关键词: 深度学习, 盐碱地, 遥感影像, 反演模型, 注意力机制

Abstract: To address the inversion problem of remote sensing images of saline-alkali land in the Tailan River Irrigation District of Wensu County, Aksu Prefecture, Xinjiang, an image classification model based on the PointRend network integrated with the CoTAttention mechanism was proposed. The model used DeepLabV3 as the backbone network and incorporated the CoTAttention module to enhance the feature extraction capability of the network. To verify the classification performance of the improved network, experiments were conducted using Sentinel-2 remote sensing images from the Tailan River Irrigation District. The results showed that the PointRend model achieved good performance metrics, with pixel accuracy, mean intersection over union (MIoU), and F1 reaching 89%, 88%, and 88%, respectively. Compared with the original PointRend model without the CoTAttention mechanism, the improved PointRend model demonstrated enhanced metrics, with pixel accuracy, MIoU, and F1 increasing by 3, 2, and 3 percentage points, respectively. The improved PointRend model significantly enhanced the ability to capture fine details and refine edges in segmentation tasks, thereby improving overall classification accuracy.

Key words: deep learning, saline-alkali land, remote sensing images, inversion model, attention mechanism