A Lightweight High-Resolution Remote Sensing Image Cultivated Land Extraction Method Integrating Transfer Learning and SENet

As one of the main types of land cover, cultivated land is the direct carrier of food production, and its changes have a very important impact on food security. Timely and accurate acquisition of cultivated land area and distribution information is of great significance to ensuring food security and promoting economic development. Classic semantic segmentation models usually have a large number of training parameters, and there are problems of inaccurate segmentation and low efficiency when performing image segmentation. Moreover, most methods of cultivating land extraction are only applicable to certain specific data sets or specific research areas, and the generalization ability of the model is poor. In response to these problems, this article proposes an improved Deeplab v3+ extraction method. First, the lightweight network MobileNetV2 is used to replace the feature extraction network Xception of the DeepLabv3+ model to reduce the number of model parameters and improve the training speed. Secondly, compare different attention mechanism modules and select SE attention to join the model to enhance the network’s extraction accuracy of cultivated land and improve the accuracy of semantic segmentation of remote sensing images. Finally, transfer learning is introduced, and the feature extraction network trained on the ImageNet data set is used as a pre-training model to enhance the model’s ability to obtain features and improve the network classification accuracy. The results of multi-model comparison experiments on the GID data set show that the method in this paper has excellent segmentation performance. It can effectively solve the problems of varying degrees of misclassification, missing classification and over-segmentation in the segmentation results of classic models, improve the phenomenon of inaccurate boundary and contour segmentation of ground objects, and provide support for subsequent cultivated land extraction tasks.