RS-MTDF: Multi-Teacher Distillation and Fusion for Remote Sensing Semi-Supervised Semantic Segmentation

This is the official PyTorch implementation of the RS-MTDF method.

Environment Preparation

First, create a new Conda environment using the environment.yaml provided in the repository:

conda env create -f environment.yaml

Once the environment is created, activate it using the following command:

conda activate Semi

Data and Pretrained Model Preparation

Download Datasets

• LoveDA: Train | Val
• ISPRS Potsdam: Data
• DeepGlobe Land Cover: Data

Expected directory structure:

./dataset
├── loveda
├── potsdam
├── deepglobe

Data Splits

We follow the experimental setting from DWL. After cropping and splitting, the structure should be:

├── splits
    ├── ./dataname(eg. loveda)
        ├── 1
            ├── labeled.txt
            ├── unlabeled.txt
        ├── 5
            ├── labeled.txt
            ├── unlabeled.txt
        ├── 10
            ├── labeled.txt
            ├── unlabeled.txt
        ├── test.txt
        ├── val.txt

Pre-trained Encoders

Download the following pretrained models and place them in the ./pretrained directory:

DINOv2-Small | DINOv2-Base | Clip-L

Expected directory structure:

├── ./pretrained
    ├── dinov2_small.pth
    ├── dinov2_base.pth
    ├── ViT-L-14.pt

🚀 Training

Once the datasets and pretrained models are prepared, start training with:

sh scripts/train.sh <num_gpu> <port>

You can modify scripts/train.sh to change the training settings, and adjust the config file for learning rate and other hyperparameters.

Eval Process

All evaluation scripts are located in the eval/ directory.

citation

If you find this project helpful, please consider citing:

@misc{song2025rsmtdfmultiteacherdistillationfusion,
      title={RS-MTDF: Multi-Teacher Distillation and Fusion for Remote Sensing Semi-Supervised Semantic Segmentation}, 
      author={Jiayi Song and Kaiyu Li and Xiangyong Cao and Deyu Meng},
      year={2025},
      eprint={2506.08772},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2506.08772}, 
}