SAR Images Change Detection Based on Maximum Classifier Difference with Multiscale Multilevel Features

Synthetic aperture radar (SAR) images can still be effective under harsh weather and lighting conditions, but the presence of scattering noise and the lack of labeled sufficient datasets have severely impacted the development of change detection (CD) methods and limited the performance of CD models. In order to solve the above problems and make full use of the existing labeled data, we propose a CD model based on the maximum classifier difference of multi-scale features. Firstly, we extract multiscale features directly from the original images and fuse them, after which the multiscale features are fed into a Siamese network for multilevel feature extraction, followed by fusion of the multilevel features for use in subsequent CD. The extracted features are then fed into the classification module and subjected to maximum classifier variance processing. While ensuring the performance of the model on the source domain dataset, the parameters of the two classifiers are updated so that the difference between their results on the target domain dataset increases, and then the parameters of the feature extractor are updated so that the difference between the results of the two classifiers on the target domain dataset decreases. Repeating the above process can achieve a reduction in the difference between the source and target domain datasets, so that the model can make full use of the labeled data and can perform well on the unlabeled data as well. We conducted comparative experiments on four publicly available SAR datasets, and the experimental results demonstrate the effectiveness and sophistication of the method.