Chen Hongruixuan ; Song Jian ; Han Chengxi ; Xia Junshi ; Yokoya Naoto

Abstract

Convolutional neural networks (CNN) and Transformers have made impressiveprogress in the field of remote sensing change detection (CD). However, botharchitectures have inherent shortcomings: CNN are constrained by a limitedreceptive field that may hinder their ability to capture broader spatialcontexts, while Transformers are computationally intensive, making them costlyto train and deploy on large datasets. Recently, the Mamba architecture, basedon state space models, has shown remarkable performance in a series of naturallanguage processing tasks, which can effectively compensate for theshortcomings of the above two architectures. In this paper, we explore for thefirst time the potential of the Mamba architecture for remote sensing CD tasks.We tailor the corresponding frameworks, called MambaBCD, MambaSCD, andMambaBDA, for binary change detection (BCD), semantic change detection (SCD),and building damage assessment (BDA), respectively. All three frameworks adoptthe cutting-edge Visual Mamba architecture as the encoder, which allows fulllearning of global spatial contextual information from the input images. Forthe change decoder, which is available in all three architectures, we proposethree spatio-temporal relationship modeling mechanisms, which can be naturallycombined with the Mamba architecture and fully utilize its attribute to achievespatio-temporal interaction of multi-temporal features, thereby obtainingaccurate change information. On five benchmark datasets, our proposedframeworks outperform current CNN- and Transformer-based approaches withoutusing any complex training strategies or tricks, fully demonstrating thepotential of the Mamba architecture in CD tasks. Further experiments show thatour architecture is quite robust to degraded data. The source code will beavailable in https://github.com/ChenHongruixuan/MambaCD