Decision Calibration Network for Semantic Labeling of High-Resolution Remote Sensing Images

Abstract

Semantic labeling of high-resolution remote sensing images is a challenging task, requiring the models to effectively distinguish different classes of ground objects while learning advanced feature representations. First of all, we propose a dual-decoder semantic labeling neural network based on the atrous spatial pyramid pooling module and attention mechanism to achieve the high-precision classification of different ground objects. The main idea is to enhance the high-level feature representation by using the complementary relationship that may exist between different decoders. Furthermore, based on this network structure, a decision calibration auxiliary loss is proposed to improve the models’s ability to classify examples of highly ambiguous output by different decoders. Finally, we conduct experimental verification on the ISPRS Vaihingen and Potsdam datasets, and the results show that the auxiliary loss can effectively improve the classification accuracy of the model.