Remote Sensing Land-Cover Classification Combining Multi-Modal and Multi-Scale Fusion with Mamba

No. 2025JC-YBQN-936), Shaanxi Provincial Department of Education Youth Innovation Team Research Program (25JP175), Shaanxi University Youth Innovation Team, Xi'an University of Posts and Telecommunications Xiyou New Star Team Project (XYT2016-01) Open source content Combining multimodal and multi-scale fusion with Mamba for remote sensing land classification 

Abstract

The rapid development of remote sensing imaging technology has brought massive and diverse data to remote sensing land classification. How to use the complementarity of multimodal data to improve classification performance has become a research hotspot. In recent years, the Mamba model has been successfully applied in the field of image processing due to its unique architecture and powerful global modeling capabilities.

Among them, the multi-scale visual Mamba model is good at dealing with complex spatial distributions, which is in line with the characteristics of large scale differences and complex orientations of remote sensing objects. To fully leverage the advantages of Mamba model in extracting and fusing remote sensing data features, this paper proposes a multimodal and multi-scale fusion model based on Mamba for remote sensing land classification (M3RS).

Firstly, the model uses a multi-scale spatial encoder to extract features from Light Detection and Ranging (LiDAR) and Synthetic Aperture Radar (SAR) images, and based on the unique data structure of hyperspectral images (HSI), proposes a multi-scale spatial encoder to extract its complex spatial spectral features. Then, a multimodal feature fusion module is proposed by combining cross Mamba and channel concatenation Mamba.

Cross Mamba efficiently fuses multimodal spatial features through interactive state space parameters, while channel concatenation Mamba fully fuses multimodal features by constructing four channel scanning methods. Finally, the model adopts an improved multi-scale feature fusion module to fuse multi-scale features layer by layer and extract classification criteria with high discriminability, which can effectively improve the accuracy of remote sensing land classification.

This article validates the effectiveness of the M3RS classification model through classification experiments conducted on three datasets: Muufl, Houston2013, and Augsburg. The attachment is the open source code researched by the author.