Improving HRNet for crop classification using Sentinel-2 time-series features

Crop distribution mapping is critical for safeguarding national food security and optimizing agricultural management strategies. However, accurately distinguishing spectrally similar crops while maintaining precise boundaries remains a significant challenge. To address this, we propose an enhanced High-Resolution Network, termed HRNet-ICoT-CNN. The key innovation of this model is the Improved Contextual Transformer (ICoT) module, which dynamically integrates local structural details with global temporal contexts, thereby enhancing multi-scale spatiotemporal feature extraction without compromising spatial resolution. Applied to Sentinel-2 time-series NDVI data from Xinhe County in the Aksu region of Xinjiang, China, the model accurately delineated the distributions of key crops – cotton, wheat, and maize – and was evaluated against leading existing models. Experimental results indicate that HRNet-ICoT-CNN outperforms baseline models across all major evaluation metrics, achieving an overall accuracy (OA) of 94.91%, average accuracy (AA) of 81.81%, and Intersection over Union (IoU) values of 0.8507 for cotton, 0.6467 for wheat, and 0.5513 for maize. Generalizability tests further affirmed the model’s robustness, yielding an OA of 81.47% and an AA of 73.20% in two untrained counties, with IoUs of 0.8750, 0.9065, and 0.6314 for the three crops, respectively. These results highlight the model’s strong transferability and reliability, offering a robust scientific foundation for crop mapping via remote sensing in arid regions.