CrossCD-Eval: a unified framework with domain-aware feature normalization for cross-domain remote sensing change detection

Abstract

Change detection (CD) in remote sensing imagery is essential for monitoring urban growth, land-use transformation, and environmental dynamics. However, cross-domain change detection remains challenging due to distribution shifts caused by differences in geographic regions, spatial resolutions, sensor/acquisition conditions, temporal periods, and annotation protocols. This paper introduces CrossCD-Eval, a unified framework for evaluating the generalization of deep learning-based change detection models under in-domain, cross-domain, and few-shot domain adaptation settings. Five recent models, namely GLAFormer1, Change-Mamba2, EHCTNet3, ELGC-Net4, and STNet5, are evaluated using CLCD, HRSCD, and LEVIR-CD. To reduce representation-level domain mismatch, we propose a Domain-Aware Feature Normalization (DAFN) module that aligns source and target feature statistics as a lightweight feature-level adaptation component. Experimental results show that all models achieve strong in-domain performance but degrade under cross-domain transfer, confirming the impact of domain shift. DAFN improves cross-domain and few-shot adaptation performance, with larger gains observed under more challenging source–target shifts. Repeated 10-shot experiments over five random target-sample selections further show that DAFN improves adaptation stability, reducing IoU standard deviation by up to 22.7%. An additional ablation and baseline-comparison study demonstrates that the observed gains are not solely due to preprocessing. Compared with full preprocessing, AdaBN, adversarial feature alignment, and the instance-normalization-based CrossCDNet baseline, DAFN provides additional feature-level adaptation benefits, especially when combined with strong backbones such as STNet.