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-poor
agricultural landscapes because of domain discrepancies and limited reference data. Traditional
vegetation index methods and supervised machine-learning approaches often lack robustness
for early-season prediction, while conventional unsupervised domain adaptation (UDA)
typically requires access to source-domain data, increasing computational and data-sharing
burdens. In this study, we introduce a source-free UDA framework, Contrastive Representation
Optimized Prototype Segmentation (CROPS), for large-scale, early-season soybean mapping
without relying on source data. CROPS utilizes NDVI-Max composites from the
QualityMosaic method to emphasize peak vegetation signals, reduce noise and redundancy, and
simplify preprocessing. A pixel-wise entropy partitioning strategy identifies high- and low-
confidence regions, enabling curriculum-based optimization within a teacher-student
architecture enhanced by Exponential Moving Average (EMA). Extensive experiments across
the USA, China, Brazil, and Argentina demonstrate that CROPS consistently surpasses
traditional indices, supervised classifiers, and established UDA methods. At the end of the
season, CROPS achieved average macro F1 scores exceeding 92%, closely matching official
agricultural statistics. Importantly, in South America, CROPS enables reliable early-season
mapping, with macro F1 above 90% in Brazil by mid-January and over 80% in Argentina by late
January. In the US Midwest, where the spectral similarity between soybean and maize makes
accurate classification particularly challenging during the growing season, CROPS achieves
robust in-season mapping for both crops. Ablation experiments reveal that this strong
performance is primarily attributed to the NDVI-Max composites’ ability to capture key
phenological features and to the progressive self-adaptive learning process, in which high-
confidence target-domain samples iteratively guide the low-confidence ones. This strategy
avoids negative transfer from source data and enhances adaptation to local characteristics.
These findings underscore the potential of CROPS as a timely, accurate, and scalable solution
for crop mapping in complex, data-limited environments.