Modeling crop yield in West-African rainfed agriculture using global and local spatial regression

Performance of crop yield models is generally evaluated without testing their ability to capture yield spatial variability across a large area. Local soil and environmental conditions or management factors usually cause significant crop yield variability. In West Africa, landscape heterogeneity and data scarcity pose yet additional challenges to crop yield modeling. In this study, conditional autoregression (CAR) and geographical weighted regression (GWR) were used to better understand spatial patterns in sorghum [Sorghum bicolor (L.) Moench.], pearl millet [Pennisetum glaucum (L.) R. Br.], and cotton (Gossypium hirsutum L.) yields in Burkina Faso. A series of SPOT satellite normalized difference vegetation index (NDVI) 1-km2 10-d composite images spanning the crop growing season and observations of rainfall, topography, soil properties, and labor availability were used as explanatory variables in the CAR and GWR models. Regression analyses revealed that crop yield was significantly related to rainfall and topography in the semiarid and subhumid agroecological zones of Burkina Faso. Soil properties and labor availability mainly affected sorghum and millet yields in the semiarid zone. By addressing spatial dependency between crop yield observations in the two zones, GWR outperformed CAR models. For CAR models, adjusted R2 (Ra2) values for the sorghum, millet, and cotton yields were 0.76, 0.70, and 0.50, respectively, for the semiarid zone and 0.54, 0.32, and 0.30, respectively, for the subhumid zone. For GWR models, Ra2 values were 0.85, 0.70, 0.78, respectively, for the semiarid zone and 0.76, 0.67, 0.65, respectively, for the subhumid zone. Thus, despite limited data availability, GWR can be used to model the spatial variability of crop yields across large areas in West Africa.