This repository contains the official implementation of our paper: "DMRS: Long-tailed remote sensing recognition via semantic-aware mixing and diversity experts"
📄 Paper: https://doi.org/10.1016/j.jag.2025.104623
We recommend using Python 3.11. Install dependencies using uv (recommended) or pip:
# Using uv (recommended)
uv sync- PyTorch >= 2.7.1
- torchvision >= 0.22.1
- CLIP (OpenAI)
- PEFT >= 0.15.2
- scikit-learn >= 1.7.0
- matplotlib >= 3.10.3
Organize your remote sensing dataset in the following structure(The program will process it as long-tail data):
dataset/
├── class1/
│ ├── image1.jpg
│ ├── image2.jpg
│ └── ...
├── class2/
│ ├── image1.jpg
│ └── ...
└── ...
Supported datasets:
- NWPU-RESISC45
- RSD46-WHU
- Custom remote sensing datasets
python CLIP_Lora.py \
--dataset_path ./NWPU-RESISC45 \
--imb_type exp \
--imb_factor 0.01 \
--epochs 40 \
--batch_size 16 \
--lr 1e-1python CLIP_Lora.py \
--dataset_path ./NWPU-RESISC45 \
--imb_type exp \
--imb_factor 0.01 \
--MME_loss True \
--num_experts 3 \
--mixrs True \
--epochs 40 \
--batch_size 16 \
--lr 1e-1--dataset_path: Path to your dataset--imb_type: Type of imbalance ('exp' for exponential, 'step' for step)--imb_factor: Imbalance factor (0.01 for severe imbalance)--MME_loss: Enable D-LoRA loss function--num_experts: Number of experts (default: 3)--mixrs: Enable MixSSS data augmentation--lora_r: LoRA rank (default: 12)--lora_alpha: LoRA scaling factor (default: 24)
If you use this code in your research, please cite our paper:
@article{WANG2025104623,
title = {DMRS: Long-tailed remote sensing recognition via semantic-aware mixing and diversity experts},
journal = {International Journal of Applied Earth Observation and Geoinformation},
volume = {141},
pages = {104623},
year = {2025},
issn = {1569-8432},
doi = {https://doi.org/10.1016/j.jag.2025.104623},
url = {https://www.sciencedirect.com/science/article/pii/S1569843225002705},
author = {Yifan Wang and Fan Zhang and Qihao Zhao and Wei Hu and Fei Ma},
keywords = {Long-tail distribution, Remote sensing, Diversity experts, Data augmentation, Foundation models},
}This project is licensed under the MIT License - see the LICENSE file for details.
For questions about the code or paper, please:
- Open an issue on GitHub
- Contact: [2024200827@buct.edu.cn]