Unsupervised Cross-Regional and Cross-Year Adaptation by Climate Indicator Discrepancy for Crop Classification

作者：Hengbin Wang, Yu Yao, Junyi Liu, Xindan Zhang, Yuanyuan Zhao, S. H. Li, Zhe Liu, Xiaodong Zhang, Yelu Zeng · 发表于：Journal of Remote Sensing · 年份：2025 · DOI：10.34133/remotesensing.0439 · 被引用次数：8 · 研究领域：Remote Sensing and Land Use

Large-scale model transfer facilitates crop classification in unlabeled sample regions. However, due to the spectral differences in the satellite image time series (SITS) of the same crop type caused by variations in a crop-growing environment between regions, cross-regional model transfer faces important challenges. Given that models trained in the source domain are affected by SITS variations and perform poorly in the target domain, in this paper, we propose an unsupervised domain adaptation method based on climate indicator discrepancy (ClimID-UDA), which addresses the problem of cross-region model transfer by mitigating SITS discrepancies using climate indicator discrepancy. In ClimID-UDA, we selected 6 climate variables representing the conditions of light, heat, water, and pressure and then constructed an index called climate indicator by calculating the barycenter of the 6 climate variables. Finally, climate indicator discrepancy was used to correct the SITS of the target domain to adapt the model trained in the source domain. The proposed method was tested in 4 regions in both China and Europe. The experiments covered different satellite sensors, different classification models, and different years. The experimental results show that ClimID-UDA achieves a more than 11% improvement in average accuracy and provides a viable option for large-scale cross-regional model transfer.