Land type classification based on U-net Using multispectral images of Gaofen-2 satellite

This article is based on the U-Net deep learning model and uses the GF-2 multispectral images to classify the land types in the Dahaize mining area. The study first preprocessed remote sensing image data to construct a sample library containing categories such as water systems, forests, residential areas, and cultivated land, and used data augmentation methods to expand the training sample size. By training the U-Net model and evaluating its accuracy, the results show that the U-Net model performs relatively evenly in multi class land type classification, with overall high classification accuracy, especially in extracting natural categories such as vegetation and bare land. Compared with traditional random forest and support vector machine (SVM) methods, the U-Net model has significant advantages in accuracy, classification consistency (Kappa coefficient), and user accuracy (U_Accuracy). The study also combined vegetation distribution changes analysis to evaluate the effectiveness of ecological restoration in the mining area, and found that vegetation restoration was significant and the ecological environment was improved. This study indicates that remote sensing image classification methods based on deep learning have strong adaptability and accuracy, and can effectively support applications in fields such as resource exploration and ecological environment monitoring.