IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (Jan 2021)

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

  • Jian Wang,
  • Yinghua Wang,
  • Bo Chen,
  • Hongwei Liu

DOI
https://doi.org/10.1109/JSTARS.2021.3122461
Journal volume & issue
Vol. 14
pp. 11903 – 11925

Abstract

Read online

Change detection (CD) in synthetic aperture radar (SAR) images faces two challenging problems limiting the detection performance: inherent speckle noise in SAR data causes the overlapping nature of changed and unchanged classes and, thus, affects the image understanding for inferring category of each image pixel; and adequate labeled samples are quite laborious and time-consuming 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 first 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 class separability is significantly increased. We construct pixel-wise and context-wise features based on the log-ratio DI and SE DI, which respectively provide image detail cue and spatial context cue, as dual input features to jointly characterize the change information at each pixel position. Second, we propose a label-consistent self-ensemble network (LCS-EnsemNet), which can take advantage of the unlabeled samples to learn discriminative high-level features for the precise identification of changed pixels. By enforcing a label consistency between dual features and a label consistency across multiple classifiers, the label-consistent self-ensemble strategy enables the proposed network to selectively transform unlabeled samples into pseudo-labeled samples in an unsupervised manner and ensures that the selected pseudo-labels are reliably and stably predicted. Finally, the cross-entropy loss is calculated with the limited labeled data and selected pseudo-labeled samples to optimize the LCS-EnsemNet in a supervised way. The proposed method is evaluated on three low/medium-resolution SAR datasets and one high-resolution SAR dataset, and experimental results have demonstrated its efficiency and effectiveness.

Keywords