https://sites.google.com/view/rainforest-challenge/multiearth-2023
Rapid Deforestation and Burned Area Detection using Deep Multimodal Learning on Satellite Imagery
Download all provided datasets
./azcopy cp "https://rainforestchallenge.blob.core.windows.net/multiearth2023-dataset-final/?sv=2021-10-04&st=2023-04-11T13%3A09%3A30Z&se=2023-06-12T13%3A09%3A00Z&sr=c&sp=racwdxltf&sig=33jltVx0OUwTPNS4PK%2F%2FZBntwjJaqAHdocj%2BpHRDp88%3D" data/multiearth2023-dataset-final --recursive
Install the required dependencies
pip install -r requirements.txt
Extract each relevant indvidual bands from the *.nc files for the flexible data loader.
python src/extract_sat_images.py
Train the final deforestation segmentation models
cd src
./forest.sh
Create the binary deforestation masks with ForestSubmission.ipynb for the test submission then upload the zipped images.
Each experiment takes 2-3 hours using 3 NVIDIA Quadro RTX 6000 cards. Our best single model achieves 90.4 Pixel Accuracy, 0.871 F1 and 0.792 IoU on the unseen test area. Training all the models with small configuration changes and blending raw predictions before binarization gives small performance boost up to 90.8 Pixel Accuracy, 0.876 F1 and 0.799 IoU.
We would like to thank for Philipp, Pascal, Christof for sharing their solution https://github.com/pascal-pfeiffer/kaggle-rsna-2022-5th-place Their easy to configure framework was useful for training the binary segmentation models.