Abstract: In order to solve the problem that the traditional full convolutional neural network could not achieve accurate extraction of cultivated land from high-resolution image, this study used the high-resolution 2 remote sensing satellite imagery as the data source, and used the improved U-Net network model （RMAU-Net network model） that integrated residual structure and multiple attention mechanisms to extract the cultivated land in the study area. The RMAU-Net network model was trained by using cultivated land samples, and cultivated land was extracted from the test set images using the trained network model. In order to verify the effect of RMAU-Net network model in extracting cultivated land, three traditional full Convolutional neural network models, DeeplabV3+, PSPNet and U-Net, were selected for comparative analysis with RMAU-Net network model. The results showed that the accuracy, recall, Intersection over Union, and F1 score of the RMAU-Net network model extraction were 90.36%, 90.78%, 82.57%, and 90.57%, respectively. Compared with DeepLabv3+, PSPNet, and U-Net network models, the RMAU-Net network model performed the best. RMAU-Net network model provided new ideas and methods for precise extraction of cultivated land, and provided basic data support for practical applications such as crop area monitoring and yield estimation.

Key words: high-resolution imaging, extraction of cultivated land, deep learning, attention mechanism, residual structure, RMAU-Net network model