Abstract: To accurately identify crop planting types in the Tailan River Irrigation District of Wensu County, Xinjiang, an improved pixel classification method for remote sensing imagery was proposed based on deep learning technology.Using Gaofen series satellite data (including GF1, GF1B, GF6, and GF2), fine annotation was performed based on GF2 fused imagery, and the annotation results were validated through field surveys. The validated annotations were then uniformly registered to other imagery data, ultimately constructing a multi-category classification dataset with a sample size of 128 px × 128 px.The EAA-UNet model significantly improved feature extraction performance by integrating the ECA module, ASSP multi-scale feature extraction module, and gated attention mechanism module.Quantitative evaluation was conducted using three metrics: Accuracy (A), Precision (P), and Mean Intersection over Union (MIoU), and a comparative analysis with current mainstream pixel classification models was performed to validate the superiority of the EAA-UNet model. The results showed that the EAA-UNet model achieved Accuracy, Precision, and MIoU values of 71.56%, 82.79%, and 69.94%, respectively, all higher than those of the FCN8s, UNet, ResUNet, and DeepLabV3+ models.The EAA-UNet model performed excellently in the crop classification task in the Tailan River Irrigation District, with its predictions highly consistent with the actual planting types of the plots. The EAA-UNet model effectively addressed the ambiguity in crop classification in the irrigation district through its introduced improvements, not only achieving precise identification of planted crop types but also providing reliable data support for subsequent water demand prediction in the Tailan River Irrigation District.

Key words: EAA-UNet model, remote sensing imagery, crop, remote sensing classification, Tailan River Irrigation District of Wensu County