SGANet: A Siamese Geometry-Aware Network for Remote Sensing Change Detection

Abstract

The significant progress in the fields of deep learning and computer vision has propelled the development of remote sensing change detection. However, we noticed that previous methods still rely on the single visual modality and cannot effectively utilize other prior information, such as elevation or depth maps. Therefore, this article presents a novel Siamese geometry-aware network (SGANet) intended for RGB-D remote sensing change detection. By incorporating both RGB data and geometry priors such as relative depth estimations derived from a monocular depth estimation model such as DepthAnythingV2, SGANet surpasses the limitations of traditional methods that primarily depend on visual data. The proposed network employs a shared siamese encoder architecture with a lightweight decoder head for efficient change map prediction. Within the encoder blocks, we integrated a local feature extraction block that excels at capturing fine-grained features and a global cross-attention block that focuses on contextual features between different modalities. Furthermore, we engineered a dual-path fusion structure that facilitates a seamless integration of vision and geometry features. Extensive experiments on the LEVIR-CD, WHU-CD, SYSU-CD, and S2Looking-CD datasets demonstrated that SGANet achieved substantial enhancements in F1-Score and intersection over union compared to benchmark methods that are in vogue. By integrating geometry priors and effective multimodal fusion mechanisms, SGANet promoted the development of geometry-aware change detection, further enhancing optimal performance.