PMNet: A Positional Mamba Network for SAR-Assisted Cloud Removal

Abstract

Synthetic aperture radar (SAR) can provide reliable observations under cloud-contaminated conditions. Therefore, SAR–optical fusion has become a promising strategy for cloud removal in remote sensing imagery. State-space models (SSMs), such as Mamba, are effective at capturing long-range dependencies. However, their sequential scanning can weaken the explicit modeling of 2-D spatial structures. This limitation may obscure the spatial identity of features during sequential modeling and hinder the exploitation of the existing correspondence between SAR and optical features. To address this issue, a positional Mamba network (PMNet) is proposed, consisting of a dual-stream encoder (DSE) and a positional Mamba fusion (PMF) module. Specifically, the DSE processes SAR and optical streams separately to extract complementary information, capturing robust structural cues from SAR data and contextual information from optical images. Furthermore, the PMF introduces a shared learnable 2-D positional prior before Mamba scanning. It provides an explicit spatial reference for SAR and optical features. This design reduces spatial-structure degradation during sequential scanning and supports feature interaction at corresponding locations. Extensive experiments on the SEN12MS-CR dataset demonstrate that PMNet outperforms existing methods. It achieves a PSNR of 32.85 dB and an SSIM of 0.926 while improving structural and textural reconstruction under cloud-contaminated conditions.