Cross-phenological-region crop mapping framework using Sentinel-2 time series Imagery: A new perspective for winter crops in China.

Precise large-area crop mapping products are the foundation for cropland monitoring, market policy decision-making, and subsequent agricultural applications. However, owing to sharp discrepancies among the temporal-spectral attributes of crops from different areas, crop mapping across phenological regions faces significant challenges. Given the inadequacy of generalizing mapping frameworks from the source to target regions, research areas are often limited, and trained models cannot be reused beyond a particular study area. In this paper, we propose a novel framework called the Phenology Alignment Network (PAN) to address the cross-phenological-region (CPR) crop-mapping problem using deep recurrent networks and unsupervised domain adaptation. Our PAN adopts a Siamese structure consisting of two identical deep models called the Temporal Spectral Network (TSNet), which serves as an adaptive multilevel phenological feature extractor for crops under various planting conditions. Specifically, the two branches of PAN accept crop samples from the source and target regions, and aim to encode them into similar deep phenological features. By constraining the distribution of hidden states from the two branches, the deep temporal-spectral features are aligned, and regional discrepancies between the source and target regions are alleviated. Consequently, the PAN adapts the deep model pretrained on the source region to the target area, improving the mapping accuracy without using additional target label information. To verify the effectiveness of our model, crop samples were collected from three plains of China, and a winter crop dataset based on Sentinel-2 time-series imagery containing over two million pixels was annotated. Experiments conducted on this dataset indicate that PAN serves as a practical CPR crop mapping framework and achieves significant improvement in the overall accuracy as well as the macro-average F1 score compared to other state-of-the-art methods. Further visual interpretation illustrates the capability of PAN to correct inaccurate predictions ascribed to phenological diversity and suggests the potential of PAN for large-area crop mapping applications. [ABSTRACT FROM AUTHOR]