Segment Anything Model for Refugee-Dwelling Extraction (SAM4Refugee) From High-Resolution Satellite Imagery
This repository contains codes for how to segment refugee dwellings (or any other types of buildings) from high-resolution satellite imagery using the SAM-Adapter model.
Feel free to read the preprint for more details.
Gao, Y. (2024, July 16). Leveraging segment anything model in identifying buildings within refugee camps (sam4refugee) from satellite imagery for humanitarian operations. arXiv.org. https://doi.org/10.48550/arXiv.2407.11381
The codes are adapted based on SAM Adapter for training and segment-geospatial for creating prediceted masks in the format of GeoTIFF and polygons in the format of ShapeFile.
These codes can be easily adapted for binary semantic segmentation applications in remote sensing. Feel free to use it for your own applications and implement in your local machine.
Due to data license issues, only data of Kutupalong refugee camp is available. Through the following link, you could download the raw image and lable data for training (small and large), testing and validation. https://drive.google.com/drive/folders/1FgD-E_2RwSeVwkgW2X7JwmmLnMU_Zo2z?usp=sharing
Link to the complete drone image dataset for Kutupalong refugee camp: https://data.humdata.org/dataset/iom-bangladesh-needs-and-population-monitoring-npm-drone-imagery-and-gis-package-by-camp-august-2018
Please Change "path_data" in /run_sam/train.py & inference_noft.py & evaluation.py to your own path.
2. Try “environment.yml” at first to create environment for both data processing and training/inferencing.
It may take quite a long time. (Change to your working directory when running commands if necessary, e.g., conda env create -f /home/yunya/environment.yml)
Link for creating environment by .yml (detailed): https://conda.io/projects/conda/en/latest/user-guide/tasks/manage-environments.html#activating-an-environment Link (short): https://shandou.medium.com/export-and-create-conda-environment-with-yml-5de619fe5a2
Then, follow the instructions in Data_Processing_SAM.ipynb (Remember to change working directory at the beginning) to start processing data and SAM_Adapter_For_Building_Extraction.ipynb to start training/inferencing.
4. Attention: if you download the packages directly from Colab, it is possible that you will miss the “pretrained” folder as shown below. If it happens, please download this folder manually.
CV2 requires Python (3.7<=Python<3.11)
conda create -n sam python==3.10 conda activate sam
conda install -c conda-forge jupyterlab -y conda install -c conda-forge nb_conda_kernels -y conda install ipywidgets -y conda install -c anaconda ipykernel -y python -m ipykernel install --user --name data --display-name "data_sam (3.10)"
conda install tqdm -y conda install rasterio -y conda install scipy -y conda install imagecodecs -y conda install scikit-learn -y conda install scikit-image -y conda install -c conda-forge opencv -y conda install -c conda-forge gdal -y conda install -c conda-forge proj geopandas -y conda install -c conda-forge geopandas -y
conda install pytorch==2.0.1 -y conda install tensorboardX -y conda install -c conda-forge segment-anything -y pip install /home/yunya/anaconda3/envs/mmcv-1.7.0.tar.gz (change path if necessary) conda install leafmap -y
Avaialable input prompts:
- ('--config', default="configs/config_sam_vit_h.yaml", help="use the hyperparameters provided by SAM-Adapter")
- ('--data', default=None, help="different datasets")
- ('--upsample', default="1024", help="1024 or SR")
- ('--size', default="small", help="small or large")
- ('--uptype', default="", help="nearest bilinear EDSR")
- ('--epoch', default=15, help="epochs for training")
- ('--model_save_epoch', default=999, help="the interval of saving trained models.")
- ('--inference_save_epoch', default=1, help="the interval of saving trained models")
- ('--thres', default=0.5, help="the threshold to determine the binary map")
Architecture of SAM-Adapter
Please check the original SAM-Adapter through this GitHub repository: https://github.com/tianrun-chen/SAM-Adapter-PyTorch (Credits to Tianrun Chen)
Study site: Djibo Refugee Camp, Burkina Faso
The benefits of upscaling for satellite images in building extraction for refugee camps or other cases
An example in the Minawao refugee camp showcases the influence of upscaling on the performance of SAM-Adapter.
A. Original image;
B. Ground truth;
C. Upscaled image;
D. Predicted masks from SAM-Adapter
