This is the official implementation of the paper Debris Segmentation using Post-Hurricane Aerial Imagery. To quickly experiment with the proposed debris segmentation model (called CLIPSeg-debris), we recommend trying this Colab demo.

Our debris segmentation dataset comprises 1,200+ manually annotated aerial RGB images from three major hurricane events:

• Hurricane Ian (2022): 685 images from Estero Island and Fort Myers, FL
• Hurricane Ida (2021): 273 images from Grand Isle, Port Fourchon, and Golden Meadow, LA
• Hurricane Ike (2008): 284 images from Galveston Island, TX

Key Features:

• Multiple annotators with consensus labeling to reduce human bias
• Balanced dataset with debris-positive and debris-free samples
• Diverse environmental conditions and sensor resolutions (5-50cm)
• Three debris classes: no debris, low-density debris, high-density debris

CLIPSeg-debris achieves state-of-the-art performance when evaluated on Hurricane Ida imagery (entirely excluded from training):

Key Achievements:

• Best overall performance: 0.86 Dice score and 0.83 IoU
• Superior precision: Lowest false positive rates across all debris classes
• Robust generalization: Tested on entirely unseen hurricane event and geographic region

# clone project
git clone https://github.com/Way-Yuhao/CLIPSeg-debris
cd CLIPSeg-debris

# [OPTIONAL] create conda environment
conda create -n clipseg python=3.10
conda activate clipseg

Install PyTorch according to instructions: https://pytorch.org/get-started/.
Install the remaining dependencies using pip:

# install requirements
pip install -r requirements.txt

• Navigate to Design-Safe to download the annotated debris dataset proposed in this work.
• Download model weights from Hugging Face.

Fine-tune the CLIPSeg model on the dataset using the provided training script.

# training on GPU 0, for example: 
python ./src/train.py experiment=clipseg_finetune trainer.devices=[0]

Experiment configuration can be found from configs/experiment/

Test the fine-tuned model on the dataset using the provided testing script.

python ./src/eval.py experiment=retest trainer.devices=[0] name=DIR_NAME ckpt_path=CKPT_PATH

Replace DIR_NAME with the name of the directory where the results will be saved, and CKPT_PATH with the path to the checkpoint file of the fine-tuned model.

Predict the segmentation mask for a single image using the fine-tuned model.

python ./src/eval.py trainer.devices=[0] data.query_images_dir=INPUT_DIR ckpt_path=CKPT_PATH name=DIR_NAME

Replace INPUT_DIR with the directory containing the input images, CKPT_PATH with the path to the checkpoint file of the fine-tuned model, and DIR_NAME with the name of the directory where the results will be saved.

If you find this code useful in your research, please consider citing:

@article{amini2025debris,
  title={Debris segmentation using post-hurricane aerial imagery},
  author={Amini, Kooshan and Liu, Yuhao and Padgett, Jamie Ellen and Balakrishnan, Guha and Veeraraghavan, Ashok},
  journal={Computer-Aided Civil and Infrastructure Engineering},
  year={2025},
  publisher={Wiley Online Library}
}