Estimating crop yields by integrating the FAO Crop Specific Water Balance model with real-time satellite data and ground-based ancillary data

Abstract

The broad objective of this research was to develop a spatial model which provides both timely and quantitative regional maize yield estimates for real-time Early Warning Systems (EWS) by integrating satellite data with ground-based ancillary data. The Food and Agriculture Organization (FAO) Crop Specific Water Balance (CSWB) model was modified by using the real-time spatial data that include: dekad (ten-day) estimated rainfall (RFE) and Normalized Difference Vegetation Index (NDVI) composites derived from the METEOSAT and NOAA-AVHRR satellites, respectively; ground-based dekad potential evapo-transpiration (PET) data and seasonal estimated area-planted data provided by the Government of Kenya (GoK). A Geographical Information System (GIS) software was utilized to: drive the crop yield model; manage the spatial and temporal variability of the satellite images; interpolate between ground-based potential evapo-transpiration and rainfall measurements; and import ancillary data such as soil maps, administrative boundaries, etc. In addition, agro-ecological zones, length of growing season, and crop production functions, as defined by the FAO, were utilized to estimate quantitative maize yields. The GIS-based CSWB model was developed for three different resolutions: agro-ecological zone (AEZ) polygons; 7.6-kilometer pixels; and 1.1-kilometer pixels. The model was validated by comparing model production estimates from archived satellite and agro-meteorological data to historical district maize production reports from two Kenya government agencies, the Ministry of Agriculture (MoA) and the Department of Resource Surveys and Remote Sensing (DRSRS). For the AEZ analysis, comparison of model district maize production results and district maize production estimates from the MoA (1989-1997) and the DRSRS (1989-1993) revealed correlation coefficients of 0.94 and 0.93, respectively. The comparison for the 7.6-kilometer analysis showed correlation coefficients of 0.95 and 0.94, respectively. Comparison of results from the 1.1-kilometer model with district maize production data from the MoA (1993-1997) gave a correlation coefficient of 0.94. These results indicate the 7.6-kilometer pixel-by-pixel analysis is the most favorable method. Recommendations to improve the model are finer resolution images for area planted, soil moisture storage, and RFE maps; and measuring the actual length of growing season from a satellite-derived Growing Degree Day product.