence. It is now generally accepted that these changes are primarily due to rising atmospheric concentrations of greenhouse gases and sulphate aerosols. The increase in atmospheric concentrations of carbon dioxide (CO 2 ) since the industrial revolution is due to human activities? combustion of fossil fuels, cement manufacture and tropical deforestation. The CO 2 concentration in the atmosphere has increased from approximately 290 parts per million by volume (ppmv) before the industrial revolution to greater than 370 ppmv today. It is predicted that atmospheric CO 2 levels will continue to rise and could exceed 500 ppmv during the 21st century (IPCC, 2001). Knowledge of the carbon cycle is a pre-requisite for predicting and controlling future atmospheric CO 2 concentrations.

22 2 Carbon on Earth is stored in three big reservoirs? atmosphere, land and ocean (Figure 1.1). Vegetation contains about 600 giga (10 9 ) tons of carbon (Gt C), comparable to the amount in the atmosphere (750 Gt C), and approximately equal to 60% of the carbon in the surface ocean layers (1000 Gt). The deeper layers contain significant amounts of carbon, about 38,000 Gt, but they can affect the atmospheric CO 2 concentrations only through the surface layers. Carbon stored in the soils is about 1600 Gt? twice as large as in the vegetation. The carbon stored in the soil and litter of forest ecosystems also makes up a significant proportion of the total carbon pool. Globally, soil carbon represents more than half of the stock of carbon in forests. More than 80% of the carbon in the boreal ecosystems is stored in the form of soil organic matter, whereas in the tropical forests the carbon is fairly equally distributed between the vegetation and soil (Dixon et al., 1994). At high latitudes, i.e., in cooler climates, soil organic matter accumulates because it is produced faster than can be decomposed. At low latitudes, however, warmer temperatures encourage rapid decomposition of soil organic matter and subsequent recycling of nutrients. Terrestrial ecosystems play a significant role in the global carbon cycle. An estimated 125 Gt of carbon are exchanged annually between the land and the atmosphere, accounting for two-fifths of the total exchange of carbon between the earth and the atmosphere (IPCC, 2000). A proportion of the carbon dioxide emitted to the atmosphere by fossil-fuel combustion, cement manufacture and deforestation is taken up by the

23 3 oceans and the land. A recent IPCC assessment updated the global budgets, shown here in Table 1.1 (Schimel et al., 2001). During the 1980s, carbon emissions totaled 5.4? 0.3 Gt C/yr from fossil-fuel burning and cement manufacture, and 1.7 (0.6 to 2.5) Gt C/yr from land-use changes. The net carbon flux into the oceans is estimated to be 1.9? 0.5 Gt C/yr, and 0.2? 0.7 Gt C/yr into the land. Because the atmospheric carbon increase is observed to be 3.3? 0.1 Gt C/yr, there is still a 1.7 Gt C missing sink per year. For the 1990s, the estimates are somewhat similar, except for a larger land carbon sink. The fossil-fuel emissions are 6.3? 0.4 Gt C /yr. Many studies, including Bousquet et al. (2000), suggest 1 to 2 Gt of carbon sequestered in pools on land in temperate and boreal regions. Such sinks represent 16 to 37% of annual carbon emissions from industrial activities. The use of carbon sinks in policies for reducing greenhouse gas emissions is presently being debated (IPCC, 2000). Thus, characterizing the location and mechanism of carbon sinks is of scientific and political importance. 1.2 Climate Change and Forests The world has about 3,870 million hectares of forests, accounting for 30 percent of the world land area (TBFRA-2000, 2000). The mutual influence of climate and forests is important in studies of global warming. I begin my discussion with how forests are impacted by changes in climate.

24 4 Climate influence forest growth through two key variables: temperature and precipitation. Any region with increased temperature and unchanged or reduced precipitation will experience significant reductions in soil moisture, which will constrain plant growth and increase the likelihood of fires. Forests response to long-term climate changes depends on the ability of the species to live in the new environment or change their geographic distributions. Climatic zones in the northern high latitudes are predicted to shift northwards by as much as 5 km per year (IPCC, 2001). Therefore, boreal forests will make gains in the northerly margins, but may be replaced by temperate vegetation at their southerly boundary. Temperate forests will be most affected by climate warming at higher latitudes and by changes in rainfall at lower latitudes. Similar to boreal forests, temperate forests also shift towards the north. Tropical forests response to climate changes depends mostly on changes in the rainfall regime. Where rainfall is reduced and temperatures increase, reduced soil moisture is expected to be the most significant threat. Changes in forest cover could feedback on the climate through modification to surface temperatures, precipitation and by influencing atmospheric CO 2 concentrations. Forests have a lower albedo than other ecosystems and, through their extensive root systems, have more access to soil water than other types of vegetation. Thus, they absorb more solar energy, which can either lead to more surface heating or cooling through changes in evapotranspiration rates. In tropical zones, evaporation processes tend to dominate and the net effect of forests is to cool and moisten the atmosphere. In the north,

25 5 sensible heating effects dominate, thereby leading to local warming (Buermann et al., 2001). 1.3 Terrestrial Carbon Processes The terrestrial carbon cycle is a highly dynamic system that includes several storage pools, such as vegetation, soil, detritus, black carbon residue from fires, harvested products, etc., that can be characterized by their turnover time (Schulze et al., 2000). Carbohydrate pools turn over on a daily basis, leaves can store carbon for several seasons, and carbon in living wood and soil pools may remain there for hundreds of years and millennia. Fire may return carbon to the atmosphere instantaneously and can produce long-lived black carbon. Any activity that changes the amount of biomass in vegetation and soil has potential to sequester carbon from, or release carbon to, the atmosphere. The terrestrial carbon cycle can be classified into the following fluxes (Figure 1.2): gross primary production (GPP), net primary production (NPP), net ecosystem production (NEP), and net biome production (NBP). About one-third of the total amount of CO 2 in the atmosphere enters into green leaves every year (Farquhar et al., 1993, Ciais et al., 1997). The amount that is fixed from the atmosphere, i.e., converted from CO 2 to carbohydrates during photosynthesis, is called GPP, which is carbon assimilation by photosynthesis ignoring photorespiration. Terrestrial GPP has been estimated to be 120 Gt C/yr based on 18 O measurements of atmospheric CO 2 (Ciais et al., 1997).

26 6 Annual plant growth is the difference between photosynthesis and autotrophic respiration (R a ), and is referred to as net primary production (NPP). NPP is the fraction of GPP resulting in plant growth, and can be measured through sequential harvesting or by measuring plant biomass, provided turnover of all components (e.g., fine roots) is included (Hall et al., 1993). Global terrestrial NPP has been estimated to be 60 Gt C/yr, that is, about half of GPP is incorporated in new plant tissue. The other half is returned to the atmospheric as CO 2 by autotrophic respiration, that is, respiration by plant tissues. All the carbon fixed as NPP may be returned to the atmospheric CO 2 pool through two processes: heterotrophic respiration (R h ) by decomposers (bacteria and fungi feeding on dead tissue) and herbivores; and combustion in natural or man-made fires. Most dead biomass enters the detritus and soil organic matter pools where it is respired at a rate that depends on the chemical composition of the dead tissue and on environmental conditions. Detritus and microbial biomass have a short turnover time (< 10 years). In contrast, soil organic carbon has decadal to centennial turnover times, because inert soil organic carbon is composed of molecules more or less resistant to further decomposition. Net ecosystem production, NEP, is the difference between NPP and heterotrophic respiration (R h ), which determines the amount of carbon lost or gained by the ecosystem without disturbances, such as harvests and fire. NEP can be estimated from measurements of CO 2 fluxes over patches of land. Global NEP is estimated at about 10

27 7 Gt C/yr. Of these, annual NEP fluxes are in the range 0.7 to 5.9 tons C/ha/yr for tropical forests, 0.8 to 7.0 tons C/ha/yr for temperate forests, and up to 2.5 tons C/ha/yr for boreal forests (Valentini et al., 2000). NEP is the most fundamental carbon flux for a natural forest ecosystem. Net biome production, NBP, is the carbon accumulated by the terrestrial biosphere when carbon losses from nonrespiratory processes are taken into account, including fires, harvests/removals, erosion and export of dissolved organic carbon by rivers to the oceans (Schulze and Heimann, 1998). NBP is estimated to have averaged 0.2? 0.7 Gt C/yr during the 1980s and 1.4? 0.7 Gt C/yr during the 1990s (IPCC, 2001). This study is limited to analysis of the carbon pool in the woody biomass of temperate and boreal forests of the northern hemisphere, roughly above the 30th parallel. 1.4 Temporal Scales Many aspects of the earth climate system vary on timescales ranging from hours to millennia. Natural variability, at interannual, decadal and longer timescales, has substantial influence on the global system. The influencing factors include changes in the atmosphere, air-sea interactions (e.g., El Niño-Southern Oscillation, ENSO), changes on land and in the oceans, solar variability, etc. Natural variability of the biosphere is an important factor influencing climate variations at longer time scales. The normalized

28 8 difference vegetation index (NDVI), an index of vegetation greenness, measured from satellite sensors is useful in characterizing vegetation variations at seasonal, interannual, and possibly decadal scales (Myneni et al., 1997, Myneni et al., 1998, Tucker et al., 1986, Tucker et al., 1999). Satellite based NDVI observations based on red reflectance, where the chlorophyll absorbs most of energy, and near-infrared reflectance, where the vegetation reflects most of the energy, were recognized early on to represent vegetation greenness. Therefore, NDVI is a valuable tool for distinguishing vegetation types and for characterizing phenology (Tucker et al., 1985, Loveland et al., 1991). Seasonal signals of NDVI have been closely related to seasonal patterns of temperature variations (Figure 1.3), as temperature is the major control factor for growth of vegetation in northerly regions. Satellite NDVI has been used to monitor the duration of active growing season of vegetation in the north (Myneni et al., 1997). Longer growing seasons were reported, in the northern high latitudes, where notable warming has occurred in the spring during the 1980s and 1990s. Large changes in both the magnitude and duration of the seasonal cycle of NDVI are observed in northern latitudes. NDVI in Eurasia (North America) increased by 12.41% (8.44%) during the growing season from 1982 to 1999, and the length of an active growing season increased by 18? 4 days in Eurasia and 12? 5 days in North America, brought by an early spring and delayed autumn (Zhou, et al., 2001). Also, the NDVI variations are consistent with variations in the seasonal cycle of atmospheric CO 2,

29 9 thus suggesting changes in biological activity of the vegetation. Year-to-year changes in seasonally integrated NDVI reflect interannual changes in vegetation greenness. Therefore, many studies have used NDVI or NDVI-based products to calculate terrestrial plant NPP (Bonan, 1996, Foley et al., 1996, Haxeltine and Prentice, 1996, Potter and Klooster, 1997, Dickinson et al., 1998, Field et al., 1998). The Global Inventory Monitoring and Modeling System (GIMMS) developed NDVI data set from the Advanced Very High Resolution Radiometers (AVHRR) on board the NOAA-7, -9, -11 and 14 satellites for the period July 1981 to December 1999 (Tucker et al., 2001). NDVI data spanning the last two decades facilitate studies of vegetation change at interannual time scales. Interannual variations in NDVI can be related to effects of interannual climate changes, such as year-to-year temperature changes and ENSO events. A recent study reported that NDVI is positively correlated with near-surface air temperature anomaly at the 1% significance level during the growing season and in spring season, in the northerly regions (Figure 1.4; Zhou et al., 2001). The El Niño-Southern Oscillation (ENSO) is the predominant mode of interannual variability in all climate variables (Cane et al., 1986, Cole et al., 1993, Dai et al., 1997, Dai and Wigley, 2000). Tropical land ecosystems have experienced substantial interannual climate variability due to frequent warm ENSO episodes in the recent decades. ENSO events during 1982/83, 1986/87, 1991/92, and 1997/98 resulted in drier and warmer conditions in the Amazon Basin, northeastern Brazil, eastern Australia, etc.

30 10 The Amazonian ecosystems acted as a source of atmospheric CO 2 during these years, because both drier weather and warmer temperatures decreased NPP and increase heteorotrophic respiration according to simulations by the Terrestrial Ecosystem Model (Tian et al., 1998). Monthly NDVI-based products, such as leaf area index (LAI) and fraction of photosynthetically active radiation absorbed by vegetation (FPAR), were used to drive the terrestrial model linked to a climate model to assess interannual variation in exchanges of energy, water, and carbon dioxide between land and the atmosphere (Potter and Klooster, 1999, Bounoua, et al., 2000, Buermann et al., 2001). Currently, there are many ways to study century-scale climate variability using proxy data. Ice cores provide the record length necessary for examination of slower oscillations. Evidence of century-scale climate variability is also preserved in lake sediments. A coupled atmosphere-ocean model has also been used to explore century scale effects of gradual increases in atmospheric CO 2 concentrations to doubling and quadrupling endpoints (Manabe and Stouffer, 1993). Changes in forest ecosystem at century scale are known as succession, the gradual and continuous process by which one habitat is replaced by another involving both biological and non- biological components. The forest is subjected to damage from agents of change, both natural and human-influenced. The older the forest, the more vulnerable it is to these forces. Fire is a major force altering existing ecosystems. Naturally occurring fires are started by lightening. Some plant species actually require fire to stimulate their regeneration. Other disturbances include diseases, insect infestation and severe weather in the form of snow, ice, and strong winds.

31 11 These environmental disturbances also help the forest to make space for new species of plants and wildlife. Timber harvesting and logging, like forest fires, are disturbances that when over done can initiate succession. Figure 1.5 illustrates centennial scale variations in temperature data through the presence of a 120-year cycle oscillation (Shabalova and Weber, 1999). If a 1000 year record of NDVI is available, this would potentially show centennial scale variations due to vegetation successional phenomenon. Forest succession is a fundamental ecological process, including regeneration, growth and mortality of individual plants. The successional stage of a forest is directly related to the ecosystem carbon budget (Schulze, 1999, Caspersen et al., 2000). Tree ring width is another good proxy for climate, not only because the variability in annual growth is correlated with summer temperature and winter precipitation (Graumlich, 1991), but also tree ring records can document relatively stable oscillatory modes when longer time series data are available. But, tree ring data cannot cover spatial detail, and also the labor for obtaining high frequency data is expensive. Traditional insitu forest woody biomass estimates, mostly from inventories, do not account for continental and global scale spatial variability. Large-scale networks do not exist for measurement of forest biomass at an ideal high frequency and fine spatial resolution that is needed because in-situ measurements are expensive and tedious. It would be ideal if remote sensing data were to fill this research gap.

32 12 If NDVI data were available for 100 years, the inter-decadal NDVI variations would be expected to represent biomass changes. Year-to-year changes in woody biomass are quite small, about two orders of magnitude smaller than the biomass pool. At decadal or longer time scales, the biomass changes can be considerable because the difference between gains and losses can accrue and may be captured accurately by low-frequency variations in vegetation greenness. GIMMS NDVI is a global 15-day data set at 8-km spatial resolution, covering two decades from July 1981 to December In this research, the difference between seasonally integrated 5-year average NDVI of the first 5 years and the last 5 years will be used to investigate changes in biomass. 1.5 Statement of the Problem Part of the puzzle of greenhouse gases and climate change is determining where carbon dioxide (CO 2 ) is absorbed, and what causes a region to become a "carbon sink". The land and oceans are known to store half of the 6.5 billion tons of carbon emitted annually from fossil-fuel burning and industrial activities. Currently the global carbon budget cannot be balanced. The CO 2 emitted by fossil fuel burning, cement manufacture and tropical deforestation is apparently greater than the amount remaining in the atmosphere and removed by the known sinks (Table 1.1). The inability to balance the carbon budget will cause considerable uncertainty on predications of future atmospheric CO 2 concentrations, climate and the debate on carbon cycle manipulations (Dixon et al, 1994).

33 13 The terrestrial sinks of carbon continue to be problematic, although there are growing indications to explain the role of the terrestrial biosphere in the missing carbon sink. Several processes may contribute to the net land sink, including the increase in plant productivity by the rising atmospheric CO 2 concentration (e.g., Tans et al., 1990; Keeling et al., 1996, Battle et al., 2000), fertilization of ecosystems by airborne nitrogen pollutants (Holland et al, 1997), changes in land management (Houghton et al., 1999, Caspersen et al., 2000, Brown and Schroeder, 1999), etc. The relative proportion of these contributions remains uncertain. Figure 1.6 shows the zonal distribution of terrestrial and oceanic carbon fluxes (Heimann, 2001). These data were deduced from eight inverse models using different techniques and sets of atmospheric observations after accounting for fossil-fuel emissions. Results are shown for the 1980s (plain bars) and for (hatched bars). Positive numbers are fluxes to the atmosphere. This figure represents our current understanding, that is, about 1 to 2 billion tons of carbon are somehow sequestered in sinks on land north of 30 o N. Elsewhere, the land is neutral, where sources nearly match sinks. However, the geographic distribution of the northerly land sink remains unknown. Robust techniques are needed for mapping carbon stocks and fluxes with lower cost in view of the stated scientific and political tones of this theme.

34 14 The land carbon sink could be distributed among various carbon pools vegetation, soil, etc. This research program will focus on the vegetation pools in the north. Satellite observations of vegetation provide global coverage with relatively high spatial resolution over the last two decades. At decadal and longer time scales, biomass can change considerablely due to cumulative differences between annual gains and losses. As mentioned before these can potentially be observed as low frequency variations in climatological greenness, in much the same way that changes in greenness at century and longer time scales suggest successional changes. The proposed research will use remote sensing data to provide a spatial detail of the woody biomass carbon pools, sources and sinks in northern forests.

35 15 Table 1.1: Contemporary carbon budget for the 1980s and 1990s (Schimel et al., 2001). Emissions (fossil-fuel buring, cement manufacture) 1980s (Gt C/yr) 1990s (Gt C/yr) 5.4? ? 0.4 Atmospheric increase 3.3? ? 0.1 Ocean-atmosphere flux -1.9? ? 0.5 Land-atmosphere flux -0.2? ? 0.7 Emissions due to land-use change 1.7 (0.6 to 2.5) Assume 1.6? 0.8 Residual terrestrial sink -1.9 (-3.8 to 0.3) -2 to -4 Negative values denote flux from the atmosphere, that is ocean or land uptake.

36 Figure 1.1: Illustration of the global carbon cycle ( 16

37 17 NBP NEP NPP CO 2 CO 2 GPP CO 2 CO 2 CO 2 Growth R a new Biomass old R h (litter) R h R h harvest Wood products Mortality Litter CWD Soil Organic matter Fire Black carbon Figure 1.2: Illustration of terrestrial carbon cycle. Arrows indicate fluxes, and boxes indicate pools. CWD represents coarse woody debris; GPP, gross primary production; NPP, net primary production; NEP, net ecosystem production; NBP, net biome production; R a, autotrophic respiration; and R h, heterotrophic respiration by soil organisms. Emissions of carbon dioxide on the right side are due to respiratory losses, and on the left due to non-respiratory losses (Schulze et al., 2000).

38 NDVI NDVI temperature Temperature ( o C) Month -10 Figure 1.3: Hypothetical seasonal patterns of NDVI (thick line) and temperature (thin line). Figure 1.4: Interannual correlation between NDVI (light color) and near-surface air temperature anomaly (black color) in Eurasia and North America (Zhou et al., 2001).

39 Temperature Anomaly ( o C) Year Figure 1.5: Centennial components of temperature variations. The high frequency is the reconstructed temperature. The low frequency line, resulted from component analysis and has about a 120-year cycle period (Shabalova and Weber, 1999).

40 Figure 1.6: Zonal distribution of terrestrial and oceanic carbon fluxes. These were deduced from eight inverse models using different techniques and sets of atmospheric observations after accounting for fossil-fuel emissions. Positive numbers are fluxes to the atmosphere. Results are shown for the 1980s (plan bars) and for (hatched bars) (Heimann, 2001). 20

41 21 Chapter 2 DATA 2.1 Satellite Greenness The NOAA series satellites were designed to operate in a near-polar, sun-synchronous orbit. The orbital period is about 102 minutes which produces 14.1 orbits per day with a repeat cycle of approximately 14 days. The launch dates of satellites are given in Table 2.1; also given are the approximate times of the ascending node (northbound Equator crossing) and the descending node (southbound Equator crossing) in Local Solar Time (LST) at launch and on March 1995 for the active satellites. A global NDVI data set at 8 8 km resolution for the period July 1981 to December 1999 was developed from about 40,000 orbits of daily data from the Advanced Very High Resolution Radiometers (AVHRR) on board the NOAA-7, 9, 11 and 14 satellites by the Global Inventory Monitoring and Modeling System (GIMMS) group. This data set consists of five subsets: Africa, Australia, North America, South America and Eurasia. It contains channels 1 ( µm) and 2 ( µm) reflectances, channels 4 ( µm) and 5 ( µm) brightness temperatures, solar and view zenith angles, and the day of compositing. The normalized difference vegetation index (NDVI)

42 22 measures the contrast between red (channel 1) and near-infrared (channel 2) reflection of solar radiation and can be used to proxy green leaf area (Myneni et al., 1998). NDVI varies between -1 and +1, and ranges between 0.2 and 0.1 for snow, inland water bodies, deserts, and exposed soils. GIMMS NDVI increases from about 0.1 to 0.75 for increasing amounts of vegetation, but saturates in the case of dense leaf canopies, for example, humid tropical forests and old growth forests. Growing season NDVI can be characterized by two metrics: growing season duration and seasonal NDVI magnitude. As such, NDVI integrated over the growing season is an ideal measure of seasonal greenness in the north (Figure 2.1). Analyses of Pathfinder AVHRR NDVI data during (Myneni et al., 1998) and GIMMS NDVI data during (Zhou et al, 2001) indicate significant changes in both the magnitude and duration of the seasonal cycle of NDVI. NDVI in Eurasia increased by 12.4% during the growing season, and 8.44% in North America from 1982 to A longer active growing season brought by an early spring and delayed autumn is seen in Eurasia (18 ± 4 days), compared with North America (12 ± 5 days). The changes in NDVI are consistent with climate changes since the early 1970s. Statistical analyses show that there is a statistically meaningful relation between changes in NDVI and land surface temperature for vegetation areas of the northern latitudes (Zhou et al., 2001). Associated with warming at high latitudes is an approximate 10% reduction in annual snow cover between 1973 and 1992, especially an earlier

43 23 disappearance of snow in spring (Groisman et al., 1994). Where snow lines have retreated earlier due to enhanced warming, an earlier start of the active growing season is expected. NDVI-based products are surrogates of plant photosynthetic activity and have been used in various land surface models (Dickinson et al., 1998, Bounoua et al., 2000, Buermann et al., 2001). The changes in NDVI are expected to reflect changes in biological activity. A photosynthetically vigorous Eurasia relative to North America during the past two decades is consistent with the idea of a large terrestrial carbon sink and forest inventory analyses (eg, Houghton et al., 1999; TBFRA-2000, 2000). Therefore, NDVI magnitude and growing season duration representative variables of vegetation activity in the north. The processing of satellite data involved cloud screening and calibration for sensor degradation and inter-sensor variations (Rosborough et al., 1994; Los et al., 1994; Vermote and Kaufman, 1995; Los, 1998). Residual atmospheric effects were minimized by analyzing only the maximum NDVI value within each 15-day interval. These data generally correspond to observations from near-nadir viewing directions (Los et al., 1994) and clear atmospheric conditions (Holben, 1986). Correction of scattering effects by aerosols remains a challenge because of its influence on both visible and near infrared channels. Stratospheric aerosols associated with volcanic eruptions tend to have longitudinally homogeneous distributions within two months of injection and then slowly decrease with time (Los et al., 2000). The data from April 1982 to December 1984 and

44 24 from June 1991 to December 1993 were corrected to remove the effects of stratospheric aerosol loadings from El Chicon and Mount Pinatubo eruptions in mid and high latitudes of the Northern Hemisphere, respectively (Vermote and El Saleous, 1994). Corrections for tropospheric aerosol are not applied because the information on aerosol properties is insufficient. 2.2 Evaluation GIMMS NDVI The NDVI data used in this research are the recently developed third-generation data from the GIMMS group. Despite the corrections described above, GIMMS data still contain variations due to orbital drift and incomplete corrections for stratospheric aerosols. In this section, GIMMS NDVI data quality is assessed The Effect of Orbital Drift The GIMMS group derived the NDVI data set from the NOAA-7, -9, -11 and 14 satellites, because they all are afternoon satellites. However, the satellites drift in orbit over time, which causes a systematic change in the local time of observation. This is a major source of non-uniformity in multi-annual satellite time series. In the case of NOAA-9 and 11, the total duration from launch date to March 1995 is 123 and 78 months, respectively. Therefore, the drift rate for NOAA-9 is approximately 3 minutes and 23 seconds per month and for NOAA-11 is about 3 minutes per month. Thus, both

45 25 satellites show similar orbital drift rates. The time series of globally averaged anomalies of deseasonalized NDVI and solar zenith angle (SZA) at times of observations indicate that the satellite derived data are not stationary (Kaufmann et al., 2000). The data increase fairly steadily within the duration of each satellite and drop sharply between two-satellite shifts. We shall assume that NOAA-7, 9, 11 and 14 have similar orbital drift rate. A correction for orbital drift can be implemented as follows: (1) The period of study, 19 years, is divided into 4 periods (7/81 to 1/85; 2/85 to 10/88; 11/88 to 1/95; and 2/95 to 12/99). In each period, the NDVI data are from one satellite only (NOAA-7, 9, 11 or 14). (2) Evaluate NDVI anomaly, pixel-by-pixel, for all months during the first year of each satellite. The NDVI mean values are obtained by averaging the NDVI data of each pixel and month of year one over the 4 satellites. (3) Similarly, evaluate NDVI anomaly during subsequent satellite years by subtracting the satellite age-specific mean values. In the first 43 months, the NDVI average is evaluated with data from four satellites. In months 44 and 45 the mean value is from three satellites (NOAA-9, 11 and 14). Finally, from months 46 to 59 the average is obtained from two satellites (NOAA-11 and 14). (4) Create the NDVI data set which is corrected for orbital effects by adding the NDVI anomaly to the mean NDVI from year one of the four satellites.

46 Evaluation for the Period June 1991 to January 1995 About 16 months, October 1993 to January 1995, of NOAA-11 data require special treatment because of the long life-span of this satellite. Also, further attention is needed for NOAA-11 data following Mount Pinatubo eruption in mid 1991 with special emphasis on tropical and Southern Hemisphere sub-tropical. Following several previous investigations on the relation between climate and NDVI (e.g., Potter and Brooks, 1998, Los et al., 2001), we will develop correction techniques and/or methods for evaluating the quality of GIMMS NDVI data, by latitude band and by biome type. We will stratify the globe into six 30 o latitude bands to cover the high, mid, tropical and sub-tropical regions. As for the biomes, we utilize the Myneni et al. (1997) cover types grass and cereal crops, shrubs, broad leaf crops, savanna, broad leaf forests, and needle forests. I utilize monthly half-degree gridded climate data, air surface temperature and precipitation, generated for the period by interpolating directly from station observations (New et al., 2000). The spatial coverage extends over all land areas, including oceanic islands but excluding the Antarctic. The precipitation data for the period of are from the NASA Global Precipitation Climatology Project (GPCP) Version 2 combined precipitation data set (Huffman, 1997), and the temperature data are from NCEP/NCAR Reanalysis Project (NCEP, 2001). The GPCP precipitation data were obtained by a gridded analysis based on gauge measurements and satellite estimates of

47 27 rainfall. Both GPCP precipitation and NCEP temperature data are processed as monthly means with spatial 2.5 o latitude 2.5 o longitude global coverage. The relations between monthly climate and NDVI require matching climate data to remote sensing NDVI data in different latitude bands and biome types. The GIMMS NDVI data are at an 8 8 km spatial resolution in an irregular equal-angle grid. The climate data are 0.5 o resolution for the period and 2.5 o for the period We use a remote sensing land cover map at a spatial resolution 8 8 km (Myneni et al., 1997) to group climate and NDVI data for each of six biomes in every latitude band. The time series of climate and NDVI data from July 1981 to December 1999 are used to develop the NDVI-climate relations. The relation between NDVI and climate is estimated from time series data, within each latitude band and for each biome type, using the following specification: NDVI(t) = α ij + β ij Temp(t-lag) + γ ij Precip(t-lag), (2.1) in which NDVI(t), Temp(t-lag), and Precip(t-lag) are vegetation NDVI, monthly mean air temperature ( o C), and monthly precipitation (millimeter) averaged over space and biome type in months t and t-lag. The lag represents lag months of NDVI variations relative to climate variables. The regression coefficients, α ij, β ij, and γ ij, are specific to latitude band i for biome type j (Tables 2.2 and 2.3). The regression coefficients in Table 2.2 are

48 28 obtained from analysis of data of the entire period, but without the June 1991 to January 1995 data. The coefficients in Table 2.3 are generated from data for the periods July 1986 to May 1991 and February 1995 to December In the tropical and subtropical regions (30 o S-30 o N), precipitation data show strong seasonality over all biomes. Precipitation is high in areas under forests, annually greater than 240 mm/month in the summer and greater than 80 mm/month during winter. The precipitation in areas under other land covers is relatively less. For example, precipitation over shrub land ranges from almost zero in the winter to 150 mm/month during summer. The temperature data show weak seasonality in the forest regions and to some extents show seasonal variation under other land covers. The tropical evergreen forests especially do not show obvious signals of seasonal variation. The NDVI data in the tropical regions exhibit a lag of one to two months to climate. In the mid and high latitudes of the Northern Hemisphere, both NDVI and climate data show strong seasonality, as is to be expected, and there are minimal differences in temperature and precipitation variations by land cover. However, the NDVI data show large differences between biome types, even if they are controlled by the same climate. Therefore, the relation between NDVI and climate will be dependent on the biome type.

49 Evaluation of GIMMS NDVI Data Quality The constant intra-annual seasonality is removed in calculating the NDVI anomalies (Zhou et al., 2001). Data used to generate the NDVI anomalies are predicted from the regression equation (2.1) with the coefficients in Table 2.2, and the input climate data are described in section Thus, the NDVI anomalies for the period June 1991 to January 1995 are generated from the modeled NDVI by subtracting the NDVI mean values over this period (Figure 2.2). The modeled NDVI values for the period of study are produced by adding the above mentioned NDVI anomalies to NDVI means generated from the regression equation (2.1) using the coefficients in Table 2.3 (Figure 2.3). Figure 2.2 shows the time series of NDVI anomalies from original GIMMS data (thin lines) and after the corrections of orbital drift effects and volcanic eruptions (thick lines), with climate predicted NDVI for the June 1991 to January 1995 period. The corrected data show smooth transitions between change of satellites. The statistically significant relations between vegetation NDVI and climate attest to the quality of GIMMS NDVI data. Spatially averaged NDVI data clearly show the need for corrections in the tropics and the Southern Hemisphere (Figure 2.3). Generally, the corrections increase the NDVI magnitude over tropical regions, possibly indicating residual cloud cover effects in the original GIMMS data. The effect of Mt. Pinatubo eruption on data from the northern latitudes can be largely ignored, that is, the original GIMMS NDVI data have satisfactory corrections for the north. I will use the original GIMMS NDVI data

50 30 in our analysis of forest biomass in the north. Further details on the quality of the GIMMS NDVI data set can be found in Kaufmann et al. (2000) and Zhou et al. (2001). 2.3 Forest Inventory Data Introduction The world has about 3,870 million ha of forests, of which 95 percent are natural forests and 5 percent are forest plantations. The dynamics of forest biomass need to be better understood, both for studies of sustainable forestry and the global carbon cycle. Specifically, the distribution of biomass and the changes associated with different management scenarios have implications for the long-term sustainability of forest resources and for sequestration of carbon emissions. The Global Forest Resources Assessment 2000, a five-year effort, was a joint endeavor carried out by the FAO in cooperation with its member countries and many other partners (TBFAR-2000, 2000). The assessment was based on country information, remote sensing surveys, mapping of global forest cover and ecological zones, and the establishment of a forestry information system. National-level data on forest resources were collected through an exhaustive survey of inventory reports and other information from participating countries. Several major challenges had to be met in order to assemble country information. For example,

51 31 over half of the developing countries had only one forest inventory, and more than onefourth of them had never carried out an inventory (FAO, 2001). About 30 percent of the world s land area is under forest. The proportion of total land area under forest varies significantly by region and country. About half the land area of South America and Europe is covered by forest, but only one-sixth of Asia s land is forested. Africa and North America fall in between, each with about one-fourth of its land covered by forest. Two-thirds of the world s forests are located in only ten countries: Russia, Brazil, Canada, USA, China, Australia, Congo, Indonesia, Angola and Peru, in descending order. As for forest cover in terms of ecological zones, the largest proportion is contained in the tropical zone (47%), followed by the boreal (33%), temperate (11%) and subtropical (9%) zones Data Sources and Quality The forest assessment is mainly based on country information. However, natural climate and management practices vary significantly among countries, which present a particular difficulty for assessment. The inventory data for the USA is from the Resource Planning Act (RPA) Assessment Database, which includes mostly FIA (Forest Inventory and Analysis) data, and also data collected by National Forests in some areas not included in the FIA

52 32 program. The RPA reports forest area and wood volume on timberland by state, and class of timber (softwood or hardwood) in The Canadian inventory has been repeated on a 5-year cycle since 1976, but only the Canada s Forest Inventory 1991 version from year 1994 (CanF91-94) is available and has been compiled to a biomass inventory (Penner et al., 1997). The average biomass (tons/ha) reported by province for timber productive forests, will be used in our analysis. Data for Russia is primarily based on the 1990 statistical forest inventory on the timber stock of Russian forests (Alexeyev and Birdsey, 1998). Data are reported in the form of stem wood volume of growing stock in administrative territories of Russia and by species (conifer, deciduous hardwood, and deciduous softwood). Data for Sweden were published in a series of statistical forestry year-books (SYF, 1988 & 1999). We utilized data collected as growing stock on forest area by species from two time periods ( and ). Data for Norway and Finland are from the European Forest Institute data set based on the national forest inventories. They reported statistics of forest conditions and resources on productive forestland by vegetation types. However, the definitions of forest types, such as lichen forest and cowberry forest, are different between the countries. National forest resource inventories for most countries have detailed and reliable information on forest distribution and changes over 5-year periods (e.g., Birdsey and Heath, 1995, Turner et al., 1995, Lowe et al., 1996, Liski and Kauppi, 2000, and Fang et al., 2001). However, the inventory data are available for certain countries and regions only, and the quality of these data varies substantially among inventories. For example,

53 33 fewer data are available for remote regions of Canada, Russia, and elsewhere. Forests contain about 60% of the carbon stored in vegetation and about 50% of the carbon stored in soil (TBFRA-2000, 2000). Of these totals, a large percentage of the vegetation (41%) and soil (72%) pools are located between 25 o N and 75 o N (Dixon, et al., 1994). Consistent with these fractions, the proposed study will analyze the carbon stored in the woody biomass of temperate and boreal forests, which cover an area of about 1.4 to 1.5 billion hectares (Liski and Kauppi, 2000) Northern Temperate and Boreal Forests Inventory data for stem wood volume from 171 provinces in six countries (Canada, Finland, Norway, Russia, Sweden and USA) that cover over one billion ha of forest area are analyzed to estimate above-stump and total biomass. The total number of provinces and states in the six countries is 182. Of these, data from 171 provinces where forest area covers more than 15% of the land area (10% in RUS) are used in the following analysis. The list of these provinces is given in an appendix to this chapter. The distribution of provincial forest area in the six countries is shown in Figure 2.4. About 44% of the provinces have forested area less than 1.0 million ha (29% less than 0.5 million ha), and about 27% have areas greater than 5 million ha (11.5% greater than 20 million ha). The dominant forest type is needle leaf (Spruce, Pine, Fir and other conifers in Canada, greater than 60%; Larch, Pine and Spruce in Russia, greater than 70%; Spruce

54 34 and Pine in Finland, Norway and Sweden, about 70-90%). The area under broad leaf forest (mostly Oak) in the USA is comparable to that of needle leaves (Pine, Fir and Spruce), about 40% (Figure 2.5). About 40% of the forest area in Canada and 55% in Russia is mature or over-mature forests. The area under immature forests in Canada is about 30% (23% middle-aged forests in Russia). The area under regeneration is less than 10% in Canada (20% in Russia) (Figure 2.6). Thus, in a broad sense, the Canadian and Russian forest age structures are comparable. In the USA, fully three fourths of the forest area includes forests younger than 85 years, but these data probably are years old. These three large countries account for 77% of the forest area north of 30 0 N. Thus, the inventory data used in our analysis represent a wide variety of inventory practice, provincial forest acreage, ecosystem types, age structures, management practices, fire and insect dynamics Evaluation of Biomass from Inventory Data of Wood Volume The inventory data have information on stem wood volume. Above-stump biomass, the oven-dry weight in tons/ha of various biological components, is estimated from inventoried wood volume for needle leaf and broad leaf forests using equation (2.2) N ( C) WVN ( P) Bcf ( C) WVB ( P) AB( P) =, (2.2) FA( P) cf +

55 35 where AB is above-stump biomass (tons/ha), N cf is the conversion factor for conifers (tons biomass/m 3 stem wood), B cf is the conversion factor for broad leaves (tons biomass/m 3 stem wood), WV N is wood volume of needle leaf forest (m 3 ), WV B is wood volume of broadleaved forest (m 3 ) and FA is forest area (ha). The values for AB, WV N and WV B are derived for inventory data for individual provinces, and the values for N cf and B cf are assigned for individual countries. Total biomass is estimated by adding the root biomass, FFN ( P) FFB ( P) TB( P) = AB( P)[1 + ( + ) Rcf ( C)], (2.3) N ( C) B ( C) cf cf where TB represent the total biomass (tons/ha), FF N is forest fraction of conifers (% of pixel area), FF B is forest fraction of broad leaves (% of pixel area) and R cf is conversion factors for roots (tons biomass/m 3 stem wood). The conversion factors are country specific and are obtained from the Temperate and Boreal Forest Resources Assessment (TBFRA)-2000 (2000).

56 Matching NDVI and Biomass Data Satellite greenness can be derived by integrating the seasonal NDVI curve over growing season, and the forest woody biomass can be obtained from inventory stem wood volume data. The relation between biomass and cumulative growing season NDVI data requires matching inventory data to remote sensing data, such that the growing season NDVI totals are evaluated from forest land cover pixels only. The methodology is illustrated here, using Sweden as an example. Sweden spans a latitude range of about 55 o N to 70 o N, with 24 provinces for which the inventory data are available (Figure 2.7). The provinces are of different land and forest areas. The data reported are stem wood volume in cubic million meters and forest area in thousand hectares for various tree types and trunk size classes. Data are published in a series of statistical hand-books. We utilized data from two periods (82-86 and 93-97). To match these provincial inventory estimates to NDVI data, the distribution of forest area in each of the provinces, not just the total forest area, is required, because the NDVI data are 8 8 km pixel data. Therefore, we use a remote sensing land cover map, shown here in Figure 2.8. This map is at a spatial resolution of 1x1 km (Hensen et al., 2000).

57 37 For each province, in a Geographical Information System, we evaluate the cumulative growing season greenness from NDVI data layers, by averaging over forest pixels, as identified from the land cover map. Forests are defined as the following remote sensing land covers: broad leaf and needle leaf forests, mixed forests and woody savannas. This assures that the resulting provincial cumulative growing season greenness is assembled from NDVI data of forested regions only. Also, the degree to which total forest area estimates from inventory and remote sensing match, provides some confidence in both inventory and remote sensing data. This comparison is shown in Figure 2.9. The inventory stem wood volume data are converted to total and above stump biomass, and plotted against the provincial growing season cumulative NDVI, as in Figure 2.10a. A similar plot for Russia is shown in Figure 2.10b. The biome types from Hansen et al. (2000) are grouped into the six biome types (Needle leaf forest, broad leaf forest, savannas, shrubs, broad leaf crops and grass) to study the zonal distribution of land vegetation types. Needle leaf forest consists of evergreen and deciduous needle leaf forests, broad leaf forest contains evergreen and deciduous broad leaf forests. Half of mixed forests and woody savannas are grouped into broad leaf forests, and the other half to needle leaf forests. The dominant biome type depends on the latitude (Figure 2.11). For example, broad leaf forests dominate tropical regions (20 o N-30 o S), about 80%, and needle forests dominate the mid and high latitudes of the Northern Hemisphere (north of 30 o N), greater than 95%.

58 38 Appendix List of Provinces/States Used in This Research The total number of provinces/states in the six countries Canada (CAN), Finland (FIN), Norway (NOR), Sweden (SWE), Russia (RUS) and the United States (USA) is 182. Data from 167 provinces where forest area is greater than 15% of the land area (10% in RUS) were used in the regression analysis (Sweden for two time periods). These are listed below: RUSSIA (57 of 71; listed in the same order as in Alexeyev and Birdsey (1998)) Kaliningrad Oblast, Arkhangel'sk Oblast, Vologoda Oblast, Mumansk Oblast, Rep. of Karelia, Rep. of Komi, Leningrad Oblast, Novgorod Oblast, Pskov Oblast, Bryansk Oblast, Vladimir Oblast, Ivanov Oblast, Tver' Oblast, Kaluga Oblast, Moscow Oblast, Ryazan' Oblast, Smolensk Oblast, Tula Oblast, Yaroslavl' Oblast, Nizhniy Novgorod Oblast, Kirov Oblast, Rep. of Mari EI, Rep. of Mordvinia, Rep. of Chuvashia, Tambov Oblast, Samara Oblast, Penze Oblast, Ul'yanovsk Oblast, Rep. of Tatarstan, Krasnodar Kray, Rep. of Kabardino-Balkaria, Rep. of North Osetia, Rep. of Checheno-Ingushetia, Kurgan Oblast, Perm' Oblast, Sverdlovsk Oblast, Chelyabinsk Oblast, Rep. of Bashkortostan, Rep. of Udmurtia, Altai Kray, Kemerov Oblast, Novosibirsk Oblast, Omsk Oblast, Tomsk Oblast, Tyumen' Oblast, Krasnoyarsk Kray, Irkutsk Oblast, Chita Oblast, Rep. of Buryatia, Rep. of Tuva, Khabarovsk Kray, Amur Oblast, Kamtchatka Oblast, Magadan Oblast, Sakhalin Oblast, Rep. of Yakutia (Sakha).

59 39 NORWAY (17 of 17) Akershus, Aust-Agder, Buskerud, Hedmark, Hordland, Møre Og Romsdal, Nord- Trondelag, Nordland, Oppland, Østfold, Rogaland, Sogn Og Fjordane, Sør- Trøndelag, Telemark, Troms, Vest-Agder, Vestfold. FINLAND (8 of 9) Lappi, Oulu, Pohjanmaa, Kymi, Pohjois-Karjala, Pohjois-Savo, Keski-Suomi, and one region which is a combination of Mikkeli, Hame, Turku Ja Pori, and Uusimaa. SWEDEN (21 of 23) Älvsborg, Blekkinge, Gävleborg, Göteborg, Halland, Jämland, Jönköping, Kalmar, Kopparberg, Kronoberg, Norrbotten, Örebro, Östergötland, Skaraborg, Södermanland, Stockholm, Uppsala, Värmland, Västerbotten, Västernorrland, Västmanland. CANADA (11 of 12) Newfoundland, Nova Scotia, New Brunswick, Quebec, Ontario, Manitoba, Saskatchewan, Alberta, Yukon Territory, Northwest Territories USA (32 of 50) % needle forest < 40: Connecticut, Massachusetts, Rhode Island, Vermont, Delaware, Maryland, New Jersey, New York, Pennsylvania, West Virginia,

60 40 Michigan, Minnesota, Wisconsin, Indiana, Kentucky, Missouri, Ohio, Virginia, Tennessee % needle forest 40: Maine, New Hampshire, North Carolina, South Carolina, Florida, Georgia, Alabama, Mississippi, Arkansas, Louisiana, Idaho, Montana, Colorado

61 41 Table 2.1: Launch date, and ascending (northbound Equator crossing) and descending (southbound Equator crossing) node times in Local Solar Time for NOAA series of satellites. Satellite Launch Date (mm/dd/yyyy) Asc. Node (Launch) Des. Node (Launch) NOAA-6 06/27/ :30 07:00 NOAA-7 06/23/ :30 02:30 NOAA-8 03/28/ :30 07:30 Asc. Node (3/95) Des. Node (3/95) NOAA-9 12/12/ :20 02:20 21:16 09:16 NOAA-10 09/17/ :30 07:30 17:53 05:53 NOAA-11 09/24/ :30 01:30 17:23 05:23 NOAA-12 05/14/ :30 07:30 19:15 07:15 NOAA-13 08/09/ :40 01:40 NOAA-14 12/30/ :30 01:30 13:30 01:30 Cited from: the NOAA Polar Orbiter Data User's Guide (November 1998 version).

62 42 Table 2.2: Regression results between NDVI and climate in different latitudinal bands for the six biome types during July 1981 to May 1991 and February 1995 to December Latitude Band S 30 0 S N N N Biome Lag Types Months α β γ R 2 B B B B B B B B B B B B B B B B B B B B B B B B B B B B B B B1: grass and cereal crops, B2: shrubs, B3: broad leaf crops, B4: savanna, B5: broad leaf forests, and B6: needle forests.

63 43 Table 2.3: Regression results for NDVI-Climate equation in different latitudinal bands for the six biome types during July 1986 to May 1991 and February 1995 to December Latitude Band S 30 0 S N N N Biome Lag Types Months α β γ R 2 B B B B B B B B B B B B B B B B B B B B B B B B B B B B B B B1: grass and cereal crops, B2: shrubs, B3: broad leaf crops, B4: savanna, B5: broad leaf forests, and B6: needle forests.

64 NDVI 0.2 magnitude growing season duration Month Figure 2.1: Growing season NDVI totals (hatched area). Growing season NDVI total is determined by its two dimensions: growing season duration and the magnitude of the observations.

65 45 Spatially Average NDVI Anomaly Year Figure 2.2: Time series of spatial averaged NDVI anomaly for different latitudinal bands. The vertical solid line denotes the time of Mt. Pinatubo eruption. GIMMS denotes results are from the original GIMMS NDVI data, and BU-GIMMS denotes results from NDVI data after corrections of satellite orbital drift and Mt. Pinatubo eruption and climatepredicted NDVI for the June 1991 to January 1995 period.

66 Figure 2.3: Time series of spatial averaged NDVI for different latitudinal bands. The vertical solid line denotes the time of Mt. Pinatubo eruption. GIMMS denotes results are from the original GIMMS NDVI data, and BU-GIMMS denotes results from NDVI data after corrections of satellite orbital drift and Mt. Pinatubo eruption and climate-predicted NDVI for the June 1991 to January 1995 period. 46

67 a) Frequency (%) Provincial Forest Area (10 6 ha) Frequency (%) Provincial Forest Area (10 6 ha) b) 26.7 Figure 2.4: Distribution of provincial and state forest area in the six countries (Canada, USA, Russia, Sweden, Finland, and Norway). a) Number 1 on x-axis represents forest area less than 1 million ha, and 2 represents forest area between 1 and 2 millions hectare, and so on. b) The first bar is forest area less than 0.25 million ha, and the second bar is forest area between 0.25 and 0.5 million ha, and so on. The last bar represents forest area above 5 million ha.

68 48 USA No stocked (1.56%) others (5.55%) larch (0.53%) fir (11.3%) birch (8.02%) spruce (8.9%) pine (27%) oak (37.1%) Forest Area (million ha) Canada Unclassified (15.3%) Unspecified broadleaves (5.08%) Other broadleaves (2.1%) Maple (2.66%) Birch (3.87%) Polar (7.85%) Unspecified conifers (6.06%) Other conifers (1.49%) Larch (0.38%) Hemlock (1.99%) Fir (8.5%) Spruce (26.3%) Pine (16.3%) Forest Area (million ha) Russia Aspen (2.4%) Oak (1.3%) Fir (2.2%) Birch (12.9%) Spruce (10.8%) Pine (21.7%) N Larch (37%) Forest Area (million ha) Figure 2.5a: Distribution of forest area by genus in the USA, Canada, and Russia.

69 49 Sweden Windthrown (2.1%) Other broadleaves (2.0%) Beech (0.5%) Oak (0.9%) Aspen (1.3%) Birch (10.1%) Spruce (44.6%) Pine (38.5%) Forest Area (million ha) Norway Class I (6.1%) Deciduous (23.5%) Spruce (40.3%) Pine (30.1%) Forest Area (million ha) Finland Treeless (1.5%) Alder (0.3%) Aspen (0.3%) Birch (7.9%) Other conifers (0.1%) Spruce (251%) Pine (64.8%) Forest Area (million ha) Figure 2.5b: Distribution of forest area by genus in Sweden, Norway and Finland.

70 Figure 2.6: Age structure of forests in Canada, Russia and the USA. 50

71 Figure 2.7: Administrative map of Sweden with 24 provinces for which the inventory data are available. Sweden spans a latitude range of about 55 o N to 70 o N. 51

72 52 evergreen needle leaf forests evergreen broad leaf forests deciduous needle forests deciduous broad leaf forests mixed forests woody savannas savannas closed shrublands open shrublands grasslands croplands barren Figure 2.8: Land cover map of Sweden with 12 landcover types at a spatial resolution of 1x1 km (Hansen et al., 2000). Forests are defined as broad leaf and needle leaf forests, mixed forests and woody savannas.

73 Forest Area (1000 ha) Remote sensing estimate Inventory report (1982:86) Inventory report (1993:97) Nbtn Vbtn Jmtl Vnrl Gavl Kopp Vrml Oreb Vstm Upps Sthm Sodm Ostg Skbg Alvs Jkpg Kron Kalm Gotl Gtbg Hall Blek Skan Provinces Land Area (1000 ha) Remote sensing estimate Inventory report (1982:86) Inventory report (1993:97) Nbtn Vbtn Jmtl Vnrl Gavl Kopp Vrml Oreb Vstm Upps Sthm Sodm Ostg Skbg Alvs Jkpg Kron Kalm Gotl Gtbg Hall Blek Skan Provinces Figure 2.9: Comparison of forest and land area between remote sensing estimates and inventory reports.

74 Figure 2.10: Plots of total woody biomass versus cumulative growing season NDVI for individual provinces in (a) Sweden and (b) Russia. 54

75 55 Pixels 90,000 80,000 70,000 60,000 50,000 40,000 30,000 20,000 10, Grass & cereal crops 2. Shrub 3. Broadleaf crops 4. Savanna 5. Broadleaf forest 6. Needleleaf forest N 10-20N 20-30N 30-40N 40-50N 50-60N 60-70N Latitude Band Pixels 35,000 30,000 25,000 20,000 15,000 10,000 5, Grass & cereal crops 2. Shrub 3. Broadleaf crops 4. Savanna 5. Broadleaf forest 6. Needleleaf forest N 10-20N 20-30N 30-40N 40-50N 50-60N 60-70N Latitude Band Figure 2.11: Zonal distribution of biome types in Eurasia and North America.

76 56 Chapter 3 METHODS 3.1 Methods for Estimating Terrestrial Carbon Processes The modeling of environmental biogeochemistry has recently received much attention because of a recognition that understanding the global cycles of carbon, methane, nitrogen, etc. is an integral part of addressing the climate change issue. The modeling activity is complemented by a large observational program, which provides crucial data for constraining the models. Typically, biogeochemical observations include measurements of concentrations of chemical constituents, such as CO 2, but also optical properties of the biosphere on land. These observations may be compared to simulation results obtained by running a particular biogeochemical model. Discrepancies between simulation and observations may then indicate the need for model modification. Often, the inverse approach is applied to investigate the unknown sink-source distribution. A classical example of inverse modeling in global biogeochemical cycles consists of the quantification of large-scale spatial and temporal variations of sources and sinks of long-lived atmospheric trace gases such as carbon dioxide. The inverse method has been widely used in this context (Enting et al., 1993, Ciais et al., 1995, Bousquet et al., 1996,

77 57 Fan et al., 1998, Bousquet et al., 1999a, Bousquet et al., 1999b, Pacala et al., 2001). A recent IPCC assessment updated the global carbon budgets. These results are estimates from an ensemble of global eight inverse models using different techniques and sets of atmospheric observations after accounting for fossil-fuel emissions. This is a method that back-calculates sources and sinks of CO 2 from the distribution of atmospheric concentrations with an atmospheric transport model that gives the best match to global set of atmospheric CO 2 observations. However, this approach is subject to some serious limitations (Heimann and Kasibhatla, 2000). Because many trace gases are long-lived, the atmospheric spatialtemporal signatures in the concentration field are relatively small and require highprecision measurements. But, the measurements are tedious and expensive. On the other hand, the surface sources typically exhibit large heterogeneity depending on type, extent and state of the various ecosystems. Another fundamental limitation is the highly diffusive property of atmospheric transport. Small-scale source structures are quickly averaged out by vigorous atmospheric transport and mixing in the troposphere. The processes of atmospheric mixing act to attenuate the spatial information in the source/sink distribution (Enting, 2000). The estimation process requires an amplification factor as the inverse of the attenuation. However such amplification also amplifies observational and modeling errors.

78 58 The land-based approaches, incorporating direct inventories of carbon on the ground, land use change, and ecosystem models, are another class of methods for quantification of sources and sinks (e.g., Houghton et al., 1999, Caspersen et al., 2000). One approach is based on contributions from land-use change (Houghton et al., 1999), that is, rates of land-use change and changes per hectare in carbon that follow a change in land use. They consider the conversion of natural ecosystems to croplands and pastures, the abandonment of croplands and pastures, harvest of industrial wood and fuel wood, and fire management. Carbon accumulations are calculated by applying a growth rate to forest areas previously harvested. The rates of growth are held constant over time. The approach based on forest inventory is largely different from the land-use change method (Birdsey et al., 1993, Turner et al., 1995). Carbon accumulations obtained from forest inventories are based on measured rates of growth, which include both recovery from earlier harvests and other factors, such as climate change and CO 2 fertilization, mortality and natural losses. The changes in carbon evaluated with the land-use change method does not account for the total net flux of carbon between land and atmosphere. It represents only the portion of the flux that can be attributed to direct human activity. Errors result from uncertainties in rates of land-use change, from aggregated estimates of biomass, growth, and decay, and from simplifying assumptions in the structures of the model.

79 59 The U.S. estimates of carbon uptake from direct measurements of forest growth (Birdsey and Health, 1995) and from changes in land use (Houghton et al., 1999) are both considerably lower that the annual sink inferred from atmospheric inverse calculations (Fan et al., 1998). It is known that the inversion results are highly sensitive to the atmospheric data used, to the spatial and temporal resolution of modeling, and to the tracer transport model used. In contrast, estimates based on land-use changes provide only part of the accumulation of carbon in forests. In particular none of the methods provide the geographic detail. Robust techniques are therefore needed for mapping carbon stocks and fluxes. Satellite measurements of vegetation provide global coverage with relatively high spatial resolution and consistent time coverage that goes back two decades. This goal of our research is to use remote sensing data to investigate a spatial detail of the woody biomass carbon stocks, sources and sinks in northern forests. 3.2 The Biomass-NDVI Equation The relation between NDVI and inventory estimates of above-stump biomass and total biomass are shown in Figure 3.2, for all seven nations. The three outliers that are associated with growing season NDVI values beyond 110 are observations from the U.S. temperate forests, where biomass is either uncharacteristically low (southeastern states) or high (pacific northwestern states).

80 60 The relationship between biomass and NDVI is estimated from the sample data (without outliers) using the following equation: 1/Biomass = α +β [(1/NDVI)/Latitude 2 ] + γ Latitude, (3.1) in which Biomass is a measure of total or above-stump biomass obtained from inventory data, NDVI is the cumulative growing season NDVI averaged over a five year period prior to inventory date, Latitude is the centriod of the area sampled by forest inventory in a province, and α, β and γ are regression coefficients. The value of these coefficients is estimated using ordinary least squares (Table 3.1). Equation 3.1 specifies the relation between NDVI and biomass [(1/NDVI)/Latitude 2 ] such that this relation can vary across space. Over large spatial scales, biomes vary by latitude, with low biomass boreal forests at high latitudes and high biomass hardwood forests at mid latitudes. This latitudinal variation probably is not linear. Biomass increases with latitude north of N, where most of the world's deserts are located. To capture this nonlinear variation, we divide NDVI by the square of latitude. This specification implies that the amount of biomass that is associated with a given level of NDVI varies with latitude with the largest values in temperate latitudes (Figure 3.3a). Similarly, the relation between biomass and NDVI varies with latitude (Figure 3.3b).

81 61 A multivariate statistical analysis based on data collected between 1996 and 1998 from 15 European forests was performed to assess the effect of the single factors (latitude, precipitation, ecosystem type, elevation, mean annual temperature, age, management type, leaf area index) on net ecosystem exchange (NEE) (Valentini et al., 2000). Results show that latitude is the most significant single variable and a good proxy for the actions of a multiplicity of factors (for example, radiation balance, length of growing season, frost events, disturbance regime). Therefore, latitude provides valuable information for generating the relation between biomass and NDVI. 3.3 Evaluation of the Biomass-NDVI Equation The equation that is used to calculate biomass from NDVI (the biomass-ndvi equation) is estimated with data from seven samples; samples from a single period for five nations (Canada, Finland, Norway, Russia, Sweden, and the United States) and two periods for Sweden ( and ). Using these samples to estimate the relation between biomass and NDVI for all of North America and Eurasia begs two related questions: (i) does the relation between biomass and NDVI vary across spatial, temporal, and ecological scales; (ii) if the relation does vary, can Equation 3.1 be used to generate accurate estimates for biomass, and changes in biomass in countries where there are no forest inventory data to generate country-specific relations?

82 62 Estimating the biomass-ndvi equation from pooled data implies that the relation between biomass and NDVI does not vary among the seven samples. That is, the relation between biomass and NDVI in Russia is the same as the relation between biomass and NDVI in the USA (i.e., β Russia = β USA ). Making this assumption to estimate the biomass- NDVI equation and using the resultant equation to calculate biomass in nations not in the regression sample implies that the value of β represents the relation between biomass and NDVI for all spatial scales, time periods, and biomes in North America and Eurasia. As noted previously, the data used to estimate the biomass-ndvi equation represent a wide variety of inventory practices, provincial forest acreage, ecosystem types, age structures, management practices, fire and insect dynamics, and time. These differences could cause the biomass-ndvi equation to indicate a relation between NDVI and biomass when in fact no relation exists and/or could bias the statistical estimates for the regression coefficients (Hsiao, 1986). Such problems would affect the reliability of our estimate for biomass and ultimately, the carbon sink Representativity One way to evaluate the ability of the biomass-ndvi equation to represent the relation between biomass and NDVI across spatial, temporal, and ecological scales is to test the null hypothesis that the regression coefficients do not vary across the seven samples used to estimate the equation. This null hypothesis is evaluated by comparing a restricted

83 63 model, in which the value of the regression coefficients do not vary among samples, against an unrestricted model, in which the values of the regression coefficients are allowed to vary. From this perspective, the biomass-ndvi equation can be considered to be a restricted model. We test the null hypothesis that the values of the regression coefficients do not vary across the seven samples (ecological and temporal scales, broadly speaking) with a F test. This test compares a restricted model, in which the values of the regression coefficients are not allowed to vary, against an unrestricted model, in which the values of the regression coefficients are allowed to vary. First, we test the assumption that the intercepts (α) do not vary by comparing an unrestricted model, in which the intercepts are allowed to vary across the seven samples (Equation 3.2) against a restricted model in which the intercept is not allowed to vary across the seven samples (Equation 3.1): 7 1/Biomass = α i + β [(1/NDVI)/Latitude 2 ] + γlatitude, (3.2) i=1 in which i corresponds to the seven samples. The set of restrictions that equalizes the values of α is tested with a test statistic (ω ), which is given by equation 3.3,

84 64 ( RSS R RSSU ) / s ω =, (3.3) RSS /( T K) U in which T is the number of observations (167), K is the number of regressors in the unrestricted equation, s is one less than number of coefficients restricted to be equal (in this case, 6), RSS R is the residual sum of squares from the restricted model (Equation 3.1) and RSS U is the residual sum of squares from the unrestricted version of the equation (Equation 4). The test statistic (ω ) is distributed as a F with s and (T-K) degrees of freedom in the numerator and denominator, respectively. If ω exceeds the critical value, this result indicates that the residual sum of squares for the restricted model increases in a manner that is statistically significant at the relevant level of significance relative to the residual sum of squares for the unrestricted model, in which case we reject the null hypothesis that the intercepts are equal across the seven samples. The results indicate that we strongly reject the set of restrictions that equalize the values of α (F(6,158) = 7.53; p < ). Although the value of α varies by sample, this will have little effect on the estimate for the carbon sink. The carbon sink is calculated by subtracting the biomass estimates for from the biomass estimates for This subtraction eliminates the value of α (and also the value of γ that is associated with latitude). For the purpose of evaluating our ability to use the regression results from equation (3.1) to calculate the change in carbon storages, we test whether the relation between NDVI and biomass varies among the seven samples. We compare a restricted model

85 65 (equation 3.2), in which α varies across the seven samples, against an unrestricted version of the model (equation 3.4), in which both α and β vary across the seven samples, 7 1/Biomass = α i + β i [(1/NDVI)/Latitude 2 ] +γ Latitude. (3.4) i=1 7 i=1 The results indicate that we reject the null hypothesis that the β s are the same across the seven samples (F(6,151) = 2.59, p > 0.03), but much less strongly than we rejected the previous restriction. This result implies that the values for β vary among nations. To assess this variation, we estimate equation 3.1 seven times with observations for an individual nation and/or period and test the null hypothesis β = 0. The results indicate that there is a statistically meaningful relation between biomass and NDVI in nearly every nation and sample period (Table 3.2). For Finland, the relation is significant at the p < 0.1 level, but not at the p < 0.05 level. This result is not surprising given the small sample from Finland (the regression equation estimated from the Finnish data has only 5 degrees of freedom). For the USA, there is no statistically meaningful relation between NDVI and biomass. This failure is due to a single observation. If we remove this observation and reestimate equation 3.1, there is a statistically meaningful relation between NDVI and biomass (p < 0.01). Together, these results indicate that there is a statistically significant correlation between NDVI and biomass within nations.

86 Stability Variations in β among nations could affect our estimate for the change in the carbon pool if the differences are systematically associated with NDVI or latitude. Therefore, we evaluate the stability of the relation between biomass and NDVI statistically. One way to evaluate the relation is to estimate equation (3.1) with sub-samples that include values of NDVI equal to or greater than a pre-defined threshold. For this sub-sample, we can evaluate whether there is a relation between biomass and NDVI by testing the null hypothesis β = 0. Rejecting this null hypothesis would indicate that there is a relation between biomass and NDVI within the range of values for NDVI that is defined by the threshold. To explore this possibility, we estimate the relation between biomass and NDVI at every possible threshold for NDVI between 47 and 127 (the minimum and maximum values are chosen so that there are enough degrees of freedom to perform the statistical tests). To do so, we estimate equation (3.1) with a sub-sample that includes observations with a value for NDVI of 127 or greater and progressively lower the threshold by one unit. The relation between NDVI and biomass for samples defined by NDVI is shown in Figure 3.4a. Line (4) gives the significance level for the t statistic associated with β (equation 3.1) estimated with data that include values of NDVI equal to or greater than the value given on the X axis. The degrees of freedom in these regressions are given by line (2). Lines (3) and (4) represent the significance level at the p < 0.05 and p < 0.10

87 67 thresholds. Line (4) shows the significance level of the t statistic for the test β = 0. A value above either line (3) or line (2) indicates that β is not statistically different from zero at threshold of p < 0.05 or p < 0.1 (i.e. there is no relation between NDVI and biomass). Line (4) moves above line (3) when the regression sub-sample includes values of NDVI equal to or greater than 113. At this point, the regression sample has less than 23 degrees of freedom (see line (1)), which reduces the reliability of the statistical estimation. These results indicate that there is a relation between NDVI and biomass for nearly all values for NDVI. Another important issue is whether the relation between biomass and NDVI for values above the threshold is the same as the relation for values below the threshold. We evaluate this question by defining a dummy variable (DUM) that is equal to 1 for values of NDVI above a pre-defined threshold and equal to zero for values of NDVI equal to or below that threshold. We use this dummy variable to modify equation 3.1 as follows 1/Biomass = α 1 + α 2 DUM + β [(1/NDVI)/Latitude 2 ] + β 2 DUM [(1/NDVI)/Latitude 2 ] + γ Latitude. (3.5) The DUM variables allow the relation between NDVI and biomass to change at a threshold. That is, if the regression coefficient α 2 *DUM is statistically significant, such a result indicates that the intercept for the relation between NDVI and biomass is α 1 for values of NDVI equal to or less than threshold and α 1 + α 2 for values of NDVI greater

88 68 than the threshold. Similarly if the regression coefficient β 2 *DUM is significant, such a result indicates that the relation between biomass and NDVI is β 1 for values of NDVI equal to or less than the threshold and β 1 +β 2 for values of NDVI greater than the threshold. We can test whether the DUM variables in equation 3.5 are statistically significant with the ω-statistic (equation 3.3). But unlike our previous use, the ω-statistic cannot be evaluated against the standard F distribution. The standard F distribution will overstate the significance of the ω statistic because we evaluate sub-samples for all potential NDVI thresholds between 49 and 127 without a priori theory as to the NDVI threshold at which the regression coefficients change (Christiano, 1992). The lack of an a priori threshold affects interpretation as follows. Using equation 3.5, we test eighty thresholds. Random chance implies that four of the eighty ω statistics will exceed the p < 0.05 critical value for the F distribution. This would cause us to argue that the regression coefficients change between sub-samples. To account for the repeated sampling, (and the uneven distribution of observations), we simulate the distribution of the ω statistic that is unique to equation 3.5 using Monte Carlo techniques (Christiano, 1992). First we generate one thousand experimental data sets. Each data set is generated using the following equation 1/Biomass = α + β [(1/NDVI)/Latitude 2 ] + γ Latitude + µ, (3.6)

89 69 in which α, β and γ are the values estimated from the full sample. The 167 values of µ for each data set are generated by drawing randomly from a normal distribution with a mean value of zero and a standard error of 0.004, which is the standard error for the regression equation estimated from the full sample. The regression coefficients remain constant to ensure consistency with the null hypothesis of the test there is no change in the value of α and/or β at any NDVI threshold. Each experimental data set is estimated using equation 3.5 with each threshold between 49 and 127. The largest ω statistic from the analysis of each experimental data set (regardless of the threshold it is associated with) is saved. These thousand values are ranked in descending order by size, such that the value at position 50 represents the p < 0.05 threshold. That is, there is less than a five percent chance of generating a value for ω that is equal to or greater than this value if equation 3.5 is searched randomly over all possible break-points. The stability of the relation between NDVI and biomass for samples defined by NDVI is shown in Figure 3.4b. Line (5) or (4) tests whether the slope (DUM*β 1 in equation 3.5) or the intercept (DUM*α 1 in equation 3.5) differs between samples above and below a given threshold for NDVI, respectively. Line (1) tests whether the intercept and slope (DUM*α 1 and DUM*β 1 ) differ between samples above and below a given threshold for NDVI. Line (2) represents the number of observations below a given level of NDVI. Line (3) represents the p < 0.05 significance level. The results indicate that equation 3.1 describes the relation between biomass and NDVI over a wide range of

90 70 values for NDVI. As indicated by line (5), we cannot reject the null hypothesis that the value for DUM*β 2 in equation 3.5 is zero for nearly every value for the 0/1 threshold (except for a threshold of 49, for which the dummy variable has only one degree of freedom). Line (4) indicates that we cannot reject the null hypothesis that DUM*α 2 is equal to zero for all NDVI thresholds. As indicated by line (1), we are unable to reject the null hypothesis that both DUM*α 2 AND DUM*β 2 for thresholds above 67 and below 80. This range generates sub-samples that are approximately equal, which generate the most reliable results. As indicated by line (2), about 25 percent of the sample has a value for NDVI below 70 while about 40 percent of the observations have a value for NDVI above 80. Together, these results indicate that the relation between NDVI and biomass is stable over a wide range of sub-samples. 3.4 The Effect of Latitude on the Relation Between NDVI and Biomass Alternately, the relation between NDVI and biomass may vary over latitude. We can explore the effect of latitude on the relation between NDVI and biomass by estimating equation 3.1 with sub-samples that are defined by latitude (rather than by NDVI). The data used to estimate equation 3.1 include observations between 29 o N and 69 o N. To see if there is a relation between NDVI and biomass within latitudinal bands, we use data from these latitudinal bands to estimate equation 3.1 and test whether β is statistically different

91 71 from zero. The first sub-sample includes all observations north of 67 o N, and the next subsample includes all observations north of 66 o N. We repeat this expansion of the subsample until all observations are included. We also repeat this sampling starting with observations from low latitudes, such that the first sub-sample includes observations from 29 o N to 31 o N, the second from 29 o N to 32 o N, and so on. Regardless of the latitude that is used to truncate the sub-sample, we strongly reject (p < 0.01) the null hypothesis that β is equal to zero. The consistent rejection of this null hypothesis indicates that there is a statistically meaningful relation between NDVI and biomass, regardless of latitude. There is no a priori evidence to indicate whether the effect of latitude on the relation between biomass and NDVI is nonlinear (as represented by the square of latitude) or linear. The use of a linear specification can be evaluated statistically by estimating the following model: 7 1/Biomass = α i + β i [(1/NDVI)/Latitude 2 ] +γ Latitude, (3.7) i=1 7 i=1 and testing whether the slopes (β i in Equation 3.7 are the same across the seven samples. The ω statistic clearly rejects the null hypothesis that the slopes ( β ) in Equation 3.7 are the same [F(6,152) = 6.72, P < ]. Similar results [the slopes (β i vary across nation] are obtained for an equation in which NDVI is not modified by latitude [F(6,152) = 8.79, P < ]. Together, these results indicate that much of the variation in i

92 72 slopes (β i ) across samples is associated with latitude and that this effect is represented more accurately by the square of latitude. 3.5 Temporal and Spatial Relations Between NDVI and Biomass To explore whether the spatial relation between biomass and NDVI is different from the temporal variation between biomass and NDVI, we use observations for Sweden to estimate the following equation, 1/Biomass = α 1 + α 2 DUM β [(1/NDVI)/Latitude 2 ] + β 2 DUM8286 [(1/NDVI)/Latitude 2 ] + γ Latitude, (3.8) in which DUM8286 is a dummy variable that is equal to 1 for observations from the 1982 to 1986 period and is equal to 0 for observations from the 1995 to 1999 period. If the spatial relation between biomass and NDVI during the and periods is different from the temporal relation between biomass and NDVI between these two periods, DUM*α 2 and/or DUM*β 2 will not be equal to zero. Conversely, if the spatial relation between biomass and NDVI during the and periods is the same as the temporal relation between NDVI and biomass between these two periods, DUM*α 2 and/or DUM*β 2 will be zero.

93 73 We test whether DUM*α 2 and/or DUM*β 2 are equal to zero with the ω statistic that is evaluated against the standard F distribution. Results indicate that we cannot reject the null hypothesis that DUM*α 2 is equal to zero (F(1,36)= 0.005, p < 0.95), DUM*β 2 is equal to zero (F(1,36) = 0.01, p < 0.91), or DUM*α 2 and DUM*β 2 are equal to zero (F(2,36) = 0.07, p < 0.94). Together, these results indicate that the spatial relation between biomass and NDVI is not statistically different from the temporal relation between biomass and NDVI.

94 74 Table 3.1: Regression results for the Biomass-NDVI equation (3.1). Total biomass (± ) Above-stump biomass (±0.0136) Values in parentheses are standard errors. α β γ Adjusted R (±902.51) (± ) (± ) (± ) Table 3.2: Regression results for the total biomass equation (3.1) with data from an individual nation and period. Countries β 1 Standard Error t statistic Sweden (p <0.004) 18 Sweden (p <0.004) 18 Norway (p<0.0001) 14 Finland (p <0.07) 5 Canada (p<0.0001) 8 Russia (p <0.002) 54 USA (p <0.19) 29 USA * (p <0.01) 28 * Result for U.S. when one outlier is removed. Values that exceed the 0.01 threshold are in bold; Values that exceed the 0.05 threshold in italics. Degrees of Freedom

95 75 Amplification of Errors CO 2 Emissions Estimated Sources Loss of Information Atmospheric Transport Model Error Measurement Error Atmospheric Transport Model Atmospheric Concentrations Measured Concentrations Figure 3.1: Ill-conditioning in the tracer inversion problem. (Enting, 2000).

96 Figure 3.2: Plot of total (A) and above-stump (B) woody biomass versus cumulative growing season NDVI. The NDVI data are five-year averages prior to the date of inventories. Outlier 1 is British Columbia (CAN) and outliers 2 are data from Washington, Oregon and (northern) California (USA). These represent 16% of North American forest area. 76

97 a) Total Biomass (tons/ha) NDVI=80 NDVI=120 0 NDVI= Latitude (degrees) b) Total Biomass (tons/ha) lat=50 0 N Lat=70 0 N lat=30 0 N Growing Season NDVI Total Figure 3.3: The relations (a) between total biomass and latitude at three levels of total growing season NDVI, and (b) between total biomass and total growing season NDVI at three latitudes.

98 Significance Level (2, 3, and 4) a) (2) (1) (4) Degrees Freedom (1) 0.00 (3) Growing Season NDVI Total Significance Level (1, 3, 4, and 5) b) (5) (1) (3) (2) (4) Number of Observations (2) Growing Season NDVI Total Figure 3.4: (a) The relation between NDVI and biomass for samples defined by NDVI. (b) The stability of the relation between NDVI and biomass for samples defined by NDVI.

99 79 Chapter 4 RESULTS AND DISCUSSION 4.1 Patterns of Woody Biomass Sinks and Pools The regression analyses indicate the coefficients associated with NDVI total are significant and stable across a large portion of the observed range for NDVI, latitude, and among nations. Thus, the regression model obtained from pooled data (equation 3.1) will be used to generate all biomass estimates discussed below. Seasonal totals for NDVI are obtained by matching provincial estimates of forest area from inventory data and those from a high-resolution (1 km) satellite vegetation map (Hansen et al., 2000) in a geographical information system. Annual data are averaged over two five year periods, and , that correspond to the start and end of the satellite record. To calculate the change in carbon storage between these periods, the first five year average biomass estimates are subtracted from the second five year average estimates. The biomass estimates are converted to carbon by multiplying by 0.5, a standard factor for converting woody biomass to carbon. Figure 4.1 shows the difference in growing season NDVI totals between the two time periods for all vegetated regions. Growing season is defined as the period when composite (15-day) NDVI values are greater than 0.1. The forest fraction, expressed as the fraction of each quarter degree box area under forest land

100 80 covers (broad leaf forests, needle leaf forests, mixed forests and woody savannas; Hansen et al., 2000; (Figure 4.2), is used as a weight to convert NDVI totals to biomass of each pixel. Because of their high spatial resolution relative to inventory measurements, biomass estimates from satellite data provide spatial detail of the carbon pool and the location and magnitude of changes. The spatial picture of pixel level changes in the biomass pool, shown in Figure 4.3, depicts carbon gains, in excess of 0.3 ton C/ha/yr, in Eurasian boreal and North American temperate forests, and carbon losses, greater than 0.1 ton C/ha/yr, in some Canadian boreal forests. The gains in Eurasia are located over a large, broad, nearly contiguous swath of land, from Sweden (about 10 0 E, north of 60 0 N), through Finland, European Russia, central Siberia to trans-baikalia (120 0 E, north of 50 0 N). In North America, similar gains occur in the eastern temperate forests of the USA and in southern Ontario and Quebec below the 50th parallel. Carbon losses occur in Canada's boreal forests, from Newfoundland to Northwest territories, except for small fragments in northern Saskatchewan and Alberta, where carbon storage increases (about 110 o W and 60 o N). The biomass map in Figure 4.4 indicates larger pools, in ton C/ha, in North America compared to Eurasia (51 vs. 39). The average pool size in Europe and the USA is larger than in Canada and Russia (54-58 vs ). Among European countries,

101 81 Austria, France and Germany have relatively large pools (60, 67 and 73, respectively). The estimates for Finland, Norway and Sweden are comparable to Russia (35-40 vs. 38). 4.2 Comparison of Remote Sensing and Inventory Estimates We estimate that the 1.42 billion ha of temperate and boreal forests stored 61±20 Gt C during the late 1990s (Table 4.1). The estimate for carbon gain during the 1980s and 90s is 0.68±0.34 Gt C/yr. This is in the mid-range of estimates by Sedjo (1992) for the mid- 1980s (0.36 Gt C/yr) and Temperate and Boreal Forest Resources Assessment-2000 (Liski and Kauppi, 2000) for the early and mid-1990s (0.88 Gt C/yr). The sequestration rate, in ton C/ha/yr, is highest in Europe (0.84) and the USA (0.66), and least in Canada and China ( ), with intermediate values for Russia (0.44). As a result, sequestration rates in Eurasia and North America ( ) are similar. This implies that nearly 70% of the sink is in Eurasia (0.47 Gt C/yr), which is consistent with its forest area but is disproportionably large relative to its pool size (Table 4.1). The above-stump biomass sink (0.64 Gt C/yr) is nearly identical to the total biomass sink (0.68 Gt C/yr) (Table 4.2). This is not surprising, and is to be expected, because these numbers are obtained from regression relations which are essentially black box

102 82 representations. That is, they translate the observed changes in growing season NDVI totals to the dependent variables. Estimates for the carbon pool and sink in the woody biomass of temperate and boreal forests in individual nations are given in Table 4.3. We evaluate uncertainties in satellite estimates of biomass pool and changes by comparing these to national, provincial, and state estimates (Figure 4.5 and Figure 4.6; also Appendix at the end of this Chapter). Figure 4.5 compares estimates of the above-stump biomass carbon pool in 46 states of the USA (Cost, 1990), 11 provinces of Canada (Penner et al., 1997), and total biomass pool in 37 Eurasian countries (Liski and Kauppi, 2000). It also shows inventory biomass estimates for 10 U.S. states for which data are from two time periods available (FIA, 2000). The inventory estimates shown here are not used to estimate the regression models. The remote sensing estimates were obtained with these regression models from pixel-level cumulative growing season NDVI averaged over five years prior to the inventory dates, and using the high resolution satellite vegetation map for identification of forest pixels (Hansen et al., 2000). Both sets of biomass estimates are converted to carbon by multiplying by 0.5, a standard factor for converting woody biomass to carbon. Estimates for Japan are divided by 2 to correspond with the axes. Figure 4.6 compares changes in the woody biomass carbon pool in 22 provinces of Sweden, 9 states of the USA (FIA, 2000) and 37 Eurasian countries (Liski and Kauppi,

103 ). The Swedish data are changes between two successive inventories ( and ) of stem wood volume, converted to woody biomass in carbon units and divided by the time interval (11 years); similarly for the 9 states of the USA. The Eurasian data are for the early to mid-1990s period. Only the Swedish data were used to estimate the regression model. The remote sensing estimates for changes in carbon storage are differences in the predicted carbon pool for the respective time periods, expressed on an annual basis. The average absolute difference (DIFF 1 ) between remote sensing and inventory estimates is 10.4 ton C/ha for above-stump biomass, 16.1 ton C/ha for total biomass, and 0.33 ton C/ha/yr for changes in pool size, or 27%, 33% and 50% relative differences (DIFF 2 ) of the mean inventory estimates, respectively (Figures 4.5 and 4.6). The DIFF 1 and DIFF 2 are defined as, N DIFF 1 = x1 x2, (4.1) i= 1 DIFF 2 = N x 1 i= 1 N i= 1 x x 2 2, (4.2) in which x 1 is the remote sensing estimate for above-stump biomass or total biomass, or estimate for changes in pool size, and x 2 is the inventory estimate. N is the total sample number. The national inventory sink estimates, in Figure 4.6 (Liski and Kauppi, 2000),

104 84 are derived from wood volume increment and loss data (natural and fellings), unlike remote sensing estimates, which are biomass differences between two time periods. The comparability of the two estimates is thus noteworthy. We also evaluated changes in the carbon pool generated by uncertainty in the biomass-ndvi relation alone. The results indicate that the standard error for our estimates of change is one to two orders of magnitude smaller than the average change, 0.48 ton C/ha/yr (section 4.4). 4.3 Bias Analysis of the Estimates If the estimates for biomass generated by the remote sensing/statistical methodology are unbiased relative to those generated from inventory data, the data in figures 4.5 and 4.6 will lie along the 45 o line. By definition, this 45 o line has an intercept of zero and a slope of 1. To test whether this 45 o line describes the relation between the biomass estimates generated by remote sensing/statistical methodology and those generated from the inventory data, we estimate the following equation, Inventory = α + β Remote Sensing + µ, (4.3) in which Inventory is the estimate for biomass from forest inventories in Figures 4.5 or 4.6, Remote Sensing is the remote sensing biomass estimate generated in Figure 4.5 or 4.6, α and β are regression coefficients, and µ is a normally distributed random error term.

105 85 To test the null hypothesis that the intercept (α) is equal to zero, we use a t statistic. This test statistic will reject the null hypothesis if its value exceeds the value associated with the p < 0.05 threshold. Failure to exceed this threshold indicates that the intercept is not statistically different than zero. To test the null hypothesis that β equals 1.0, we use the ω statistic (equation 3.3). This test statistic will reject the null hypothesis if imposing a value of 1 on β causes the residual sum of squares for equation (4.3) to increase in a statistically significant fashion relative to the version of equation (4.3) in which β is allowed to assume the value that minimizes the residual sum of squares. Failing to exceed this threshold would indicate that β is not statistically different from one. Lastly, both of these hypotheses (α = 0, β = 1) can be tested jointly with the ω statistic. This test statistic will reject this null hypothesis if imposing the restrictions on α and β cause the residual sum of squares for equation (4.3) to increase in a statistically significant fashion relative to the version of equation (4.3) in which α and β are allowed to assume the values that minimize the residual sum of squares. Failing to exceed this threshold would indicate that α is not statistically different zero and β is not statistically different from 1.0. Results indicate that we fail to reject any of these hypotheses. The intercept of a line fit to data in Figure 4.5 is not statistically different from zero (t = 0.83, p < 0.42). Similarly, the slope of the line is not statistically different from 1.0 (F(1,112)= 0.40, p < 0.53). Finally, we cannot reject the null hypothesis (α = 0, β = 1) for the data in Figure

106 (F(2,112)= 0.35, p < 0.71). Similar results are obtained for the data in Figure 4.6. The intercept of a line fit to data in Figure 4.6 is not statistically different from zero (t = 0.05, p <.97). Nor is the slope of the line statistically different from 1.0 (F(1,66)= 1.28, p < 0.27). Finally, we cannot reject the null hypothesis (α = 0, β = 1) for the data in Figure 4.6 (F(2,66)=1.07, p < 0.35). Together, these results indicate that the biomass estimates generated by the remote sensing/statistical methodology are unbiased relative to the biomass estimates generated from inventory data. 4.4 Uncertainty of the Carbon Sink Estimate We estimate the effect of uncertainty in the statistical relation between biomass and NDVI on the estimate for the carbon sink by running a Monte Carlo simulation. Ideally, this Monte Carlo experiment would be simulated with the entire data set. For each pixel and period in North America and Eurasia, we would use the values of NDVI to calculate a value for biomass that includes an error. This error would be determined by the standard error for our estimate of biomass, which can be derived from the regression results. This process would be repeated a thousand times to generate a confidence interval for our point estimate of the carbon sink. Unfortunately, this process is not computationally feasible because the North American and Eurasian data set includes tens of millions of pixels.

107 87 To avoid these difficulties, we evaluate the effect of uncertainty in biomass-ndvi equation on our estimates for the size of the sink by running a Monte Carlo experiment for a simulated landscape in which NDVI does not change. To do so, we create a hypothetical landscape of 10,000 pixels (640,000 km 2 ) where NDVI is identical for each pixel for both periods. We use statistical techniques to calculate the standard error associated with the point estimate for biomass that is generated by the biomass-ndvi equation (equation 3.1). This standard error increases as the values for NDVI and latitude move away from the sample mean (83 and 54, respectively). This standard error is multiplied by a normally distributed random variable that has a mean value of zero and a variance of 1. The resultant estimate for the regression error is added to the point estimate to calculate a value of biomass for each pixel. This process is repeated for each pixel to generate a second value for biomass for each pixel. These two values are subtracted from each other and divided by two to calculate each pixel's change in carbon pool. These values are summed over the 10,000 pixels to calculate the total change in the carbon pool in the hypothetical landscape. The total is divided by 10,000 to calculate the mean change in carbon storage per pixel. This process is repeated one thousand times. We use these thousand observations to calculate a mean (and standard error) change in the carbon pool for the 640,000 km 2 hypothetical landscape where NDVI does not change. The results (Table 4.4) indicate that the mean estimate for the per pixel change in the carbon pool is statistically indistinguishable from zero and that the standard error of this mean is one or two orders of magnitude smaller than the positive per ha change in carbon

108 88 pool (sink) reported in the previous section (0.48 ton C/ha/yr). The small size of the standard error relative to the per pixel carbon sink does not vary greatly if we change the values for latitude and/or NDVI that are associated with the hypothetical 640,000 km 2 landscape (the size of standard error decreases as we increase the number of pixels included in the Monte Carlo simulation). The generality of this result indicates that it is highly unlikely that the size the carbon sink reported in the text is a statistical artifact of the uncertainty in the statistical relation between biomass and NDVI. 4.5 Comparison of Estimates for Canada, China, Russia and the USA The four large countries, Canada, China, Russia and the USA, account for 84% of the pool, 78% of the sink and 87% of the forest area. Therefore, it is important to compare remote sensing estimates for these four countries with independent estimates (Table 4.5). These comparisons are difficult because of differences in the definition of forest areas, periods for which the estimates are valid and large uncertainties associated with all estimates. The TBFRA-2000 estimates are generally valid for the mid-1990s. Our sink estimate for the USA (0.142 Gt C/yr) is comparable to the TBFRA-2000 estimate (0.166 Gt C/yr). It is greater than estimates for the 1980s, from both inventory (0.063 Gt C/yr by Turner et al., 1995, and Gt C/yr by Birdsey and Heath, 1995) and

109 89 land-use change studies (0.02 Gt C/yr by Houghton et al., 1999). Our estimates of pool size (12.5 Gt C) and forest area (215 Mha) for the late 1990s are comparable to those of TBFRA-2000 (13.85 Gt C, and 217 Mha, respectively). Since the 1970s, Canadian forests suffered fires and insect damage (Kurz and Apps, 1999), which is consistent with the loss of carbon in Figure 4.3. Our sink estimate, Gt C/yr, is similar to estimates reported by TBFRA-2000 (0.093 Gt C/yr) and the Canadian Forest Service (about Gt C/yr for the period). Our estimate is greater than the total terrestrial sink in Canada between 1980 and 1996 inferred by Chen et al. (2000) (0.053 Gt C/yr). Our estimates for pool size (10.6 Gt C) and forest area (239 Mha) for the late 1990s also are comparable to the TBFRA-2000 estimates (11.9 Gt C and 244 Mha). Our estimate for pool size in China (3.68 Gt C for the period) is slightly lower than the estimate by Fang et al. (2001) (4.75 Gt C for the period), and the remote sensing estimate of forest area (142.6 Mha) is greater than the Fang et al. estimate ( Mha for the period). These differences cause estimates of average pool sizes to differ. Our estimate (25.77 tons/ha for the period) is lower than that of Fang et al. (2001) (44.75 tons/ha for the period). Our sink estimate (0.039 Gt C/yr for the period period) is greater than Fang et al. s estimate (0.024 Gt C/yr for the period). These differences may be due to afforestation and reforestation programs (12.74 Mha in the period and Mha in the period).

110 90 The remote sensing estimate of Russian forest area, 642 Mha, is lower than estimates by TBFRA-2000 (816 Mha), Alexeyev and Birdsey (1998) (771 Mha) and Nilsson et al. (2000) (764 Mha). These differences may be associated with the definition of a forest. Forest and other wooded land (FOWL) in the FAO statistics is equivalent to what in Russia is defined as "forest land", which includes "forested area" and "unforested area". "Forested area" is defined by stocking density. "Unforested area" is land where stocking density is temporarily below the forested area threshold. In 1993, "Forest land" was 887 Mha, "Forested area" was 764 Mha and "Unforested area" was 123 Mha. Estimates by Alexeyev and Birdsey (1998) and Nilsson et al. (2000) may have included "forested area" only, which is not comparable to the remote sensing land cover definition of forests. These inventory eatimates fluctuate. For example, the area of stocked stands ("forested area") was estimated to be Mha in 1988, Mha in 1993, and Mha in The low estimate for forest area that is generated by the remote sensing methodology (642 Mha) may be caused by the coarse resolution of satellite data (8 8 km). This resolution may not capture tree stands in the forest-tundra of Russia, where small lots of sparse, open larch stands with extremely low growing stock (30-50 m 3 /ha) are located throughout the vast peatlands. In addition, Russia has about 35 Mha of dwarf shrub communities (Betula nana and others) which are counted as forests in inventory studies. The total area of plain and mountain forest-tundra forests is about 108 Mha, which may

111 91 not be classified as forest land cover in remote sensing data (recent unpublished analysis of V. Alexeyev). There is an additional Mha difference between remote sensing and inventory estimates. When expressed on a per ha forest area basis, the various pool estimates are comparable (38-43 tn C/ha). The difference in sink estimates between remote sensing and the TBFRA-2000 is smaller (0.44 vs. 0.53; in tn C/ha/yr). Nilsson et al.'s estimate for the biomass sink, Gt C/yr, is smaller than our (0.292 Gt C/yr) and the TBFRA-2000 estimates (0.423 Gt C/yr). Nilsson et al. could not have derived their sink estimate from stem wood volume data because the increment they quote (816 Mm 3 /year) on 760 Mha of forested area in 1990 is comparable to TBFRA-2000 estimate of 1134 Mm 3 /year on 886 Mha of FOWL area during the same period. If they used data for stem wood volume, the three sink estimates would be comparable on a per unit forest area basis. Alexeyev and Birdsey (1998) do not provide a sink estimate. It is not clear why the TBFRA-2000 estimate for forest area (816 Mha) is different than the remote sensing estimate, considering that the two agree well for Canada, USA and other countries.

112 Comparison of Pooled vs. Country-Specific Relations Equation (3.1) is estimated seven times with observations from individual nations and two time periods for Sweden. As discussed in section 3.3.1, there is a statistically meaningful relation between biomass and NDVI in nearly every nation and sample period (Table 3.2). The regression coefficients, α, β, and γ, are estimated for each nation and two time periods for Sweden (Table 4.6). The regression slopes for Norway and Russia ( vs ) are larger than the slope estimated from the pooled data (3809.7), but the slopes for the USA (1370.6) and Canada (1631.0) are much smaller than the pooled estimate. The slopes for other countries are comparable to the pooled estimate. The intercept is largest in the USA ( ), and least in Finland ( ). The different slopes and intercepts, estimated with observations from an individual nation to represent varying forest types, management practices, and time periods, will cause differences in the estimates of carbon pools and sinks. This issue is discussed below. The country-specific relations were evaluated with data from individual countries. There were available for one time period, except in the case of Sweden, for two time periods. Therefore, it is meaningful to compare the carbon pool sizes between the pooled estimates and country-specific estimates. The ranges of growing season NDVI totals vary among the countries, because of differences in location, dominant biome type, management practices, and time periods. I perform the comparison only in the range of sampled NDVI used in the regression analysis for each country (Table 4.7). Three

113 93 latitudes (35 o N, 55 o N, and 65 o N) are selected in the comparison to match different countries. Results show that there are nearly linear relations between woody biomass carbon storage and growing season NDVI total for both the pooled estimates and the countryspecific estimates (Figure 4.7). There are non-linear relations at lower latitudes for the pooled estimates and with large regression slopes for the country-specific estimates (Russia and Norway). The pooled regression model slightly underestimates in the range of low NDVI values and overestimates at large NDVI values, for the USA. Conversely, the pooled regression model for Russia overestimates in the range of low NDVI values and underestimates at large NDVI values. The pooled estimates overestimate for Canada, Finland, and Norway in their NDVI ranges, as compared to the country-specific estimates. Both the pooled relation and the country-specific relations are used to estimate carbon storage for the seven sample observations (Table 4.8). Total carbon storage in the six countries is comparable to the pooled estimate (48.1 vs Gt C). The carbon storage in North America is slightly larger from the pooled estimate than from the country-specific models (22.9 vs Gt C). Russian carbon storage is similarly overestimated (26.6 vs Gt C). The estimates for Scandinavian countries are comparable to the pooled estimates (1.815 vs Gt C). The pooled estimate of change in carbon storage is Gt C/yr for , which is comparable to the

114 94 country-specific estimate ( Gt C/yr). Comparisons of carbon pool and changes in carbon pools indicate that the pooled relation between biomass and NDVI is consistent with the country-specific relations. Therefore, it is not unreasonable to use the pooled relation for estimating carbon storage and changes in view of the fact that this allows working with countries for which the inventory data are not available. 4.7 Comparison of Pooled vs. Forest Type-Specific Relations The data used to estimate the relation between forest woody biomass and satellite greenness represent a wide variety of inventory practices, provincial forest areas, ecosystem types, age structures, and time periods (Figure 3.2). As stated in Chapter 3, the relation between forest biomass and NDVI can vary by latitude. Although latitude was suggested to be the most significant single variable to influence on net ecosystem exchange (Valentini et al., 2000), it is important to evaluate the effect of different forest cover types on the relation between biomass and NDVI. To explore these effects, I estimate the regression equation 3.1 for the provinces with dominant needle leaf forest (the fraction of needle leaf forest greater than 70%) and dominant broad leaf forest (the fraction of broad leaf forest greater than 70%). The values of regression coefficients estimated with data from dominant needle leaf forests and dominant broad leaf forests, together with data from total forests, are listed in

115 95 Table 4.9. The regression slope for needle leaf forests ( ) is larger than the slope estimated from the pooled estimate ( ), but the slope for broad leaf forest ( ) is smaller than the pooled estimate. The intercept is larger in the estimate from broad leaf forests (-0.021), least in the estimate from needle leaf forests (-0.06), and the pooled estimate in between ( ). The regression coefficient for latitude also varies among the different forest cover types. The different slopes and intercepts, estimated with sample data from dominant forest covers, will cause differences in the estimates of forest woody biomass. To evaluate these differences, I performed two control tests with the three regression models (coefficients are shown in Table 4.9) for two specific latitudes (40 o N and 70 o N). Both the pooled relation and the forest type-dependent relations are used to estimate forest woody biomass storage (Figure 4.8). In the boreal regions, such as 70 o N, the estimates by the regression models from total forests and needle leaf forests are comparable (Figure 4.8b). The biomass storage from the pooled estimate is slightly larger than the estimate by the regression model from needle leaf forests at nearly all NDVI values below 115. For the broad leaf forests in boreal regions, however, there is larger difference between the pooled estimate and broad leaf-dependent estimate for large values of NDVI. In the temperate regions, such as 40 o N, the pooled regression model generates smaller values for biomass at lower NDVI values (below 100), and larger values for values of NDVI greater than 100 (Figure 4.8a). The estimates for the pooled relation are consistent with the needle forest-dependent estimates for the NDVI below

116 However, there is large difference between these two estimates for the high NDVI values. The relation between biomass and NDVI for the needle leaf forest is evaluated with most sample provinces from boreal regions, and for the broad leaf forest with most sample provinces from temperate regions. The dominant forest type in boreal regions is needle leaf forest covers, and is broad leaf forests in temperate and tropical regions. As described above, there are small differences in biomass estimates generated by the pooled relation and needle forest type-dependent relation in boreal region. Similarly, the estimates from the pooled relation and the broad leaf forest relation show small differences in temperate regions. Based on these results, the errors introduced by using the pooled regression model to estimate forest woody biomass among different types may be small. 4.8 Reasons for the observed changes The reasons for the observed changes are not known but the spatial patterns seen in Figure 4.3 offer some clues. Increased incidence of fires and infestations in Canada (Figure 4.7), fire suppression and forest regrowth in the USA (Figure 4.8), declining harvests in Russia (Figure 4.9), improved silviculture in the Nordic countries (Figure 4.10), forest expansion (afforestation and reforestation) and regrowth in China (Figure

117 ), woody encroachment and longer growing seasons from warming in the northern latitudes possibly explain some of the changes (Houghton et al., 1999; Kurz and Apps, 1999; Caspersen et al., 2000; Keeling et al., 1996; Myneni et al., 1997). This implies uncertainty regarding the future of biomass sinks and therefore the need for monitoring. 4.9 Limitations of this Research How robust are these results? The conclusion may be weakened by the uncertainties from the accuracy and representativity of satellite and inventory data, the models, and so on. As stated in section 2.2, GIMMS NDVI data have satisfactory corrections for the north. However, residual atmospheric effects and calibration errors in satellite data cannot be ruled out because it is not enough to minimize the residual atmospheric effects by using only the maximum NDVI value within each 15-day interval. In addition, there are still errors from other processes, such as orbital drift and stratospheric aerosols from volcanic eruption, as discussed in chapter 2. Further corrections need to be performed when more ground observations are available. Thus far, the inventory data are the best survey of forests. However, the inventory data are available for certain countries and regions only, and representativity is weakened by coarse spatial sampling and heterogeneity of land cover. Moreover, natural climate

118 98 and management practices vary significantly among countries. Uncertainties in inventory data are country-specific and difficult to quantify (Liski and Kauppi, 2000). Simple models are used to convert wood volume and satellite greenness data to biomass, as discussed in chapter 3. The validity of the biomass-ndvi relation at all scales is open to question. The differences in estimates of forest area generated from remote sensing and forest inventories are not easy to reconcile because of definition problems Significance of this Research The scientific contributions of this research are in the context of global carbon cycle research. This work provides spatial detail on the location of biomass carbon pools and where changes in this pool have occurred at a resolution that permits direct validation with ground data. The NDVI data, when used in inversion studies provide additional constraints to inferences regarding the distribution of sources and sinks for atmospheric CO 2 and isotopic concentration data. These inversion studies cannot partition the inferred sink between vegetation, soil and other pools. For example, if the vegetation is a sink and the soil is a source, estimates for changes in the vegetation pool would complement inversion results.

119 99 The political implications of this research are in the context of developing policies for reducing emissions of greenhouse gases. For example, debate is currently underway regarding which of the forest biomass sinks can be used by the Annex 1 parties, the industrialized nations, to meet their commitment to reduce greenhouse gas emissions under the Kyoto Protocol of the UNFCCC. Satellite estimates of biomass changes can be used to verify compliance (Nilsson et al., 2000), if the uncertainty of the remote sensing/statistical estimates can be further reduced. Improved observations of greenness levels from a new generation of spacecraft sensors such as MODIS and MISR, and possibly direct biomass measurements with lidars, offer promise for the future. The economic implication of this research is that the wood volume maps are valuable information to the forest industry, because wood from forests is the raw material for a multi-billion dollar global industry. This study suggests the possibility of surveying forests from space and making wood volume maps across a wide variety of forests Analysis of Emissions and Sinks of Industrialized Nations Carbon sinks can have an important and measurable effect on atmospheric carbon dioxide concentration. If these can be measured and monitored with sufficient accuracy, it allows sink management, accounting, and financial incentives for carbon sequestration services

120 100 (Sedjo and Toman, 2001). Well-established methods for monitoring carbon stocks typically involve sample plots. As discussed previously, remotely sensed vegetation greenness can be a good surrogate for forest biomass, especially in the north where there is presently a large carbon sink (Figure 4.14). NASA developed satellite data indicate that most northern forests were storing carbon. Forests in America, Europe and Russia have been storing nearly 700 million metric tons of carbon a year. American forests stored 140 million tons carbon a year, and Russia, the country with most forests, account for almost 40 percent of the biomass carbon sink. Although Canada s boreal forests were found to be losing carbon, our estimates show that Canada as a whole is a carbon sink and ranks in the top three among the Annex 1 countries. Since the 1770s, roughly 270 Gt carbon have been released to the atmosphere from combustion of fossil fuels and cement production (Marland et al., 2001). Half of these emissions have occurred since the mid 1970s (Figure 4.15). North America, including the United States and Canada, is the highest fossil-fuel CO 2 emission region in the world with 1.61 Gt carbon emitted in 1998 (Marland et al., 2001). Emissions from the United States account for 92 percent of current fossil-fuel CO 2 emissions from the North America. Fossil-fuel CO 2 emissions from Western Europe (United Kingdom, Italy, France, and Spain contribute 65%) are virtually unchanged since 1973, and fell 10.5% from Eastern Europe during the 1990s to the early 1980s level.

121 101 Carbon emissions can be treated as a surrogate of a nation s degree of industrialization. Countries may be ranked by their sinks to emissions ratio to possibly assess their activities towards controlling greenhouse gas emissions (Figure 4.16). Carbon sinks from our estimates, 700 million tons of carbon per year in America, Europe and Russia, are about 12% of annual global carbon emissions from industrial activities during 1980s and 1990s. The USA forest sink is about 11% of USA s annual emissions. Among the Annex 1 countries, the sinks to emissions ratio is above 60% in Sweden, Russia, and Canada, and is close to 40% for Finland and Norway. All these countries have extensive forest area. The rank based on sinks to emissions ratio per capita will assess the efficiency of industrialization weighted by the nation s forest area (Figure 4.17). High values of this ratio are seen in Sweden, Norway, and Finland. The most industrialized nations, with high population and less forest area, such as USA, Japan, Germany, and France, rank at the bottom of this group. Under the Kyoto Protocol, certain human-induced activities in the land-use, land-use change and forestry (LULUCF) sector that remove greenhouse gases from the atmosphere, namely afforestation, reforestation and reduced deforestation, may be used by Annex I countries to offset their emission targets (KPR, 1998). Conversely, changes in these activities that deplete carbon sinks will be subtracted from the amount of emissions that an Annex I country may emit over its commitment period. However, many issues remained unresolved in the Kyoto Protocol and have been the subject of continuing negotiations. The outstanding issues include the elaboration of agreed definitions for

122 102 "afforestation", "reforestation" and "deforestation". The Bonn Agreements set out a number of principles to govern the treatment of LULUCF activities under the Kyoto Protocol, and also register agreement on definitions of "forest", "afforestation", "reforestation" and "deforestation" (Bonn, 2001). The protocol allows counting certain carbon sinks as part of a nation's emissions reduction commitment, within some limits, and even trading of carbon sinks between nations. This raises the question of how these Kyoto sinks can be identified and separated from all sinks. Some of the carbon sources and sinks identified in our work are key to realizing the Kyoto Protocol. This research provides some clues. More research is however needed to identify the Kyoto sinks. According to the definitions in Kyoto Protocol, only the human-induced sinks can be taken into account. When new generation spacecraft sensors with a higher spatial resolution are used together with improved ground observations, satellite greenness may be used in monitoring programs. This research is a first step in that direction.

123 103 Appendix: Lists of provinces, states and countries in Figure 4.5 and 4.6 a) The following provincial (Canada), state (USA) and national data were used in Figure 4.5. CANADA (11) Alberta, Manitoba, New Brunswick, Newfoundland, Northwest Territories, Nova Scotia, Ontario, Quebec, Saskatchewan, Yukon Territory. Not included: British Columbia USA (46) Alabama, Alaska, Arizona, Arkansas, Colorado, Connecticut, D. Columbia, Delaware, Florida, Georgia, Idaho, Illinois, Indiana, Iowa, Kansas, Kentucky, Louisiana, Maine, Maryland, Massachusetts, Michigan, Minnesota, Mississippi, Missouri, Montana, Nebraska, Nevada, New Hampshire, New Jersey, New Mexico, New York, North Carolina, North Dakota, Ohio, Oklahoma, Pennsylvania, Rhode Island, South Carolina, South Dakota, Tennessee, Texas, Utah, Vermont, Virginia, West Virginia, Wisconsin, Wyoming. Not included: California, Hawaii, Oregon, Washington EURASIAN COUNTRIES From TBFRA-2000 (37) Albania, Armenia, Austria, Azerbaijan, Belgium, Bosnia, Bulgaria, Byelarus, Croatia, Czech, Denmark, Estonia, Finland, France, Georgia, Germany, Greece, Hungary,

124 104 Ireland, Italy, Japan, Kazakhstan, Latvia, Lithuania, Netherlands, Norway, Poland, Portugal, Romania, Slovakia, Slovenia, Spain, Sweden, Switzerland, Turkey, Ukraine, United Kingdom. Countries not included: (1) Russia and China are given in Table 4.1 (2) Kyrgyzstan, Macedonia and Uzbekistan: Sink data not given in TBFRA-2000 (3) Cyprus, Iceland, Israel, Luxemburg, Moldova, Turkmenistan, and Uzbekistan: Forest area less than 0.1 million ha USA (10) Arkansas (1988, 1995), Florida (1987, 1995), Georgia (1989, 1997), Mississippi (1987, 1994), North Carolina (1984, 1990), South Carolina (1986, 1993), Texas (1986, 1992), Virginia (1986, 1992), Wisconsin (1983, 1996). b) The following provincial (Sweden), state (USA) and national data were used in Figure 4.6. USA (9) Arkansas (1988, 1995), Florida (1987, 1995), Georgia (1989, 1997), Mississippi (1987, 1994), North Carolina (1984, 1990), Texas (1986, 1992), Virginia (1986, 1992), Wisconsin (1983, 1996). South Carolina not included because of shows decrease in biomass.

125 105 EURASIAN COUNTRIES From TBFRA-2000 (37) Albania, Armenia, Austria, Azerbaijan, Belgium, Bosnia, Bulgaria, Byelarus, Croatia, Czech, Denmark, Estonia, Finland, France, Georgia, Germany, Greece, Hungary, Ireland, Italy, Japan, Kazakhstan, Latvia, Lithuania, Netherlands, Norway, Poland, Portugal, Romania, Slovakia, Slovenia, Spain, Sweden, Switzerland, Turkey, Ukraine, United Kingdom. Countries not included: (1) Russia and China are given in Table 4.1 (2) Kyrgyzstan, Macedonia and Uzbekistan: Sink data not given in TBFRA-2000 (3) Cyprus, Iceland, Israel, Luxemburg, Moldova, Turkmenistan, and Uzbekistan: Forest area less than 0.1 million ha SWEDEN (22) Älvsborg, Blekinge, Gävleborg, Göteborg, Gotland, Halland, Jämtland, Jönköping, Kalmar, Kronoberg, Norrbotten, Örebro, Östergötland, Skän, Skaraborg, Södermanland, Stockholm, Uppsala, Värmland, Västerbotten, Västernorrland, Västmanland. Not included: Kopparberg because of data quality issues

126 106 Table 4.1: Remote sensing estimates of biomass carbon pool ( ) and sinks in temperate and boreal forests of North America and Eurasia. Country Average Pool (tons/ha) Carbon Pool (Gt C) Carbon Sink (Gt C/yr) Forest Area (Mha) Canada USA North America China Finland Japan Russia Sweden Other Eurasia Total Albania, Armenia, Austria, Azerbaijan, Belarus, Belgium, Bosnia and Herzegovina, Bulgaria, Croatia, Cyprus, Czech Republic, Denmark, Estonia, France, Georgia, Germany, Greece, Hungary, Italy, Kazakhstan, Kyrgyzstan, Latvia, Lithuania, Netherlands, Norway, Poland, Portugal, Romania, Slovakia, Slovenia, Spain, Switzerland, Tajikistan, Turkey, Turkmenistan, United Kingdom, Ukraine, Uzbekistan.

127 107 Table 4.2: Remote sensing estimates of carbon pools ( ) and sinks in the above stump biomass of temperate and boreal forests in North America and Eurasia. Country Carbon pool, Carbon Sink, Forest area, Gt C Gt C/yr Mha Canada USA North America China Finland Japan Norway Russia Sweden Other Eurasia Total Albania, Armenia, Austria, Azerbaijan, Belarus, Belgium, Bosnia and Herzegovina, Bulgaria, Croatia, Cyprus, Czech Republic, Denmark, Estonia, France, Georgia, Germany, Greece, Hungary, Italy, Kazakhstan, Kyrgyzstan, Latvia, Lithuania, Netherlands, Norway, Poland, Portugal, Romania, Slovakia, Slovenia, Spain, Switzerland, Tajikistan, Turkey, Turkmenistan, United Kingdom, Ukraine, Uzbekistan.

128 108 Table 4.3: Country estimates of carbon pools and sinks in the woody biomass of temperate and boreal forests. Country Carbon Pool Carbon Sink Forest Area (Mt C) (Mt C/yr) (million ha) Albania Armenia Austria Azerbaijan Belgium Bosnia Bulgaria Byelarus Canada China Croatia Czech Denmark Estonia Finland France Georgia Germany Greece Hungary Italy Japan Kazakhstan Kyrgyzstan Latvia Lithuania Macedonia Netherlands Norway Poland Portugal Romania Russia Slovakia Slovenia Spain Sweden Switzerland Turkey Turkmenistan United Kingdom Ukraine USA

129 109 Table 4.4: Monte Carlo simulation results to evaluate the change in carbon storage that is generated by the uncertainty in biomass-ndvi equation. Growing Season NDVI Total Latitude E ( ) ( ) ( ) E-04 ( ) ( ) ( ) E-04 ( ) ( ) ( ) The number is mean per pixel changes in carbon pool (tons C/yr/pixel); the values in parenthesis are the standard errors for this estimate.

130 110 Table 4.5: Comparison of estimates for Canada, China, Russia, and the USA. Country Remote Sensing Estimates Pool Gt C Sink Gt C/yr Area Mha Inventory & Other Estimates Pool Gt C Sink Gt C/yr Area Mha Canada China Russia USA TBFRA-2000 (Liski and Kauppi, 2000); estimates for the period early to mid 1990s. 2. From inventory data (CFS, 1993); for the period. 3. Fang et al. (2001); for the period. 4. Nilsson et al., (2000); for Alexyev and Birdsey (1998); for Turner et al. (1995); for the 1980s. 7. Birdsey and Heath (1995); for the 1980s. 8. Houghton et al. (1999); for the 1980s; land-use study. 9. Pacala et al. (2001); for the period; forest trees in coterminous USA only.

131 111 Table 4.6: Regression results for the total biomass equation (3.1) with data from individual country and for Sweden with two time periods. Countries β α γ (*10-4 ) Sweden Sweden Norway Finland Canada Russia USA Table 4.7: Growing season NDVI total range and latitude for the six countries. Country Growing Season NDVI Total Minimum Values Maximum Values latitude Canada o N USA o N Finland o N Norway o N Russia o N Sweden o N

132 112 Table 4.8: Comparison of carbon storage estimated by the pooled relation and the country-specific relations for the six countries. Country Time Period Carbon Pool (Gt C) Pooled Estimates Country-specific estimates Relative Difference of Carbon pool estimates (%) 1 Canada USA Finland Norway Russia Sweden Sweden The relative difference in carbon pool estimates is defined as: (Pooled estimates Country-specific estimates)/country-specific estimates. Table 4.9: Regression results for the total biomass equations (3.1) with data from different biome types (Needle forests, broad leaf forests, and total forests). α β γ R 2 DF Needle forests Broad leaf forests Total forests Values that exceed the 0.01 threshold are in bold; Values that exceed the 0.05 threshold in italics.

133 113 Figure. 4.1: Difference in growing season NDVI totals between two time periods, and , for all vegetated regions. Figure 4.2: Map of forest fraction defined as the fraction of each quarter degree pixel area under forest land covers. Forests include broad leaf forests, needle leaf forests, mixed forests and woody savannas, land covers in Hansen et al. (2000) classification.

134 114 Figure 4.3: Spatial detail of changes in woody biomass carbon pool of northern temperate and boreal forests between late 1990s and early 1980s. Biomass estimates were converted to carbon by multiplying by 0.5, a standard factor for converting woody biomass to carbon (TBFRA-2000, 2000).

135 115 Figure 4.4: Spatial detail of pool size in the northern temperate and boreal forests during late 1990s. Biomass estimates were converted to carbon by multiplying by 0.5, a standard factor for converting woody biomass to carbon (TBFRA-2000, 2000).

136 116 Figure 4.5: Comparison of remote sensing and inventory estimates of the biomass carbon pool. We show estimates at the provincial, state and national level, rather than on per unit forest area basis, to include uncertainties associated with differences in respective estimates of forest area.

137 117 Figure 4.6: Comparison of remote sensing and inventory estimates of the rate of biomass carbon pool change. We show estimates at the provincial, state and national level, rather than on per unit forest area basis, to include uncertainties associated with differences in respective estimates of forest area.

138 118 Woody Biomass Carbon (tons/ha) Canada (55 o N) USA (35 o N) Russia (65 o N) Sweden (65 o N) Norway (65 o N) Finland (65 o N) Growing Season NDVI Total Figure 4.7: Comparison of the pooled estimates (dash lines) and country-specific estimates (solid lines) for six countries (Canada, USA, Russia, Sweden, Norway, and Finland). The values of latitude are set to 35 o N for the USA, 55 o N for Canada, and 65 o N for the other countries.

139 needle leaf forest broad leaf forest total forest Total Biomass (tons/ha) a Growing Season NDVI Total 140 Total Biomass (tons/ha) needle leaf forest broad leaf forest total forest 20 b Growing Season NDVI Total Figure 4.8: Comparison of the pooled estimates and forest type-specific estimates at (a) the latitude of 40 o N, and (b) the latitude of 70 o N.

140 120 Figure 4.9: Detailed map of changes in the carbon pool for Canada. Figure 4.10: Detailed map of changes in the carbon pool for the USA.

141 121 Figure 4.11: Detailed map of changes in the carbon pool for Russia. Figure 4.12: Detailed map of changes in the carbon pool for European countries.

142 Figure 4.13: Detailed map of changes in the carbon pool for China and Japan. 122