CM3Mamba: Preprocessed multisource remote sensing datasets and model code

This deposit contains preprocessed input datasets and the CM3Mamba model implementation for multisource land cover classification. It accompanies the manuscript "CM3Mamba: Robust multisource land cover classification via three-slot state-space fusion and shared pure-read training" by Jingbo Zhou, Shuran Jia, Shaoqun Qi, Dousheng Zhang, and Zhenping Jiang. The data include 11 × 11 training, validation, and test patches and labels for Houston 2013, Augsburg, and MUUFL Gulfport, together with fixed spatial split coordinates, sample ordering, evaluation subset definitions, and the normalization and PCA parameters used in preprocessing. HSI inputs contain 20 principal components. Auxiliary inputs comprise one LiDAR channel for Houston, four SAR channels and one DSM channel for Augsburg, and two LiDAR channels for MUUFL. Labels use zero-based class indices. The code archive provides the proposed CM3Mamba model (configuration A6), its training entry point, environment requirements, and examples of fused and unimodal inference. The implementation is also available at https://github.com/nanhe-cmd/CM3Mamba. README.md documents the file structure, array conventions, and installation and training commands; dataset_metadata.json provides machine-readable dataset descriptions. Original data sources are the IEEE GRSS 2013 Data Fusion Contest (https://www.grss-ieee.org/community/technical-committees/2013-ieee-grss-data-fusion-contest/), the S2FL Augsburg benchmark (https://github.com/danfenghong/ISPRS_S2FL), and MUUFL Gulfport (https://github.com/GatorSense/MUUFLGulfport). Third-party observations and labels retain their original providers' rights, terms, and citation requirements. The CM3Mamba source code is distributed under its included MIT license.