Label refinement for change detection in remote sensing

Abstract

Change detection in remote sensing aims to detect changes occurring in the same geographical area over time. Existing methods present two main challenges: (1) relying on single-scale features to capture multi-scale object changes, which limits their ability to effectively handle multi-scale change; and (2) misclassification issues caused by prediction uncertainty, particularly in regions near decision boundaries, leading to reduced overall detection performance. In this study, to address these limitations, we propose LRNet, a multi-scale change detection framework designed to enhance the perception of objects at varying scales and refine change region details during decoding. Abandoning the use of fixed thresholds for classification, LRNet incorporates a Label Refinement (LR) strategy that propagates information from high-confidence regions to low-confidence regions by evaluating feature-space similarity, enabling precise grouping of pixels within change regions. Extensive experiments on benchmark datasets — SYSU-CD, LEVIR-CD+, and SECOND-CD — demonstrate that LRNet outperforms state-of-the-art methods, with significant improvements of 7.9% in F1 and 12.38% in IoU on the challenging SECOND-CD dataset.