EDG-Net: Edge-Enhanced Dynamic Graph Convolutional Network for Remote Sensing Scene Classification of Mining-Disturbed Land

Scene classification and mapping of surface mining-disturbed land can attain semantic-level information that is useful for monitoring mine geo-environment. Mining land’s complex characteristics makes it difficult to extract key features, restricting the accuracy improvements. This study first constructed a 5-class dataset based on multispectral, synthetic aperture radar, and topographic images. Subsequently, a novel model of edge-enhanced dynamic graph convolutional network (GCN) (EDG-Net) was proposed to learn the discriminative features for classification of mining land with irregular edges, different sizes, a relatively small proportion, and sparse spatial distribution. (1) Edge-enhanced multiscale attention module: it is designed to capture key multiscale features and edge details using parallel dilated convolutions with attention fusion and edge enhancement, which facilitates the identification of objects with irregular edges and different sizes. (2) Downsampling fusion module: it integrates the features obtained through spatially split learning and max-pooling to overcome the information loss issue of small objects. (3) Patch-based dynamic GCN: the input images were split into several patches as nodes, and a graph was constructed and dynamically updated by connecting the nearest neighbors. It is beneficial to progressively explore the inherent attributes within node feature maps and fully utilize long-range information, which helps address the sparse distribution issue of mines. Finally, the convolutional and graph features are fused using a residual structure to obtain richer feature representations. The proposed EDG-Net achieved an overall accuracy of 78.08% ± 0.22% and acceptable regional-scale mapping performance, indicating that the proposed dataset and model were beneficial for classification and mapping of mining land.