LCS-EnsemNet: A Semisupervised Deep Neural Network for SAR Image Change Detection With Dual Feature Extraction and Label-Consistent Self-Ensemble

Abstract

Change detection (CD) in synthetic aperture radar (SAR) images faces two challenging problems: 1) inherent speckle noise in SAR data causes the class confusion and affects the image understanding for inferring category of each image pixel; and 2) adequate labeled samples are quite hard to collect, which is the major limitation for supervised methods. In this article, we develop a novel deep learning-based semisupervised method to address these challenges. The method firstly incorporates a pixel-wise log-ratio difference image (DI) and its saliency map to produce a spatially enhanced (SE) DI using a reweighting scheme based on the fact that changed pixels exhibit higher saliency than unchanged pixels. As a result, prominent changed regions are highlighted and the interclass separability is increased. We construct pixel-wise and context-wise features based on the log-ratio DI and SE DI to jointly express the dissimilarity at each pixel. Second, a two-branch deep neural network taking the dual features as input is devised to reveal and refine underlying label information from abundant unlabeled samples with the help of a two-stage refinement strategy for improving generalization ability, called LCS-EnsemNet. The refinement strategy, namely label-consistent self-ensemble, exploits label consistency between dual features and across multiple classifiers to ensure the reliability of pseudo labels. Finally, the cross-entropy loss is calculated with the limited labeled data and selected pseudo labeled samples to optimize the LCS-EnsemNet. The proposed method is evaluated on three low/medium- resolution and one high-resolution SAR data sets, and experimental results have demonstrated its effectiveness.