MsRE: Towards Efficient Remote Sensing Segmentation via Vision Foundation Models

Bin Wang ¹, Shun Lv ¹, Zhi Li ¹, Fei Deng ², Yiguang Liu^1†

¹ Sichuan University ² Chengdu University of Technology

^† Corresponding author.

Network Overview

🔍️🔍️ NEWS

• [2026/07/14] 🤗🤗 The Training Code has been updated.
• [2026/07/06] 🎉🎉 Our paper is accepted by IEEE TGRS (Link)!
• [2026/05/04] 🧨🧨 Update project page && readme.md.
• [2026/04/13] ✨✨ Init Repo.

📖 Clone Repo

---

We add mmsegmentation as our repository submodule .

So, one should clone this repository use the script as follows:

clone repository

git clone --recurse-submodules https://github.com/woldier/MsRE

Tips

If one already cloned the project and forgot --recurse-submodules,

 # cloned the project and forgot clone submodules 🥲🥲
 git clone https://github.com/woldier/Bridge 

 # initialize and update each submodule in the repository 🥰🥰
 git submodule update --init

after that, we link mmsegmentation/mmseg $\to$ mmseg:

soft link

$ pwd work_dir/MsRE
ln -s mmsegmentation/mmseg mmseg

🛠️️ 1. Creating Virtual Environment

This repo use python-3.8, for nvcc -v with cuda >= 11.6.

torch 2.1.1, cuda 12.1, mmcv 2.1.0, mmengine 0.9.1

Install script

conda create -n  MsRE  python==3.8 -y
conda activate MsRE


pip install torch==2.1.2+cu121  torchvision==0.16.2+cu121 -f https://download.pytorch.org/whl/torch_stable.html
# for CN user use follow script
pip install torch==2.1.2+cu121  torchvision==0.16.2+cu121 -f https://mirrors.aliyun.com/pytorch-wheels/cu121/  

pip install mmcv==2.1.0 mmengine==0.9.1 -f https://download.openmmlab.com/mmcv/dist/cu121/torch2.1/index.html

pip install -r submodule-mmseg/requirements/runtime.txt

Installation of the reference document refer:

Torch and torchvision versions relationship.

📂 2.Preparation of datasets

We selected Postsdam, Vaihingen and LoveDA as benchmark datasets.

2.1 Download of datasets

ISPRS Potsdam

The Potsdam dataset is for urban semantic segmentation used in the 2D Semantic Labeling Contest - Potsdam.

The dataset can be requested at the challenge homepage. The '2_Ortho_RGB.zip' and '5_Labels_all_noBoundary.zip' are required.

ISPRS Vaihingen

The Vaihingen dataset is for urban semantic segmentation used in the 2D Semantic Labeling Contest - Vaihingen.

The dataset can be requested at the challenge homepage. The 'ISPRS_semantic_labeling_Vaihingen.zip' and 'ISPRS_semantic_labeling_Vaihingen_ground_truth_eroded_COMPLETE.zip' are required.

LoveDA

The data could be downloaded from Google Drive here.

Or it can be downloaded from zenodo, you should run the following command:

loveda download

cd /{your_project_base_path}/Bridge/data/LoveDA

# Download Train.zip
wget https://zenodo.org/record/5706578/files/Train.zip
# Download Val.zip
wget https://zenodo.org/record/5706578/files/Val.zip
# Download Test.zip
wget https://zenodo.org/record/5706578/files/Test.zip

2.2 Data set preprocessing

Place the downloaded file in the corresponding path The format is as follows:

file structure

MsRE/
├── data/
│   ├── LoveDA/
│   │   ├── Test.zip
│   │   ├── Train.zip
│   │   └── Val.zip
├── ├── Potsdam_RGB/
│   │   ├── 2_Ortho_RGB.zip
│   │   └── 5_Labels_all_noBoundary.zip
├── ├── Vaihingen_IRRG/
│   │   ├── ISPRS_semantic_labeling_Vaihingen.zip
│   │   └── ISPRS_semantic_labeling_Vaihingen_ground_truth_eroded_COMPLETE.zip

after that we can convert dataset:

details

• Potsdam

python tools/convert_datasets/potsdam.py data/Potsdam_RGB/ --clip_size 512 --stride_size 512 -o data/potsdam

• Vaihingen

python tools/convert_datasets/vaihingen.py data/Vaihingen_IRRG/ --clip_size 512 --stride_size 256 -o data/vaihingen

• LoveDA

python tools/convert_datasets/loveda.py data/LoveDA/ --tmp_dir data -o data/loveda

🔥 3. Training

training scripts

# loveda
python tools/train.py configs/msre/msre_dinov3_vit-large_segmentor_1xb2-amp-40k_loveda-512x512.py

# potsdam
python tools/train.py configs/msre/msre_dinov3_vit-large_segmentor_1xb2-amp-40k_potsdam-512x512.py

# vaihingen
python tools/train.py configs/msre/msre_dinov3_vit-large_segmentor_1xb2-amp-40k_vaihingen-512x512.py

Acknowledgements

This project is built upon OpenMMLab. We thank the OpenMMLab developers.

Citation

If you use Geoad in your research, please cite:

        @ARTICLE{11599658,
          author={Wang, Bin and Lv, Shun and Li, Zhi and Deng, Fei and Liu, Yiguang},
          journal={IEEE Transactions on Geoscience and Remote Sensing},
          title={MsRE: Towards Efficient Remote Sensing Segmentation via Vision Foundation Models},
          year={2026},
          volume={},
          number={},
          pages={1-1},
          keywords={Modeling;Remote sensing;Semantic segmentation;Training;Tuning;Decoding;LoRa;Visualization;Vegetation;Head;Vision Foundation Models;Semantic Segmentation;Parameter-Efficient Fine-Tuning;Remote Sensing},
          doi={10.1109/TGRS.2026.3711219}
        }