Li et al. (2026) Interpretable Cotton Mapping Across Phenological Stages: Receptive-Field Enhancement and Cross-Domain Stability

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Remote Sensing
• Year: 2026
• Date: 2026-03-25
• Authors: Li Li, Jinhua Wang, Keke Jia, Jianli DING, Xiangyu Ge, Zhihong Liu, Zheng Zhang, Hongzhi Xiao
• DOI: 10.3390/rs18070980

Research Groups

Not explicitly provided in the text. The study focuses on the Wei-Ku Oasis, Xinjiang, China.

Short Summary

This study developed an interpretable semantic segmentation framework for cotton mapping in arid irrigated agroecosystems using multi-source remote sensing data, achieving high classification accuracy and robust generalization while explicitly quantifying the importance of different predictors across phenological stages.

Objective

• To develop an interpretable semantic segmentation framework for accurate and timely cotton-field mapping in the Wei-Ku Oasis, Xinjiang, China, under multi-source remote sensing conditions, addressing limitations of existing deep-learning approaches regarding interpretability and performance variability across phenological stages.

Study Configuration

• Spatial Scale: Wei-Ku Oasis, Xinjiang, China.
• Temporal Scale: Cotton growing season, with monthly analysis highlighting August as the most discriminative acquisition window, and June–July also maintaining high accuracy. Cross-year evaluation was also performed.

Methodology and Data

• Models used: An interpretable semantic segmentation framework integrating a receptive-field enhancement mechanism and an embedded feature-attribution module. Compared against U-Net, DeepLabV3+, and SegFormer baselines.
• Data sources: Sentinel-2 surface reflectance, Sentinel-1 VV/VH backscatter, Digital Elevation Model (DEM), vegetation indices (NDVI, EVI), and Gray-Level Co-occurrence Matrix (GLCM) texture features.

Main Results

• The proposed model achieved a mean Intersection over Union (mIoU) of 85.62% and an F1-score of 92.96% on the test set, outperforming U-Net, DeepLabV3+, and SegFormer baselines.
• Monthly classification results indicated August as the most discriminative acquisition window (mIoU = 85.54%, F1 = 92.83%), with June–July also maintaining high recognition accuracy.
• Feature attribution revealed varying predictor importance across phenological stages: Sentinel-2 red-edge bands were consistently influential, NDVI/EVI contributions increased during June–August, SAR VH was more important during peak canopy development, and DEM maintained stable information contribution.
• The model demonstrated strong generalization capability, achieving an mIoU of 82.81% in same-region cross-year evaluation and 74.56% under cross-region transfer.

Contributions

• Developed an interpretable deep-learning framework for cotton-field mapping that not only improves classification accuracy but also explicitly models and quantifies feature importance, providing insights into optimal acquisition timing and predictor relevance.
• Offers a methodological reference for cotton-field mapping and acquisition timing selection in arid irrigated regions, enhancing reliability for operational deployment.

Funding

Not provided in the text.

Citation

@article{Li2026Interpretable,
  author = {Li, Li and Wang, Jinhua and Jia, Keke and DING, Jianli and Ge, Xiangyu and Liu, Zhihong and Zhang, Zheng and Xiao, Hongzhi},
  title = {Interpretable Cotton Mapping Across Phenological Stages: Receptive-Field Enhancement and Cross-Domain Stability},
  journal = {Remote Sensing},
  year = {2026},
  doi = {10.3390/rs18070980},
  url = {https://doi.org/10.3390/rs18070980}
}