National-scale mapping of staple crops in Nepal with multi-year Sentinel-2 L2A imagery and binary transformer models

Accurate mapping of staple crops using satellite imagery is essential for sustainable agricultural management, particularly in data-limited regions. Despite recent advances in Earth observations and deep learning, using multi-year time series satellite data and binary deep learning models to produce national-scale multi-class crop type map remains under-explored. In this study, we attempt to map staple crops, other crops, and non-crop at 10m resolution that leverages multi-year, unsmoothed Sentinel-2A/B MSI Level-2A time series imagery, binary transformer models, and a probability-based thresholding method. Nepal was selected as the case study due to its complex topography and heterogeneous landscapes. We developed six experimental setups, and models achieved an average overall accuracy of 91.39%, with class-wise F1-scores ranging from 0.81 to 0.96. The best-performing model was then applied to 28 tiles covering Nepal, and a probability-based thresholding approach was used to assemble binary results into a multi-class map. The map overall accuracy was higher than 94%, when validated with field data. Sensitivity analyses further confirmed the robustness and spatial consistency of the proposed framework across diverse land-cover and agro-ecological zones. Finally, the produced national-scale maps indicated 3.91% increase in staple crops areas from 2021 to 2023 in Nepal.