Yang et al. (2025) Phenology-Guided Wheat and Corn Identification in Xinjiang: An Improved U-Net Semantic Segmentation Model Using PCA and CBAM-ASPP

Identification

• Journal: Remote Sensing
• Year: 2025
• Date: 2025-10-28
• Authors: Wei Yang, Xian Guo, Yiling Lu, Hsiao-Wei Hu, Fei Wang, Rongrong Li, Xiaojing Li
• DOI: 10.3390/rs17213563

Research Groups

• School of Architecture and Civil Engineering, Chengdu University, Chengdu 610106, China
• School of Earth Resources, China University of Geosciences, Wuhan 430074, China

Short Summary

This study developed an improved U-Net semantic segmentation model, integrating Principal Component Analysis (PCA), a Convolutional Block Attention Module (CBAM), and an enhanced Atrous Spatial Pyramid Pooling (ASPP) module, guided by phenological analysis, to accurately identify wheat and corn in Xinjiang, achieving an overall accuracy of 90.91%.

Objective

• To improve the accuracy of wheat and corn distribution identification in complex environments, specifically addressing challenges posed by spectral similarity and complex spatial distribution in Xinjiang, China.

Study Configuration

• Spatial Scale: Qitai County, Changji Hui Autonomous Prefecture, northeastern Xinjiang, China.
• Temporal Scale: Data acquired between April and August 2022, with an identified optimal identification window from day 156 to day 176 of the year.

Methodology and Data

• Models used: Improved U-Net (with ResNet50 backbone, CBAM, and depthwise separable ASPP), DeeplabV3+, PSPnet, HRnet, Segformer, standard U-Net.
• Data sources: Sentinel-2 satellite imagery (10 m resolution), kNDVI and EVI time series, ground truth data from field surveys (April-August 2022), Google Earth images. Data pre-processing included atmospheric correction, band resampling, and Principal Component Analysis (PCA) to extract PCA1, PCA2, and PCA3.

Main Results

• The optimal time window for distinguishing wheat and corn based on phenological differences (kNDVI and EVI time series) was identified as days 156–176 of the year.
• The proposed improved U-Net model achieved superior performance on the PCA-transformed Sentinel-2 dataset, with a mean Intersection over Union (mIoU) of 83.03%, mean Pixel Accuracy (mPA) of 91.34%, F1-score of 90.73%, and Overall Accuracy (OA) of 90.91%.
• This represents an OA improvement of 5.97% over DeeplabV3+, 4.55% over PSPnet, 2.03% over HRnet, 8.99% over Segformer, and 1.5% over the standard U-Net.
• Using the PCA dataset improved recognition accuracy by approximately 2% compared to the traditional RGB dataset.
• The integration of the CBAM attention mechanism improved mIoU by 2.40% and F1-score by 1.47%, while the improved depthwise separable ASPP module improved mIoU by 1.28% and F1-score by 0.79%.
• The depthwise separable ASPP module significantly reduced model parameters (from 191.7 million to 91.1 million), computational load (from 15.97 GFLOPS to 12.75 GFLOPS), and memory usage (from 931.91 MB to 548.87 MB).
• The model demonstrated strong generalisability, with prediction errors less than 2% in Qitai County, Xinjiang.

Contributions

• Developed a novel "time series + data + model" technical path for accurate crop identification by integrating phenological analysis, PCA-transformed Sentinel-2 data, and an improved U-Net model.
• Introduced an improved U-Net semantic segmentation model incorporating ResNet50 as a backbone, a Convolutional Block Attention Module (CBAM), and a depthwise separable Atrous Spatial Pyramid Pooling (ASPP) module, enhancing accuracy and computational efficiency for crop identification.
• Demonstrated the effectiveness of Principal Component Analysis (PCA) in constructing a high-efficiency dataset that integrates visible and near-infrared band information, outperforming traditional RGB datasets for crop discrimination.
• Identified an optimal phenological window (days 156–176) for distinguishing wheat and corn in Xinjiang, leveraging kNDVI and EVI time series to mitigate spectral similarity interference.
• Validated the model's strong generalisability and potential for application in arid grain-producing areas for crop mapping, area statistics, and yield estimation, contributing to national food security.

Funding

• National Natural Science Foundation of China (42101363)
• Research Initiation Fund of Chengdu University (2081923044 and 2081923045)

Citation

@article{Yang2025PhenologyGuided,
  author = {Yang, Wei and Guo, Xian and Lu, Yiling and Hu, Hsiao-Wei and Wang, Fei and Li, Rongrong and Li, Xiaojing},
  title = {Phenology-Guided Wheat and Corn Identification in Xinjiang: An Improved U-Net Semantic Segmentation Model Using PCA and CBAM-ASPP},
  journal = {Remote Sensing},
  year = {2025},
  doi = {10.3390/rs17213563},
  url = {https://doi.org/10.3390/rs17213563}
}