Demeter

Proposed solution : Our project, Demeter, aims to accurately classify crop types using a combination of Sentinel-2 multispectral and SAR observations, fine-tuned for the unique agricultural landscape of Northern India. We would also take other relevant data from Sentinel-1/EOS-4, ResourceSAT-2/2A satellites. We propose a semantic segmentation-based approach (computer looks at every single pixel and decides its class) using a U-Net architecture enhanced for 12-band Sentinel-2 data and supported by temporal composites. Each satellite chip (256×256 pixels) contains pixel-wise crop labels and field IDs, making this problem naturally suited for segmentation rather than bounding-box detection. Our workflow involves: Data preprocessing: stacking multispectral bands, normalizing, and creating seasonal composites (early, peak, late growth). Modeling: training a U-Net with a 12-channel input to produce per-pixel crop classifications, supported by class-weighted cross-entropy and Dice loss to handle class imbalance. Post-processing: aggregating predictions at the field level using majority voting, then generating bounding boxes or polygons for submission. Evaluation: mean Intersection-over-Union (mIoU) and field-level macro-F1, with uncertainty estimation for low-confidence predictions. This approach is robust to fragmented fields, intercropping, and seasonal variation. It also supports extension to foundation models (e.g., Prithvi, SegFormer) for improved generalization if computational resources allow. By leveraging both optical and SAR signals, our method can mitigate cloud cover challenges and improve predictions in complex cropping scenarios such as Zaid season or underground crops. Our approach is perfectly tailored for India’s complex agriculture: it classifies every pixel using both optical and radar satellites, tracks crop growth over time, and aggregates predictions at field level , ensuring unmatched accuracy and robustness even under clouds, small plots, and diverse crop cycles.