Self-Supervised Confident Learning For Hyperspectral Image Change Detection

Hyperspectral images (HSIs) contain rich spectral information required by detection of multiple ground changes. HSI change detection (CD) via deep learning attracts lots of attention in recent years for its advantages in deep feature representation and nonlinear problem modeling. However, unexpected changes can be hardly labelled beforehand, which limits the application of supervised models for HSI CD. In this paper, we propose a novel self-supervised framework based on confident learning (SSCL) for HSI CD. It mainly comprises two branches, binary classification and label correction. For starters, two unsupervised methods are combined to select credible samples as the initial training set. It is used to train a feature extractor in the classification branch while being corrected by the CL and prototype selection in the correction branch. These two branches are updated alternatively as the former share the feature extractor with the latter and the latter provides credible pseudo labels for the former. Experiments on real HSI datasets have shown that our method outperforms many state-of-the-art methods.