Neural network models for contrail detection and segmentation

Introducing an open-source project that implements contrail segmentation neural network models in PyTorch.

These ResUNet models are constructed using fewshot augmented transfer learning, wherein multiple image augmentations are applied to a pre-trained ResUNet model. Consequently, the model can be efficiently fine-tuned using only a small set of labeled satellite images.

To enhance contrail detection through contrail information in Hough space, a new loss function, SR Loss, has been developed, aiming at further optimizing the contrail detection process.

Contrail detection examples

The following set of images shows examples of contrails detected with different models trained with Dice, Focal, and SR loss functions:

The following set of images shows examples of contrails detected from different image sources, which are different from the GOES-16 images that are used in the model training.

Dependencies

PyTorch
segmentation_models_pytorch
opencv
scikit-learn

Setup

It is recommended to use a mamba/conda environment. To setup mamba/conda, follow this tutorial: https://youtu.be/Ket0WUTm5JU?t=47

Create conda environment

mamba create -n contrail python=3.11 -c conda-forge

Install dependencies

For PyTorch users with CUDA 12.1 (recommended):

conda activate contrail
mamba install pytorch pytorch-cuda=12.1 -c pytorch -c nvidia
pip install segmentation-models-pytorch albumentations

For PyTorch users without CUDA:

conda activate contrail
conda install pytorch cpuonly -c pytorch
pip install segmentation-models-pytorch albumentations

Models

Use pre-trained models

You can also download the already trained model weights from: https://surfdrive.surf.nl/files/index.php/s/n1b0L2qfu2PZ6d3

Save the downloaded models in a folder called models under the data directory.

Train models with own data

If you want to train the model, you can train models with following examples:

# train with own dataset, for 1000 epoch, SR loss function
python train.py --dataset own --epoch 1000 --loss sr

python train.py --dataset own --epoch 1000 --loss dice

# train model for 60 minutes
python train.py --dataset own --time 60 --loss focal

Train models with Google contrail dataset

You can also use Google contrail dataset for training the models

Download the data from: https://www.kaggle.com/competitions/google-research-identify-contrails-reduce-global-warming/data
Then put the uncompressed files in google-goes-contrail folder under data directory
Run process_google_data.py to generate mask statistics file mask_stats.csv inside of the previous folder

Then, use the following examples to train models

# train with google dataset, 30 minutes, SR loss
python train.py --dataset google --time 30 --loss sr

# fewshot training, 400 images, 30 minutes
python train.py --dataset google:fewshot:400 --time 30 --loss dice

Detection and visulization

The detect.py provides examples for loading models and detecting contrails in the testing images.

junzis / contrail-seg