Estimate yield at parcel level from S2 time serie in sub-Saharan smallholder farming systems

In Southern Mali, high yield variability caused by rainfall distribution, local soil fertility and farmer management impacts the vulnerability of smallholder farmers. To better understand yield variation on a yearly basis and in a food security context, it is necessary to adequately estimate the crop yield. As yield assessment requires accurate crop type maps, the first step of the present paper is to classify crop types with the 10-20 m resolution time series of Sentinel-2 satellite. A random forest classifier applied under the cropland area produces a crop type map with an overall accuracy of 85%. The best performances are met for cotton (Fscore 93.86%), maize (Fscore 87.52%) and millet (82.5%) while sorghum (Fscore 45.7%) and peanut (Fscore 71%) show lower results. In a second phase, the yield estimation is based on linear regressions with vegetation indices (VI) or Leaf Area Index (LAI). The best performance is met with the maximum LAI (obtained from Sentinel-2) accross the growing season and reaches an average R
²
of 0.61 for all crop types. This very simple proxy is ambitious for food security as crop yield map can be produced with high accuracy. Thus yield variation can easily be assessed and could help to better target agricultural interventions at the farm or village scale.