Prediction of crop yield in Sweden based on mesoscale meteorological analysis

Transcription

1 Meteorol. Appl. 7, (2000) Prediction of crop yield in Sweden based on mesoscale meteorological analysis Valentin L Foltescu, Environmental Physics, Chalmers University of Technology, Göteborg, Sweden (Permanent affiliation: Swedish Meteorological and Hydrological Institute, Norrköping, Sweden) This paper presents a prediction system for regional crop growth in Sweden, recently set up at SMHI (Swedish Meteorological and Hydrological Institute). The system includes a state-of-the-art crop growth model, WOFOST (WOrld FOod STudies) and inputs from meteorological mesoscale analysis. The simulated crops are spring barley, spring rape, oats and winter wheat, and the period of investigation is The simulated water-limited grain yield is used as a predictor in the yield prediction procedure. The technological time trend describing the yearly increase of the production level is accounted for as well. Yield prediction based on crop growth modelling is justified since the ability to forecast the yield is higher compared to that using the technological time trend alone. The prediction errors are of the order of 8 to 16%, with the lowest errors for winter wheat and spring barley. 1. Introduction Recent developments of mesoscale analysis systems result in higher spatial resolution and better accuracy. The analysis can be performed in real time, consequently allowing a timely access to its products (i.e. gridded information on a variety of meteorological variables). Multivariate observational data from synoptic stations, automatic stations, satellites and weather radars can be integrated into the mesoscale analysis scheme. In addition, weather forecast fields from limited area models may be used as first guess initial fields. Considering these latest developments, it should be advisable to use gridded meteorological information in the context of crop growth modelling, in order to increase the accuracy of crop yield calculations. The aim of the present study was to design and assemble a prediction system for regional crop growth in Sweden, including a state-of-the-art crop growth model and inputs from the mesoscale analysis. The use of gridded fields from the mesoscale analysis is expected to increase the accuracy of the crop yield calculations. The values in the gridded fields are more realistically used for crop growth simulations. The values no longer represent point data. They describe the average conditions prevalent in the grid cells during the time step. Model-based crop yield estimation and prediction has gained recognition in recent years as a cost-efficient and complementary method to statistical field sampling and remote sensing. The MARS (Monitoring Agriculture with Remote Sensing) project of the European Commission has devoted a special activity aimed at developing and using agro-meteorological models at the European level (Vossen, 1996). The present paper describes the newly established prediction system in Sweden, and its performance. The predicted yields are for winter wheat, spring barley, spring rape and oats in two agricultural regions of high importance for agricultural grain production in Sweden, namely Skåne and Skaraborg. Skåne and Skaraborg together produce about 40% of Sweden s grain production. About two-thirds of the total land area in Skåne and one-third that in Skaraborg is arable. The growing season in Skåne is more than 200 days while for Skaraborg it is limited to days. Figure 1 shows the two counties for which the model is set up in southern Sweden and their associated crop yield districts for which official statistics are collected. 2. Model-system outlook 2.1. The crop growth model and its meteorological input The Swedish system for crop growth predictions is built around the Crop Growth Monitoring System (CGMS), developed at the JRC (Joint Research Centre). Implicitly, the engine of the system is the WOFOST (WOrld FOod STudies) agro-meteorological model for crop growth. WOFOST is a complex mechanistic model that simulates biomass accumulation in combination with phenological development, on the basis of underlying 313

2 V L Foltescu partitioned among the various plant organs. The partitioning factors vary with the development stage of the crop. The simulation of phenological development and biomass formation and accumulation also takes account of other processes, such as maintenance of the existing living biomass (maintenance respiration), senescence of leaves, transpiration and extension or roots. WOFOST keeps track of the soil moisture content in order to determine if a crop is exposed to water stress. Production is diminished if the availability of soil moisture is reduced. In the present implementation, the model expects non-limiting plant nutrient status and it does not take into account effects of fertilisers and pesticides. The Swedish application is run on an 11 km rotated latitude longitude grid, and it requires daily grid weather. More than two years of gridded information (from 1997 and onwards) have been processed using MESAN (MESoscale ANalysis system) at the SMHI (Swedish Meteorological and Hydrological Institute). An overview of MESAN is given by Häggmark et al. (1997). Making yield predictions necessitates many years of simulations. Since MESAN was not available before 1997, it was decided to use station data for the years and apply the spatial interpolation routine supplied with the CGMS itself (approach based on a study carried out by van der Voet et al., 1994). Figure 1. Location of Skaraborg (top) and Skåne in Southern Sweden. Thick lines indicate the extent of the two counties used for agro-meteorological modelling in the present system set-up. Thin lines indicate the borders of the yield districts. Shaded areas represent lakes. processes (e.g. photosynthesis and respiration) and how these processes are influenced by environmental conditions (including those of weather and soil moisture). Details regarding the theoretical background of the model can be found elsewhere (e.g. van Diepen et al., 1989; Supit et al., 1994). The daily meteorological variables used as inputs to WOFOST are maximum temperature, minimum temperature, global radiation, wind speed, vapour pressure, precipitation and evapo-transpiration. For the calculation of the last variable the Penman method is used. The time-step in the model is one day and the Euler integration method is used. Crop growth depends on the level of incoming radiation, the crop leaf area and the photosynthetic characteristics of the leaves. With these inputs the daily rate of CO 2 assimilation is calculated. Part of the carbohydrates produced is converted into biomass, which is GIS layers and soil parameters The Swedish system for crop growth prediction is linked to the MESAN grid, the areas for the administrative regions (for which statistical surveys are produced) and a soil map from the Swedish Geological Survey. The forested areas are filtered out from the soil map. The linkage is provided by GIS (Geographic Information System). The three different GIS layers mentioned above were intersected in order to create the elementary mapping units (EMUs), for which the simulations are performed. The intersection gave rise to 785 EMUs for the 157 grid cells in Skåne. In Skaraborg, the numbers are 479 and 102 respectively. The large number of simulation units is attributed to the detailed soil map, which has a direct impact on the size of the system. The scale of the map is 1: and it contains six soil types: peat, clay-silt, silt-sand-gravel, glacial sediment, clayey till mixed with till, and moraine. In the absence of measured soil water retension curves it was decided to use soil texture for setting up the soil mapping units in the model. The soil water behaviour is mainly governed by soil texture. However, the soil

3 classes were rather wide (in terms of texture), making it difficult to use the pedotransfer functions from literature for estimating the hydraulic properties (e.g. waterholding capacity). Instead, the soil parameters were retrieved from the classes of the European Community s soil map (Reinds et al., 1992), slightly modified as a result of sensitivity analyses in a few areas, homogeneous with respect to soil type Crop parameters and calendar Prediction of crop yield based on mesoscale meteorological analysis (a) The temperature analysis scheme The observation systems involved in the temperature analysis consist of synoptic (SYNOP) stations, automatic stations and climatic stations. The vertical profile of temperature is retrieved from HIRLAM and used for interpolation of the first guess to the analysis grid. The autocorrelation functions are not isotropic. They are corrected with respect to the difference in elevation of the stations and also take account of the land sea fraction. The corrections are empirical linear functions. Historical information on sowing dates was obtained from SCB (Statistiska centralbyrån, Statistics Sweden) and mean sowing dates were calculated. Crop parameters were adopted as follows. For winter wheat and spring rape the built-in parameter file of CGMS was used. For spring barley values from Fagerberg et al. (1998) were used. Oats was parameterised with independent data from SLU (Sveriges lantbruksuniversitet, Swedish University of Agricultural Sciences), used in Carlsson et al. (1993) Spatial interpolation of meteorological data The system requires MESAN input of the following gridded daily meteorological variables: maximum temperature, minimum temperature, vapour pressure, wind speed, precipitation and total cloud cover. In the MESAN-approach, these variables or similar ones (from which the above variables can be calculated) are analysed by employing multivariate statistical (optimal) interpolation. The optimal interpolation (OI) technique has been widely used in meteorological applications and will not be dealt with in detail here (see, for example, Daley, 1991). The interpolation is based on the minimum-mean-squared-error of the analysis. A key aspect of the OI technique is the use of so-called autocorrelation functions that provide important information on how the variable being interpolated is correlated in space. In the development of MESAN much effort has been devoted to modelling these functions using a large body of historical data (Häggmark et al., 1999). The interpolation employs the derived predetermined autocorrelation functions. Observational data from synoptic and climate stations as well as satellites and weather radars are assimilated into MESAN. Weather forecast fields from HIRLAM (High Resolution Limited Area Model) are used as first-guess fields. Quality control is performed before submission to OI in MESAN. Thereby large or systematic errors can be identified and eliminated. The following subsections outline some important aspects of the analysis of three significant variables: temperature, cloudiness and precipitation. More detail on MESAN s interpolation schemes can be found in Häggmark et al. (1999). (b) The cloud analysis scheme The observation systems involved in the cloud analysis consist of SYNOP-stations, METAR (airport data), automatic stations and satellites. The polar NOAA satellite is used together with the SCANDIA classification scheme (Karlsson, 1996). The geostationary METEOSAT is used as well. The NOAA satellite provides high spatial resolution and extensive spectral information but coarse temporal resolution. METEOSAT provides high temporal resolution instead. Observations within ±1 hour from the analysis time are being considered. In addition to the cloudiness reporting stations, the difference between the temperature at the 2 m level and the brightness temperature (which is the effective black body radiation temperature retrieved from satellites) is used as an indicator for the presence/absence of clouds. This is supplemented by HIRLAM information on stratification, relating it to the chance of encountering clouds for well-defined stability criteria. (c) The precipitation analysis scheme Observational data from SYNOP and climatic stations as well as weather radars are assimilated into MESAN. Precipitation can be highly variable, both temporally and spatially. The autocorrelation functions vary with accumulation time (3, 12 and 24 hours), and are steepest for the 3-hour analysis. Orographic effects on precipitation (such as topographic enhancement and enhancement by coastal convergence) are indirectly taken into account. If the wind direction is known, a statistical model is used to create a field using standard deviations of precipitation. The model is based on a climatological regression analysis in which the following are used as predictors: Frequency of wind directions multiplied by the upslope gradient of topography. Component of the roughness length perpendicular to wind direction. Latitude. The result is a field that shows the climate of the current weather situation. If the wind direction is not known, the climatological field of the standard 315

4 V L Foltescu deviations is used. Both observations and the firstguess field are normalised with respect to the standard deviation, and the interpolation is performed in the normalised variable. The normalisation is inverted at the end of the analysis. Isotropic analysis in the normalised variable is possible due to the fact that the annual precipitation, normalised by the standard deviations of the daily precipitation, is nearly constant, independent of station. For the period when MESAN was not available, i.e. before 1997, the CGMS-approach was used for the interpolation. The basis of the analysis is the selection of suitable meteorological stations for determination of the representative meteorological conditions in a grid cell. A score is calculated for each station taking into account the distance to the grid centres, the difference in altitude and the distance to coast. Once this selection is made, a grid-cell average is taken for most of the meteorological variables, with the exception of precipitation. The precipitation for a grid cell is taken from the weather station that is most similar to the grid centre (in terms of the above-mentioned score). Details of the method can be found in van der Goot (1998) Calculation of global radiation Global radiation is a controlling factor in the crop growth process. It is however measured at only 12 sites, sparsely spread over Sweden. The number of sites is too small, and the stations are situated too far apart for performing a reliable interpolation. Instead, global radiation (expressed in J m 2 d 1 ) is calculated using the Supit formula (see equation (1)) if the total cloud cover is available (Supit, 1994). The calculation of global radiation (R gs ) is based on the extra-terrestrial Angot radiation (R a in J m 2 d 1 ), minimum and maximum temperature (T min and T max in ºC), the total cloud cover (CC in oktas) and a set of three coefficients: A s and B s (both dimensionless) and C s (in J m 2 d 1 ). The three coefficients are empirical regression constants and depend on the geographical location. They have been established for five regions in Europe (Supit, 1994). Rgs = Ra As T T + Bs CC C (1) + max min 1 / 8 s When the total cloud cover was not available, the Hargreaves formula (from Supit, 1994) was used to calculate global radiation (R gh ): where A h and B h (both dimensionless) are the Hargreaves coefficients. The two coefficients used in equation (2) are empirical regression constants established by Supit (1994) for six regions in Europe. 316 R = R A T T + B gh a h max min h (2) Since global radiation is crucial to crop growth, it is important to assess the quality of the calculated radiation. A grid cell in Skåne was selected corresponding to a location (Lund, 55.7ºN, 13.2ºE) where global radiation was measured. A whole year, 1998, was investigated on a daily basis. The correlation between the calculated and the observed radiation was high: 0.93 using the Supit formula and 0.82 using the Hargreaves formula Sensitivity of the model The simulations were preceded by a variety of sensitivity studies of which two studies are presented briefly. A sensitivity study for homogeneous soils over the simulation units gave insight into the importance of the rooting depth parameter of the soils. The model sensitivity was higher with respect to it than towards the soil classes (mainly determined by texture). For simulations it is was concluded that it would be important to specify rooting depth very accurately. During the sensitivity analysis another important fact was revealed, which relates to the importance of using meteorological fields with high spatial resolution. A transect was taken over a region with a climatologically high gradient of precipitation (Northern Skåne). The yield calculated in grid cells along the transect could increase by 52% over a short distance of 15 km, as a result of a 35% increase in the yearly precipitation. Therefore, it is expected that the use of the MESANanalysed meteorological variables, interpolated on a high-resolution grid (as in the present case), will allow us to better resolve yield variations over short distances. 3. Crop growth simulations The crop yield variables calculated by the model are biomass and grain production under actual precipitation conditions (water limited situation) and under conditions of optimum moisture availability (potential situation). Spring barley was one of the simulated crops. Simulated yearly growth curves of spring barley are shown in Figure 2 for the water limited case. The interannual variation is significant. However, since the simulation is stopped at maturity, the year-to-year variations in the harvest time are implicitly taken into account. Figure 3 depicts the simulated dry matter grain yields (water limited) of spring barley and the corresponding dry yields surveyed and reported by SCB. The results illustrate how the model is able to mimic a large part of the patterns in the time trend of the survey. The mean grain yields in each county are presented with a continuous line in Figure 3. The dashed lines

5 Prediction of crop yield based on mesoscale meteorological analysis in the historical series of the surveyed yield and b, c and d are regression constants determined from 14 years of survey statistics and simulations. Statistics on the yields per hectare are based on surveys carried out by SCB. Farms all over the country are selected by stratified random sampling. SCB made regular surveys (objective crop yield surveys) of crop cutting from sample plots prior to After 1995 interviews with farmers became the basic method for data collection, but the objective crop yield surveys were still carried out in certain yield-districts as a reference. From 1998 the yield statistics for cereals are solely based on interviews with farmers (Hagblad, 1998). Figure 2. Simulated spring barley (dry matter grain yield) in Skaraborg. Ten-day periods (decades) are numbered from 1 January. illustrate the variability (within two standard deviations) with respect to the average yield, calculated for the previously reported amount of simulation units. The variability in the survey results is among the reporting crop yield districts for the official statistics. 4. Grain yield estimation The simulated grain yields are not to be compared directly with the surveyed grain yields, because the latter are often considerably lower than the water-limited grain yields. This is due to sub-optimal cultivation practices, losses during harvesting, pests, weeds, etc. The simulated grain yields represent optimum production, which can be reached under optimum conditions of nutrient supply, control of weed, plant pests and diseases, as well as optimum farm management. Note also that the simulations do not take into account the introduction of new grain varieties or other improved agricultural practices introduced after tuning of the model. The concept of yield estimate is inferred and is calculated based on a model reported by Koning et al. (1993). The water-limited grain yield (Y w ) has been used as a predictor. The basis of the prediction model is the annual variation in the simulation results as opposed to the annual variation in the surveyed yields. The technological time trend describing the yearly increase of the production level (due to gradually improved agricultural practices) is accounted for as well. The estimated yield is calculated with the following expression: Y = Y + b ( t t)+ c ( Y Y )+ d (3) s i w w where Y is the estimated yield, Y s is the surveyed yield, Y w is the water limited grain yield. t i is the year number Sweden is divided into 106 yield districts out of which 17 are in Skåne and 10 in Skaraborg (Figure 1). For a separate yield district the standard error for a survey estimate could be rather high, especially since no cutoff is used so that yield districts with few observations in a certain year would be omitted. As an example at the level of the individual districts, the standard errors in Skåne and Skaraborg in 1998 varied between 2 and 19% for winter wheat, 2 and 22% for barley, 3 and 34% for oats, and 3 and 31% for spring rape. Standard errors in the lower parts of the intervals were dominant though. On a county level, the standard errors in both Skåne and Skaraborg were of the order of 5% for all crops of interest to this study. Figure 4 shows the discrepancy between the estimated and the surveyed grain yields. The surveyed grain yields are taken as the reference, although they are also marred by deviation from the true yields, as just stated. The median relative deviations between estimates and surveys are of the order of 6 10%. The best estimates are obtained for winter wheat and spring barley. Notably high deviations can be identified at certain times for some crops. Both the simulation results and the estimates are much in excess over the surveyed yields in 1992, which is regarded as a problematic year for all of the investigated crops except for winter wheat. The beginning of summer in 1992 was extremely dry. Less than 10% of the normal monthly precipitation was registered in June Although the model simulated the drought situation through a significant drop in yield, the yield decrease according to simulation was not too significant. Figure 5 shows the simulated relative soil moisture under a barley crop in Skåne calculated as a surface average over the county and displayed for 1992 alone and as a long-term mean ( ). The relative soil moisture is defined as the percentage of the difference between field capacity and wilting point. The exceptionally dry summer of 1992 and its repercussion on the soil moisture emerged distinctly from the model output. Wheat appears to have responded better to drought conditions, possibly due to the way the model describes uptake of water from the soil as a function of 317

6 V L Foltescu Figure 3. Time series of the simulated and surveyed dry matter grain yields of spring barley in Skåne and Skaraborg. the crop s rooting system. Winter wheat has a developed root system that promotes water uptake from deep soil layers in periods of drought. A limiting factor is winter damage and an example of this is the poorer harvest of 1987 in Skåne after a cold winter. By comparison with the other investigated crops, the results for spring rape and oats are of poorer quality. This might indicate that the crop parameters require further calibration and testing. However, it may also signal that oats and spring rape are more sensitive to different pests. In 1995 severe attacks of plant pests reduced the rape grain yield. Different fungal pests attacked spring rape in Skåne in 1991 causing some 30% yield reduction. In 1992 both Skaraborg and Skåne suffered from attacks of frit fly in oats. This came on top of the drought that year. Spring rape is also sensitive to wet soils in the spring. At our northerly moist latitudes this situation occurs from time to time. The spring of 1991 in Skaraborg was reported as very wet causing water damage to the rape fields. 5. Grain yield prediction The comparison between estimated yields and surveys serves as an indicator of the capability of the modelling effort to calculate yields. It allows us to identify the outlier situations described previously, with discrepancies resulting from either extreme weather conditions 318 (causing crop injuries) or due to harvest-related problems or pest attacks. Such effects are clearly not taken into account by the model since the model is not conceived to deal with those sort of stresses. The estimated grain yields, however, should not be confused with the predicted grain yields. The predicted yields for the same series of years and for subsequent years are calculated with a similar formula as the estimated yields but with a different set of the regression coefficients b, c, d, determined separately for each year. The leave-one-out method employed resembles a cross-validation approach. First of all, years previously identified as outliers (that is, when the model cannot be made responsible for the discrepancy) are discarded from the time series. Secondly, the yields of the year subject to prediction are kept out of sight in the calculation of means, and the prediction is based solely on the remaining years. The procedure is repeated for each year in turn. A prediction error can be formulated by following an example from Wallach & Goffinet (1989). The RRMS (Relative Root Mean Square) prediction error is calculated as a percentage of the mean surveyed yield: RRMS( e1 K, ek )= 1 k i Y i= k = 1 s e i (4)

7 Prediction of crop yield based on mesoscale meteorological analysis Figure 4. Relative difference (%) between the estimated and the surveyed dry matter grain yields of winter wheat, spring barley, spring rape and oats in Skåne and Skaraborg between 1985 and in which k is the number of prediction years, Y s is the mean surveyed yield of the predicted years and e i is the error of the ith prediction. The RRMS error given by equation (4) indicates what accuracy to expect when making predictions of future yields, based on the present system set-up and the historical trends (i.e. predictions of the past). Table 1 summarises the results. An attempt was made to evaluate the gain of making predictions based on model simulations and trends, 319

8 V L Foltescu Table 2. Squared Pearson correlation coefficient (R 2 ) emerging from the regression between predictions and surveys. Predictions are either based on crop growth modelling (last column) or are calculated as an extrapolation of the technological trend inferred from the surveys. Figure 5. Simulated relative soil moisture status under a spring barley crop in Skåne 1992 and during the period (indicated as the period mean). Ten-day periods (decades) are numbered from 1 January. using equation (3), as opposed to considering the simple technological trend alone, derived from the survey results as: In equation (5) Y s is the mean surveyed yield of the predicted years, b is the yearly increase of the surveyed yield and (t i t) is the time change from the ith prediction year to the reference time t taken as the mean of the predicted years. Table 2 reveals that the explained variability (quantified as R 2, i.e. the squared Pearson correlation coefficient) increases for all the investigated crops in both Skåne and Skaraborg when using the modelling-based prediction. This proves that the modelling effort is justified. 6. Concluding remarks The following conclusions can be drawn from this study. (a) Yield prediction based on crop growth modelling is justified since the ability to forecast the yield is 320 Y = Y + b t t s ( i ) Table 1. Relative Root Mean Square Error (%) for prediction of spring barley, spring rape, oats and winter wheat in Skåne and Skaraborg. County Relative Root Mean Square Error (%) Spring Spring Oats Winter barley rape wheat Skåne Skaraborg (5) County/Crop higher compared to using only the historical trend in the surveys (Table 2). (b) The prediction errors (RRMS errors) are of the order of 8 16%, with the lowest errors for winter wheat and spring barley in Skåne. (c) Notably high discrepancies between predicted and surveyed grain yield can occur at times for any of the four crops. These may be the effect of extreme weather conditions resulting in crop injuries, but also harvest-related problems or pests attacks. Such effects are not taken into account by the model. (d) Discrepancies may appear also due to the year-toyear variations in the sowing time. A yearly differentiated crop calendar is not implemented in the system. (e) It would be desirable to implement the information on actual land use for cultivation of the various crops in order to determine exactly which areas should undergo simulation for a particular crop. (f) Yearly-based crop calendars, measurement of the exact extent of the cultivated areas for each crop and a soil map with narrow soil texture classes might result in improvement of the predictions. (g) The continuing use of gridded meteorological data in years to come will increase the time series of MESAN-based simulations and will eventually enable calculation of a new set of regression constants in the prediction model. The predictions will then no longer suffer from the lack of homogeneity in the time series caused by using station data before Acknowledgements Survey-based predictions Correlation Coefficient Modelling-based predictions Skåne Spring barley *0.39* **0.48** Spring rape Oats Winter wheat 0.15 *0.34* Skaraborg Spring barley 0.18 *0.28* Spring rape *0.29* *0.30* Oats 0.11 *0.33* Winter wheat * Correlation is significant at the 0.05 level ** Correlation is significant at the 0.01 level The author wishes to acknowledge several institutes for making the realisation of this project possible. SMHI

9 Prediction of crop yield based on mesoscale meteorological analysis for providing its infrastructure and GIS resources. SAI (Space Application Institute) of the JRC for providing the CGMS software. SLU and particularly the Department of Ecology and Crop Production Science, for assisting in the creation of a crop-knowledge database. SCB for providing the historical survey-based statistics on the yields. The author would like to thank several collaborators for the feedback and support, in particular: B. Dahlström, A. Gyllander, E. van der Goot, P. Nyman and G. Wahlstedt. Special thanks also go to all the dedicated persons involved in the development of the WOFOST and CGMS over the past two decades. References Carlsson, H., Larsson, S. & Magnét, B. (1993). Stråsäd Trindsäd Oljeväxter, Speciella Skrifte,r 49, Sveriges lantbruksuniversitet Uppsala (Swed. Univ. of Agric. Sci.). Daley, R. (1991). Atmospheric Data Analysis. Cambridge University Press. Fagerberg, B., Nyman, P., Sigvald, R. & Torssell, B. (1998). Estimating regional crop yields in Sweden with a simulation model and field observations. Swedish J. Agric. Res., 28: Hagblad, L. (1998). Crop cutting versus farmer reports review of Swedish findings. Environment and Regional Statistics. MR/LP 1998:2, Statistics Sweden, Örebro. Häggmark, L., Ivarsson, K-I. & Olofsson, P-O. (1997). MESAN. Mesoskalig analys. RMK 75 Report (in Swedish), SMHI, Norrköping, Sweden. Häggmark, L., Ivarsson, K-I., Gollvik, S. & Olofsson, P-O. (1999). MESAN, an operational mesoscale analysis system. Tellus 25B: Karlsson, K-G. (1996). Validation of model cloudiness using satellite-estimated cloud climatologies. Tellus, 48A: Koning, G. H. J., Jansen, M. J. W., Boons-Prins, E. R., van Diepen, C. A. & Penning de Vries, F. W. T. (1993). Crop growth simulation and statistical validation for regional yield forecasting across the European Community. Simulation Reports CABO-TT 31, CABO-DLO, Wageningen Agricultural Universiy, The Netherlands. Reinds, G. J., de Koning, G. H. J. & Bulens, J. D. (1992). Crop production potential of rural areas within the European Communities. III: Soils, climate and administrative regions. Working Documents W67, Netherlands Scientific Council for Government Policy, The Hague. Supit, I. (1994). Global radiation. EUR 15745, Office for Official Publications of the European Communities, Luxembourg. Supit, I., Hooijer, A. A. & van Diepen, C. A. (1994). System description of the WOFOST 6.0 Crop Simulation Model implemented in CGMS. Volume 1: Theory and algorithms. EUR 15956, Office for Official Publications of the European Communities, Luxembourg. van Diepen, C. A., Wolf, J. and van Keulen, H. (1989). WOFOST: a simulation model of crop production. Soil Use and Management, 5: van der Goot, E. (1998). Spatial interpolation of daily meteorological data for the Crop Growth Monitoring System (CGMS). Data Spatial Distribution in Meteorology and Climatology. EUR 18472, Office for Official Publications of the European Communities, Luxembourg. van der Voet, P., van Diepen, C. A. & Oude Voshaar, J. (1994). Spatial interpolation of meteorological data. A knowledge-based procedure for the region of the European Communities. SC-DLO Report 53.3, DLO Winand Staring Centre, Wageningen, The Netherlands. Vossen, P. (1996). Early crop production assessment of the European union: The systems implemented by the MARS- STAT project. Agrometeorological Models: Theory and applications in the MARS project. Proceedings. EUR 16008, Office for Official Publications of the European Communities, Luxembourg. Wallach, D. & Goffinet, B. (1989). Mean squared error of prediction as criterion for evaluating and comparing system models. Ecological Modelling, 44: