A source-free unsupervised domain adaptation framework for large-scale, in-season soybean mapping

Reliable in-season soybean mapping remains difficult in heterogeneous and data-poor agricultural regions because phenology, management, and cloud-driven sampling differences induce substantial domain shifts, causing models trained in one region to generalize poorly to new landscapes. Since collecting dense, region-specific labels every season is often impractical, unsupervised domain adaptation (UDA) offers a practical route for adapting models using unlabeled target imagery. However, standard UDA commonly assumes access to source-domain imagery during adaptation, which increases data-sharing, storage, and computational burdens and limits scalable deployment. We propose a source-free unsupervised domain adaptation framework, Contrastive Representation Optimized Prototype Segmentation (CROPS), to deliver large-area, in-season soybean maps using only a pretrained model and unlabeled target imagery. CROPS constructs NDVI-based QualityMosaic maximum-NDVI composites that anchor each pixel to a phenology-informed high-canopy observation, mitigate cloud-related sampling noise, and reduce temporal redundancy. It then builds a crop-specific easy/hard reliability partition from the NDVI-QM target patches, combining entropy-based confidence, semantic dominance, boundary credibility, and class-balanced quotas. Reliable easy patches define target-domain semantic prototypes, whereas hard patches are incorporated later through prototype-guided alignment and teacher-confidence-modulated supervision. Across six target sites spanning the USA, China, Brazil, and Argentina, CROPS consistently outperforms other baselines and UDA methods. In end-of-season evaluation, CROPS achieves an average macro-F1 of 92.44 percent with IoU of 79.73 percent, closely matches official statistics at administrative-unit scale, and remains close to an Oracle upper bound trained with the full target reference data. In-season analyses show that CROPS advances the operational high-accuracy milestone, macro-F1 at least 0.80, by about 35 to 41 days in strongly shifted regions and remains robust under persistent cloud contamination. Additional experiments in selected three-class settings demonstrate promising within-season separation of soybean, corn, and background and improved cross-sensor transfer from Sentinel-2 to Landsat 8. Overall, CROPS provides a timely, scalable, and privacy-compatible pathway for operational crop mapping under limited labels and restricted source-data access.