- Notifications
You must be signed in to change notification settings - Fork2
Beyond the Visible: Multispectral Vision-Language Learning for Earth Observation
License
IBM/MS-CLIP
Folders and files
| Name | Name | Last commit message | Last commit date | |
|---|---|---|---|---|
Repository files navigation
Multispectral Vision-Language Learning for Earth Observation
Clive Tinashe Marimo,Benedikt Blumenstiel,Maximilian Nitsche,Johannes Jakubik,Thomas Brunschwiler
IBM Research Europe, IBM Germany
Llama3-MS-CLIP has been accepted at ECML PKDD 2025 🎉
The CLIP model consists of two encoders for text and images. We extended the RGB patch embeddings to multispectral input and initialized the weights of the additional input channels with zeros. During the continual pre-training, the images and texts of each batch are encoded and combined. The loss increases the similarity of matching pairs while decreasing other combinations.
We evaluated Llama3-MS-CLIP with zero-shot classification and text-to-image retrieval results, measured in accuracy (%) ↑ and mAP@100 (%) ↑, respectively.The following Figure compares our model with the OpenCLIP baselines and other EO-VLMs.We applied a smoothed min-max scaling and annotated the lowest and highest scores.Our multispectral CLIP model is outperforming other RGB-based models on most benchmarks.
How to run evaluation on benchmark datasets:
- First clone this repository
git clone https://github.com/IBM/MS-CLIPcd MS-CLIP- Prepare a new environment
python -m venv venvsource venv/bin/activatepip install -e .Alternatively, you can install this repo to another venv with
pip install git+https://github.com/IBM/MS-CLIP.git
To run classification and retrieval on custom images, classes and text queries use theinference.py file. The script automatically downloads the weights from Hugging Face. You just need to provide a path to a folder with Sentinel-2 L2A files (all 12 bands) andclass_names or aquery.
python inference.py --run-classification \ --model-name Llama3-MS-CLIP-Base \ --images ./examples \ --class-names ocean agriculture snow rural# --classes-file classes.txt # Alternative via txt filepython inference.py --run_retrieval\ --model-name Llama3-MS-CLIP-Base \ --images ./examples \ --query"Agricultural fields in a rural area" \ --top-k 3# --queries-file queries.txt # Alternative via txt file
Example output for classification:
╒════╤═══════════════════╤═════════╤═════════╤════════╤═════════╕│ │ Image │ Class │ Ocean │ Snow │ Rural │╞════╪═══════════════════╪═════════╪═════════╪════════╪═════════╡│ 0 │ 282D_485L_3_3.tif │ Ocean │ 0.857 │ 0.000 │ 0.143 │├────┼───────────────────┼─────────┼─────────┼────────┼─────────┤│ 1 │ 637U_59R_1_3.tif │ Rural │ 0.000 │ 0.000 │ 1.000 │├────┼───────────────────┼─────────┼─────────┼────────┼─────────┤│ 3 │ 609U_541L_3_0.tif │ Snow │ 0.000 │ 1.000 │ 0.000 │╘════╧═══════════════════╧═════════╧═════════╧════════╧═════════╛Example output for retrieval:
Image SimilarityQuery Rank Agricultural fields in a rural area 1 637U_59R_1_3.tif 0.243049 2 38D_378R_2_3.tif 0.189760 3 433D_629L_3_1.tif 0.144854 4 282D_485L_3_3.tif 0.011940 5 609U_541L_3_0.tif -0.004457If you like to save the results in a csv, you can provide file path, with--save-path results/your_retrieval_results.csv.
If you installmsclip as a package, you can use the inference functions in python.
frommsclip.inferenceimportrun_inference_retrieval,run_inference_classificationfrommsclip.inference.utilsimportbuild_model# Init Llama3-MS-CLIP from Hugging Facemodel,preprocess,tokenizer=build_model()results=run_inference_classification(model=model,preprocess=preprocess,tokenizer=tokenizer,image_path="path/to/folder",class_names=["class1","class2","class3"])results=run_inference_retrieval(image_path="path/to/folder",queries=["A satellite image of a rural area."])
We provide more information on the expected data structure of the benchmark datasets inDATASETS.md.
To run evaluation on all benchmarking datasets use theevaluation.py file, run:
python evaluation.py# For other models, change the following default values:python evaluation.py \ --model-name Llama3-MS-CLIP-Base \ --pretrained True \ --save-name Llama3-MS-CLIP-Base \ --templates msclip \ --dataset-dir benchmark_datasets \ --batch-size 64 \ --workers 0 \ --precision amp \ --save-path results/This will create a csv file for each dataset showing metrics like accuracy or mean average precision.
For testing Llama3-MS-CLIP on a new dataset, you can add the data loading and processing tosrc/inference/benchmark_tool.pyor run the inference script with--save-path and evaluate the saved results.
Please cite the following paper, if you use the caption dataset and/or the Llama3-MS-CLIP model in your research:
@article{marimo2025beyond, title={Beyond the Visible: Multispectral Vision-Language Learning for Earth Observation}, author={Marimo, Clive Tinashe and Blumenstiel, Benedikt and Nitsche, Maximilian and Jakubik, Johannes and Brunschwiler, Thomas}, journal={arXiv preprint arXiv:2503.15969}, year={2025}}Built with Meta Llama 3.
While the model itself is not based on Llama 3 but OpenCLIP B/16, it is trained on captions generated by a Llama 3-derivative model. Therefore, the model name starts with Llama 3 following its license (https://github.com/meta-llama/llama3/blob/main/LICENSE).
About
Beyond the Visible: Multispectral Vision-Language Learning for Earth Observation
Resources
License
Security policy
Uh oh!
There was an error while loading.Please reload this page.
Stars
Watchers
Forks
Releases
Packages0
Contributors3
Uh oh!
There was an error while loading.Please reload this page.