Can Sentinel-2 imagery support the identification of legume fields in maize-dominated smallholder farming systems?

Identifying trends in crop diversification is critical for assessing progress toward sustainable intensification in smallholder farming systems. This study investigates the potential of remote-sensing data for identifying common bean and groundnut in a maize-dominated agricultural landscape in the Southern Highlands of Tanzania. Using 5-day interval Sentinel-2 images, we built season-adjusted and phenology-based predictors to classify crops, addressing the challenges posed by the variability in the onset of the rainy season, wide planting windows, and diverse cropping practices. Field-level Enhanced Vegetation Index (EVI) profiles were extracted and analysed by detecting key phenological events, including vegetation emergence, peak, and maturity. Random forest models were used to identify crop types and their different cropping practices from field-level spectral bands, vegetation indices, and phenology-based metrics. In maize and common bean fields, we identified contrasting cropping practices, characterised by one or two successive vegetation peaks during the growing season. In fields characterised by one vegetation peak, producer accuracy for maize reached 86 % for maize, 53 % for bean, and 59 % for groundnut. The use of phenology-based predictors improved model transferability to untrained areas and seasons, compared to predictors based on vegetation indices tied to satellite acquisition dates. Still, legume identification accuracy remained too low for mapping. Improving legume identification in smallholder farming systems hinges upon an understanding of local cropping practices, along with the development of predictors applicable across heterogeneous agroclimatic conditions.