HCFFormer: a hybrid CNN–transformer framework for heterogeneous remote sensing image change detection

Heterogeneous change detection is concerned with identifying surface changes based on multi-temporal remote sensing images, taken using a variety of sensors. The main problem lies in the fact that one has to reconcile non-linear radiometric differences and modality-specific distortions in order to maintain semantic consistency. In order to overcome the aforementioned problems, we propose Heterogeneous Cross-modal Fusion Transformer (HCFFormer), a unified hybrid CNN-Transformer architecture for heterogeneous change detection across domains with high quality, low cost and semantic consistency. HCFFormer integrates the capabilities of convolutional neural networks for modelling textures at local level with the global contextual reasoning capabilities of transformers. A cross-attention fusion module will dynamically align the heterogeneous features in the latent space, while a multi-scale graph aggregation block will capture hierarchical spatial dependencies to increase the precision of boundaries. Moreover, we introduce a domain consistency regularizer to guarantee the robustness of the cross-domain feature alignment in a non-adversarial manner. Experiments were performed on three benchmark datasets to evaluate the performance of HCFFormer, showing that it outperforms all the state-of-the-art methods in terms of accuracy, spatial coherence and computational complexity. Therefore, the proposed framework presents a new perspective for scalable and semantically consistent heterogeneous change detection in multimodal remote sensing applications. The source code is publicly available at: https://github.com/x1xSAS/HCFFormer.