SeCoR: Evidence-Guided Selective Correction for Lightweight Remote Sensing Change Detection

Abstract

Remote sensing change detection (RSCD) aims to identify land-cover changes from bitemporal images and is widely used in urban monitoring, disaster assessment, and land-resource analysis. Lightweight RSCD models usually rely on compact backbones and efficient temporal interaction, but most of them update bitemporal features in a dense or implicit manner, without explicitly verifying whether opposite-temporal evidence is reliable. This may amplify pseudochanges under appearance disturbances or miss weak true changes when valid support is limited. To address this issue, we propose SeCoR, an evidence-guided selective correction framework for lightweight RSCD. In the encoder, reliability-aware local support correction selectively aggregates reliable local support from the opposite temporal branch to correct features before fusion. In the decoder, adaptive prototype-prior repair converts confident coarse priors into image-specific prototypes to repair ambiguous low-level features. Experiments on four public optical RSCD benchmarks demonstrate a strong accuracy-efficiency tradeoff with only 2.50 M parameters and 2.66G FLOPs. SeCoR obtains the highest F1 and IoU scores on all four datasets, with only marginal gains on LEVIR-CD and larger gains on WHU-CD, SYSU-CD, and UAV-CD, while remaining competitive with the strongest method on LEVIR-CD. Five-seed evaluations further provide statistical evidence of reproducible improvements over the architecture-matched Baseline. A half-width variant reduces the model to 0.99M parameters and 1.99G FLOPs while retaining competitive accuracy.