Adaptive Adversarial Cross-Domain Segmentation Network for High-Resolution Remote Sensing Images

Convolutional neural networks (CNNs) have dominated the field of semantic segmentation of high-resolution remote sensing images (HRSIs). These models often assume that training and test data follow the same distribution. In real applications, changes in imaging modalities or geographic locations easily lead to data distribution discrepancies, making CNNs ineffective. Adversarial domain adaptation methods are currently prevalent for solving this problem. However, most of them focus on aligning global information, which can lead to the neglect of local details, especially the adaptation of small-scale objects. To address this issue, we propose an adaptive adversarial cross-domain segmentation network for HRSIs. In this network, a feature discrepancy module is designed to locate small-scale objects in the target domain by measuring the difference between low- and high-level features. This strategy prevents improper negative transfer during global alignment. Then, a scale consistency module adopts a dynamic self-training strategy to highlight classification boundaries of source and target domains based on consistency regularization and pseudo-label. Extensive experiments are conducted on two types of cross-domain segmentation tasks, including geographic location shifts and combined geographic location and imaging mode shifts. Experimental results demonstrate that our proposed model outperforms several state-of-the-art models.