FarSLIP: Discovering Effective CLIP Adaptation for Fine-Grained Remote Sensing Understanding
**Paper**: [FarSLIP: Discovering Effective CLIP Adaptation for Fine-Grained Remote Sensing Understanding](https://huggingface.co/papers/2511.14901)
**Code**: [https://github.com/NJU-LHRS/FarSLIP](https://github.com/NJU-LHRS/FarSLIP)
## Introduction
We introduce FarSLIP, a vision-language foundation model for remote sensing (RS) that achieves fine-grained vision-language alignment. FarSLIP demonstrates state-of-the-art performance on both fine-grained and image-level tasks, including open-vocabulary semantic segmentation, zero-shot classification, and image-text retrieval.
We also construct MGRS-200k, the first multi-granularity image-text dataset for RS. Each image is annotated with both short and long global-level captions, along with multiple object-category pairs.
## Checkpoints
You can download all our checkpoints from [Huggingface](https://huggingface.co/ZhenShiL/FarSLIP), or selectively download them through the links below.
| Model name | Architecture | OVSS mIoU (%) | ZSC top-1 accuracy (%) | Download |
|-------------|--------------|---------------|-------------------------|----------------|
| FarSLIP-s1 | ViT-B-32 | 29.87 | 58.64 | [FarSLIP1_ViT-B-32](https://huggingface.co/ZhenShiL/FarSLIP/resolve/main/FarSLIP1_ViT-B-32.pt?download=true) |
| FarSLIP-s2 | ViT-B-32 | 30.49 | 60.12 | [FarSLIP2_ViT-B-32](https://huggingface.co/ZhenShiL/FarSLIP/resolve/main/FarSLIP2_ViT-B-32.pt?download=true) |
| FarSLIP-s1 | ViT-B-16 | 35.44 | 61.89 | [FarSLIP1_ViT-B-16](https://huggingface.co/ZhenShiL/FarSLIP/resolve/main/FarSLIP1_ViT-B-16.pt?download=true) |
| FarSLIP-s2 | ViT-B-16 | 35.41 | 62.24 | [FarSLIP2_ViT-B-16](https://huggingface.co/ZhenShiL/FarSLIP/resolve/main/FarSLIP2_ViT-B-16.pt?download=true) |
## Dataset
FarSLIP is trained in two stages.
+ In the first stage, we use the [RS5M](https://github.com/om-ai-lab/RS5M) dataset. A quick portal to the RS5M dataset: [link](https://huggingface.co/datasets/omlab/RS5M).
+ In the second stage, we use the proposed MGRS-200k dataset, which is available on [Huggingface](https://huggingface.co/datasets/ZhenShiL/MGRS-200k).
Examples from MGRS-200k
## Usage / Testing
Below is a sample usage for zero-shot scene classification, taken directly from the [official GitHub repository](https://github.com/NJU-LHRS/FarSLIP#zero-shot-scene-classification).
### Zero-shot scene classification
+ Please refer to [SkyScript](https://github.com/wangzhecheng/SkyScript?tab=readme-ov-file#download-benchmark-datasets) for scene classification dataset preparation, including 'SkyScript_cls', 'aid', 'eurosat', 'fmow', 'millionaid', 'patternnet', 'rsicb', 'nwpu'.
+ Replace the `BENCHMARK_DATASET_ROOT_DIR` in `tests/test_scene_classification.py` to your own path.
+ Run testing (e.g. FarSLIP-s1 with ViT-B-32):
```
python -m tests.test_scene_classification --model-arch ViT-B-32 --model-name FarSLIP1 --force-quick-gelu --pretrained checkpoints/FarSLIP1_ViT-B-32.pt
```
Comparison of zero-shot classification accuracies (Top-1 acc., %) of different RS-specific CLIP variants across multiple benchmarks.
## Citation
If you find our work is useful, please give us ⭐ in GitHub and consider cite our paper:
```tex
@article{li2025farslip,
title={FarSLIP: Discovering Effective CLIP Adaptation for Fine-Grained Remote Sensing Understanding},
author={Zhenshi Li and Weikang Yu and Dilxat Muhtar and Xueliang Zhang and Pengfeng Xiao and Pedram Ghamisi and Xiao Xiang Zhu},
journal={arXiv preprint arXiv:2511.14901},
year={2025}
}
```
