Mamba-CR: A State-Space Model for Remote Sensing Image Cloud Removal

Abstract

Cloud removal is a critical step to improve the usability of cloud-contaminated remote sensing images. Deep learning methods based on convolutional neural networks (CNNs) and transformer architectures are widely used for cloud removal tasks. However, the localization of convolutional operations limits the capture of global features, and Transformer is prone to forget local spatial details when processing high-resolution images, which affects the further improvement of cloud removal accuracy. This study proposes a novel cloud removal model, Mamba-CR, based on the state-space model. In this model, the Convolution State-Space Group (CSSG) is designed to extract and reconstruct shallow spatial features of remote sensing images to alleviate the sequential dependence of Mamba on the data input of image processing sequences; the Transformer Structured State-Space Group (TSSG) is designed to enhance the model’s ability to sense global context and capture long-range spatial dependencies. Compared with other SOTA cloud removal methods on three publicly available datasets, Mamba-CR achieves a definite performance advantage while keeping the computational effort low.