A Source-Free Unsupervised Domain Adaptation Framework for Large-scale, in-season Soybean Mapping

Achieving reliable in-season soybean maps is challenging in heterogeneous and data-pooragricultural landscapes because of domain discrepancies and limited reference data. Traditionalvegetation index methods and supervised machine-learning approaches often lack robustnessfor early-season prediction, while conventional unsupervised domain adaptation (UDA)typically requires access to source-domain data, increasing computational and data-sharingburdens. In this study, we introduce a source-free UDA framework, Contrastive RepresentationOptimized Prototype Segmentation (CROPS), for large-scale, early-season soybean mappingwithout relying on source data. CROPS utilizes NDVI-Max composites from theQualityMosaic method to emphasize peak vegetation signals, reduce noise and redundancy, andsimplify preprocessing. A pixel-wise entropy partitioning strategy identifies high- and low-confidence regions, enabling curriculum-based optimization within a teacher-studentarchitecture enhanced by Exponential Moving Average (EMA). Extensive experiments acrossthe USA, China, Brazil, and Argentina demonstrate that CROPS consistently surpassestraditional indices, supervised classifiers, and established UDA methods. At the end of theseason, CROPS achieved average macro F1 scores exceeding 92%, closely matching officialagricultural statistics. Importantly, in South America, CROPS enables reliable early-seasonmapping, with macro F1 above 90% in Brazil by mid-January and over 80% in Argentina by lateJanuary. In the US Midwest, where the spectral similarity between soybean and maize makesaccurate classification particularly challenging during the growing season, CROPS achievesrobust in-season mapping for both crops. Ablation experiments reveal that this strongperformance is primarily attributed to the NDVI-Max composites’ ability to capture keyphenological features and to the progressive self-adaptive learning process, in which high-confidence target-domain samples iteratively guide the low-confidence ones. This strategyavoids negative transfer from source data and enhances adaptation to local characteristics.These findings underscore the potential of CROPS as a timely, accurate, and scalable solutionfor crop mapping in complex, data-limited environments.