SCRS-Mamba

Official implementation of SCRS-Mamba for remote sensing scene classification.

• Author: Zaichun Yang

Overview

SCRS-Mamba introduces a Scale-aware State Space Model (SA-SSM) with spatially continuous multi-view scanning for robust remote sensing scene recognition.

This repository provides:

• Training and evaluation scripts based on MMEngine/MMPretrain
• Model and dataset definitions for SCRS-Mamba
• Feature visualization utilities (Fine/Coarse/Fusion)

Environment

This codebase is built on the OpenMMLab ecosystem.

1) Create environment

conda create -n scrsmamba python=3.10 -y
conda activate scrsmamba

2) Install PyTorch

Please install PyTorch following the official instructions for your CUDA version.

3) Install OpenMMLab dependencies

Recommended (with OpenMIM):

pip install -U openmim
mim install "mmcv>=2.0.0,<2.4.0"
pip install "mmengine>=0.8.3,<1.0.0" "mmpretrain>=1.2.0"

4) Install SCRS-Mamba

pip install -e .

5) Install Mamba dependencies

pip install "transformers>=4.39.0"
pip install mamba-ssm causal-conv1d

Data Preparation

This repository follows the common folder structure:

data/
  AID/
  UCMerced_LandUse/
  NWPU-RESISC45/

Prepare train/val splits as plain text lists (relative paths) under datainfo/.

Training

Example: AID (Base) with SA-SSM enabled.

python tools/train.py configs/scrsmamba/scrsmamba_aid_b_sa_ssm.py --amp

Checkpoints and logs will be saved to work_dirs/.

Evaluation

python tools/test.py configs/scrsmamba/scrsmamba_aid_b_sa_ssm.py \
  work_dirs/scrsmamba_aid_b_sa_ssm/best_*.pth

Feature Visualization (Fine/Coarse/Fusion)

python tools/visualization/vis_sa_ssm_features.py \
  --config configs/scrsmamba/scrsmamba_aid_b_sa_ssm.py \
  --checkpoint work_dirs/scrsmamba_aid_b_sa_ssm/best_*.pth \
  --images path/to/image1.jpg path/to/image2.jpg \
  --out-dir outputs/sa_vis \
  --img-size 224

Citation

If you find this work useful, please cite:

@unpublished{yang2026scrsmamba,
  title  = {SCRS-Mamba: Scale-aware and Spatially Continuous Multi-View Scanning Mamba for Remote Sensing Scene Classification},
  author = {Yang, Zaichun},
  year   = {2026},
  note   = {Manuscript under review at iScience}
}

Acknowledgements

This project is built upon the OpenMMLab ecosystem (MMEngine/MMPretrain) and related open-source efforts.