Cross-modal change detection using historical land use maps and current remote sensing images

Using bi-temporal remote sensing imagery to detect land in urban expansion has become a common practice. However, in the process of updating land resource surveys, directly detecting changes between historical land use maps (referred to as “maps” in this paper) and current remote sensing images (referred to as “images” in this paper) is more direct and efficient than relying on bi-temporal image comparisons. The difficulty stems from the substantial modality differences between maps and images, presenting a complex challenge for effective change detection. To address this issue, in this paper, we propose a novel deep learning model named the cross-modal patch alignment network (CMPANet), which bridges the gap between different modalities for cross-modal change detection (CMCD) between maps and images. Our proposed model uses a vision transformer (ViT-B/16) fine-tuned on 1.8 million remote sensing images as an encoder for images and trainable ViTs as the encoder for maps. To bridge the distribution differences between these encoders, we introduce a feature domain adaptation image-map alignment module (IMAM) to transfer and share pretrained model knowledge rapidly. Additionally, we incorporate the cross-modal and cross-channel attention (CCMAT) module and the transformer block attention module to facilitate the interaction and fusion of features across modalities. These fused features are then processed through a UperNet-based feature pyramid to generate pixel-level change maps. These fused features are then processed through a UperNet-based feature pyramid to generate pixel-level change maps. On the newly created EVLab-CMCD dataset and the publicly available HRSCD dataset, CMPANet has achieved state-of-the-art results and offers a novel technical approach for CMCD between maps and images.