Unbalanced Class Learning Network With Scale-Adaptive Perception for Complicated Scene in Remote Sensing Images Segmentation

The semantic segmentation of wide-field remote sensing images (RSIs) plays a significant role in many fields. However, due to the complexity of the content of RSIs, the dataset often has an uneven distribution of land type between different classes and large gaps in the scales of different objects. This often creates great problems for fine segmentation. To solve the issues, an unbalanced class learning network with scale-adaptive perception (UCSANet) is proposed, which can adaptively cope with multiscale objects and unbalanced classes. The design can be inserted in any convolution network easily and can enrich features without increasing too many parameters. The network groups feature and use atrous convolutions with different dilated rates on different groups to extract multiscale features while separable convolutions reduce the amount of network parameters. Then, the fusion of features between different scales is achieved through the self-attention mechanism. Furthermore, a weight map is designed to adaptively combine the predictions of two segmentation heads with cross-entropy loss and Lovasz-Softmax loss, respectively, which enable the network to focus on learning low-frequency classes without affecting high-frequency classes. Experimental results on GF-6 MSI datasets demonstrate that the proposed UCSANet performs significantly better than others and achieves multiclass segmentation more accurately.