Estimating smallholder crops production at village level from Sentinel-2 time series in Mali's cotton belt

In Mali's cotton belt, spatial variability in management practices, soil fertility and rainfall strongly impact crop productivity and the livelihoods of smallholder farmers. To identify crop growth conditions and hence improve food security, accurate assessment of local crop production is key. However, production estimates in heterogeneous smallholder farming systems often rely on labor-intensive surveys that are not easily scalable, nor exhaustive. Recent advances in high-resolution earth observation (EO) open up new possibilities to work in heterogeneous smallholder systems. This paper develops a method to estimate individual crop production at farm-to-community scales using high-resolution Sentinel-2 time series and ground data in the commune of Koningue, Mali. Our estimation of agricultural production relies on (i) a supervised, pixel-based crop type classification inside an existing cropland mask, (ii) a comparison of yield estimators based on spectral indices and derived leaf area index (LAI), and (iii) a Monte Carlo approach combining the resulting unbiased crop area estimate and the uncertainty on the associated yield estimate. Results show that crop types can be mapped from Sentinel-2 data with 80% overall accuracy (OA), with best performances observed for cotton (Fscore 94%), maize (88%) and millet (83%), while peanut (71%) and sorghum (46%) achieve less. Incorporation of parcel limits extracted from very high-resolution imagery is shown to increase OA to 85%. Obtained through inverse radiative transfer modeling, Sen2-Agri estimates of LAI achieve better prediction of final grain yield than various vegetation indices, reaching R2 of 0.68, 0.62, 0.8 and 0.48 for cotton, maize, millet and sorghum respectively. The uncertainty of Monte Carlo production estimates does not exceed 0.3% of the total production for each crop type.