CSTN: A cross-region crop mapping method integrating self-training and contrastive domain adaptation

Abstract

Crop mapping is essential for agricultural management and food production monitoring, but challenges like limited crop labels and poor model generalization significantly hinder large-scale crop mapping. Here, we introduce a novel Contrastive Self-Training Network (CSTN), integrating a self-training strategy and contrastive domain adaptation (CDA) for cross-region crop mapping. CSTN uses pseudo-labels in the target region generated by the self-training strategy to assist supervised learning, and aligns features across regions using class-aware prototypes. Qualitative and quantitative evaluations demonstrate that CSTN significantly outperforms state-of-the-art methods with a 12.29 % increase in average F1-score, particularly in maize identification. Moreover, CSTN also enables early-season crop classification for pre-harvest decision-making applications. The interpretability of the model is demonstrated through an in-depth analysis of feature map visualizations, attention map visualizations, and the effectiveness of the modules. This study provides a robust method for enhancing large-scale crop mapping and facilitating more accurate and timely agricultural practices.