Diversity Matters: Co-training for Semi-Supervised Change Detection in Remote Sensing Images

Abstract

General change detection (CD) methods require extensive annotated data to ensure effective performance, yet the annotation of remote sensing (RS) bi-temporal images is significantly time-consuming and labor-intensive. While numerous RS methods have employed semi-supervised (SS) learning to tackle the issue of insufficient labeled data, they encounter challenges related to confirmation bias. In this paper, we propose a novel two-branch co-training framework for SSCD, which aims to capture diversity predictions from heterogeneous representations to boost SSCD performance. Specifically, we first adopt a change-aware cutmix strategy to enlarge the diversity of input for a more diverse output. Then, we develop a correlation suppression to generate distinctive features, aiming to reduce homogenization. This involves weakening features by randomly masking the intermediate high-dimensional feature similarity between two branches of the same sample. Furthermore, we introduce a discrepancy supervision scheme to amplify the diversity of co-training cross-supervision, which alleviates the confirmation bias during the learning process from the views of the two branches. Extensive experiments conducted on the WHU-CD, and GZ-CD datasets demonstrate that the proposed approach performs favorably against the state-of-the-art methods.