Remote sensing as a tool to detect and quantify vegetation properties in tropical forest-savanna transitions Edward Mitchard (University of Edinburgh) Presentation to Geography EUBAP 10 th Oct 2008 Supervisor: Patrick Meir (Edinburgh) 2 nd supervisors: Sassan Saatchi (NASA JPL & UCLA) Iain Woodhouse (Edinburgh) Tim Malthus (Edinburgh) France Gerard (CEH Wallingford)
Why Budongo? Unique dataset from early 30s Remeasure Plot 7 (diameter, species and mapping of all stems > 10cm), as 'unlogged'! GPS location of other plots to help calibrate/ground-truth historical remote sensing products
2008
2000
1986
1973
2008 1986 1973
Budongo Plot 7 Biomass 1939-2008 Mg ha -1
Key questions Can woody cover / biomass of tropical forests be quantified from space? Change detection? Scaling up to continental scales? How accurately? Realworld applications.
Estimation of biomass changes in a forest-savanna transition region of central Cameroon Mbam Djerem National Park Interface Outline between of rainforest talk: and savanna Fieldwork Optical 1650 mm change annual detection precipitation Radar Very low change human detection population within national Future work park
Field data Spent 6 weeks in central Cameroon Set up eight 20 x 200 m transects running from forest to savanna Access to data from ten 1 hectare (100 m x 100 m) plots
How can you non-destructively weigh >7000 trees? For every tree with a diameter >5 cm measured: - Position - Height - Species - Diameter - Canopy breadth Collated wood density values for 130 species Calculated biomass using dry forest allometric equation from Chave et al. (2005) AGB - 2.187 e (0.916 x ln( D H
Reflectance (%) How does a satellite see trees? 1. Optical satellite Reflection: leaves soil - absorb strongly in red and blue - reflect in near infrared Vegetation Indices used, e.g. Normalised Difference Vegetation Index (NDVI) Wave Length (nm)
NDVI change detection - methods Took Landsat TM image from 1986, Landsat ETM+ image from 2000, and ASTER scenes from 2006 geo-referenced them all to the 2000 image, re-sampled to 28.5 m pixels 1986 10 km 2000 2006 Landsat TM 30 th Dec 1986 Landsat ETM+ 12 th Dec 2000 ASTER 27 th Nov 2006
Methods Sensor-specific radiometric and atmospheric correction Calculated NDVI values for these dry season images response to woody vegetation only (grasses all dead) Cross-calibration using known unchanged targets
Landsat 1986 NDVI 0.5 0.4 0.3 0.2 0.1 Woody savanna Gallery forest / rainforest 0-0.3-0.2-0.1 0 0.1 0.2 0.3 0.4 0.5 0.6-0.1 Lake/river -0.2-0.3 Dry grassland -0.4-0.5 Landsat 2000 NDVI
Excellent relationship found between NDVI and field-measured canopy cover 0.55 0.50 ASTER 2006 NDVI 0.45 0.40 0.35 0.30 r 2 = 0.87 p < 0.0001 0.25 0 1 2 3 4 Canopy cover (m 2 /m 2 )
Change-detection Calculated the ratio: Change NDVI = new old new + old for 86 00 & 00 06 Resulting images had values between -1 and 1 Thresholded resulting two greyscale images using standard deviations Histogram showing change 00 06
Change-detection > +4 S.D.s +2 to +3 S.D.s +1 to +2 S.D.s No Change ( 1 S.D) -1 to -2 S.D.s -2 to -3 S.D.s < -3 S.D.s Water / no data 12.57 % significant positive change 86 00 7.76 % significant positive change 00 06 < 0.04 % significant negative change in either image ETM+ 2000 Δ NDVI 1986 2000 Δ NDVI 2000 2006
> +4 S.D.s +2 to +3 S.D.s +1 to +2 S.D.s No Change ( 1 S.D) -1 to -2 S.D.s -2 to -3 S.D.s < -3 S.D.s Water / no data 1986 to 2000 2000 to 2006 1986 to 2000 1986 2000 2006
Fire frequency over the study area Number of fires 16 14 12 10 8 6 4 r 2 = 0.57 p < 0.01 2 0 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 Year Derived from ATSR-2 and AATSR World Fire Atlas, using the more sensitive Algorithm 2.
Rainfall changes Average monthly rainfall (mm/month) 400 300 200 100 0 1900 1910 1920 1930 1940 1950 1960 1970 1980 1990 2000 Annual average CRU Annual average TRMM December-February average CRU December-February average TRMM July-September average CRU July-September average TRMM Year Average monthly rainfall for rainiest three months, annual average, and driest three months over study area.
How does a satellite see trees? 2. Radar satellite Leaves, twigs, and trunks of trees all cause radio waves to scatter back to the satellite Simple log relationship often found Not that well understood Active radar pulse Return signal (backscatter)
Radar backscatter data Comparing two L-band satellite SAR datasets for a 15 000 km 2 area: - JERS HH from 1996 (100 m) - ALOS PALSAR HH/HV from 2007 (12.5 m) Georeferenced ALOS HV to Quickbird (60 cm) JERS 1996 ALOS 2007
Radar backscatter data Comparing two L-band satellite SAR datasets for a 15 000 km 2 area: - JERS HH from 1996 (100m) - ALOS PALSAR HH/HV from 2008 (12.5m) Georeferenced ALOS HV to Quickbird (60cm) Resample to 25 m resolution Extracted backscatter values over all the fieldsites Compared sigma 0 values with structural features and then biomass estimates from field sites
Relationship between estimated plot BIOMASS and 2007 ALOS backscatter ALOS HV Sigma 0 (db) -14-15 -16-17 -18-19 -20 ALOS HV against biomass r 2 = 0.84 p < 0.0001 0 100 200 Biomass (Mg ha -1 300 400 ) ALOS HH Sigma 0 (db) -6-7 -8-9 -10-11 ALOS HH against biomass Biomass map at 25m resolution for 2007 using ALOS HV data r 2 = 0.55 p < 0.0001 0 100 200 Biomass (Mg ha -1 300 400 )
Calibration of JERS to Biomass Use this to produce biomass map from 1996 Georeferenced JERS 1996 to ALOS HV 2007, masked areas not found in ALOS No field data from 1996, so cross-calibration using 26 x 7km 2 suspected unchanged areas Regressed the biomass of these plots against JERS data JERS HH Sigma 0 (db) -6-8 -10-12 -14-16 -18 r 2 = 0.99 p < 0.00001 0 50 100 150 200 250 300 350 ALOS HV-derived biomass (Mg ha -1 )
Biomass decreased, over whole study area by 17 %, equivalent to 2.27 Mg ha -1 yr -1. BUT increased by 0.98 Mg ha -1 yr -1 in 1996 the 5 300 km 2 lowpopulation density area. Decreased by 4.13 Mg ha -1 yr -1 in the 2007 Change detection Low population High poplulation Change ratio: remaining 9 600 km 2. 2007 1996 Change = 2007 + 1996
Total area (km 2 ) Total area (km 2 ) Total area (km 2 ) Changes in biomass class JERS 1996 ALOS 2007 a) all data b) less-populated area c) Populated area 6000 2500 4000 5500 5000 4500 4000 3500 Av. 1996 = 168.9 Mg ha -1 Av. 2007 = 143.9 Mg ha -1 2250 2000 1750 1500 Av. 1996 = 170.4 Mg ha -1 Av. 2007 = 181.2 Mg ha -1 3500 3000 2500 Av. 1996 = 168.0 Mg ha -1 Av. 2007 = 122.6 Mg ha -1 3000 1250 2000 2500 2000 1000 750 1500 1500 1000 500 500 250 1000 500 0 0-10 10-30 30-70 70-110 110-200 Biomass Class (Mg ha -1 ) > 200 0 0-10 10-30 30-70 70-110 110-200 Biomass Class (Mg ha -1 ) > 200 0 0-10 10-30 30-70 70-110 110-200 Biomass Class (Mg ha -1 ) > 200 0-10 10-30 30-70 70-110 110-200 > 200 Biomass Class (tons/hectare)
Tibati Ngaundal Roads increasing human population deforestation/degradation reduction in human population reduction in burning woody encroachment Mbam Djerem National Park Railway line
Optical vs. Radar: less sensitivity over forest Δ NDVI 2000 2006 Change detection: Landsat 2000 NDVI with cross-calibrated ASTER 2006 NDVI Change detection: JERS 1996 with ALOS 2007
Future work Add Landsat MSS 1973 and Spot Vegetation 1995 Fieldwork in Uganda & Tanzania; data from Gabon, Burkina Faso, Mali, Ghana, Ivory Coast, Central African Republic & Congo Scale up to the whole of tropical Africa using AVHRR and MODIS time series REDD; UNFCCC; Prince s Rainforest Trust MSS 1973 TM 1986 SPOT 1995 ETM+ 2000 ASTER 2006 JERS-1 1996 ALOS 2007
TROBIT (TROpical Biomes In Transition, NERC-funded consortium project) Edward.mitchard@ed.ac.uk Any questions? blications: - Mitchard et al., Measuring biomass changes from woody encroachment and restation in the forest-savanna boundary region of central Africa from multi-temporal L-band r backscatter, in review, JGR Biogeosciences - Mitchard et al., Detecting Woody Encroachment from 1982-2006 along a Forestnna Boundary in Central Cameroon, in review, Earth Interactions - Freedman, A.H., Smith, T.B., Mitchard, E.T.A., Saatchi, S.S. & Buermann, W. Loss of otypic variation due to human alteration of the rainforest-savanna gradient in West Africa, in w, Science
Spectral unmixing upscaling LOCAL SCALE (hundreds of km 2 ) Spectral data Texture Object-orientated classification Field studies Quickbird (0.6 m) 2007/8 Spectral unmixing rules producing 15m gridded % tree cover product for training REGIONAL SCALE (thousands of km 2 ) Spectral data Field studies ASTER (15 m) 2007 Spectral unmixing rules producing 250 m gridded % tree cover product for training CONTINENTAL SCALE (millions of km 2 ) Spectral data Seasonality MODIS (250/500 m) 2000-2007 Aerial photographs (1-4 m) 40s 90s ALOS 12.5 m 2007 JERS-1 100 m 1996 Landsat (30-60 m) 1972-2003 ERS c- band 1992-2008 JERS-1 100 m 1996 AVHRR (1 km/8 km) 1982-2008