A Semi-supervised Hybrid Multi-scale Network for Synthetic Aperture Radar Image Change Detection

With its unique imaging characteristics, synthetic aperture radar (SAR) image is increasingly attracting interests in the fields of remote sensing. Existing SAR image change detection methods mainly train network by pre-classification to select reliable labels for each pixel point. However, the training samples obtained by these pre-classification methods on noise-heavy datasets are devoid of diversity, which may lead to poor generalization performance of the network. In this research, a semi-supervised hybrid multi-scale network (SHMNet) is proposed for SAR image change detection. Specifically, this method first analyzes the difference maps using a hierarchical clustering algorithm to classify all pixel points into samples with labels and samples without labels. Then we gain labels with low-entropy for the unlabeled samples and then mix the labeled and unlabeled samples by using MixUp. and the semi-supervised loss function takes into account the information contained in the samples without labels, which can enhance the generalization performance of the model. The experimental results on two SAR datasets confirm the efficiency of the proposed SHMNet.