Land Cover Classification Based on PSPNet Using Remote Sensing Image

In recent years, land cover classification has been modeled as semantic segmentation of remote sensing images, and significant advances have been achieved. Labeled examples for some categories are difficult to obtain manually by photo interpretation or ground survey, thereby causing category imbalance problems. In some categories, intraclass variability is large, but interclass difference is small, causing hard discrimination. To segment pixels accurately, we proposed an improved land cover classification network based on Pyramid Scene Parsing Network. In our network, an adaptation loss based on focal loss is proposed to increase the focus on indistinguishable pixels for category imbalance. Moreover, the network aggregates multiscale features to obtain fused local and global context information using multiple dilated convolutions with various dilation factors, avoiding information loss causing by large intraclass variability and small interclass difference. Experiments were conducted on real land cover datasets. These experiments confirmed the superior performance of the proposed network compared with the state-of-the-art land cover classification models.