Bi-temporal Convolutional Neural Network based on U-Net Architecture for Wildfire Burnt Area Detection. The purpose of this project is to provide more accurate prediction results compared to publicly available products, utilizing a relaxed dataset requirement and achieving higher spatial resolution. The model utilizes two temporal images of the same location, pre and post-wildfire occurrences, in the form of false-color images for better distinction between burnt and unburnt areas.
The train_AttentionUnet.py contains scripts for model training and testing. The neural network's hyperparameters can be adjusted using different parameters. It supports the following models and loss functions:
- U-Net
- Bi-temporal U-Net
- U-Net with attention gate
- Bi-temporal U-Net with attention gate
- Various loss functions
The AttentionUnet.py file contains the architecture of the neural network model. This script defines the structure of the Attention U-Net used for burnt area detection caused by wildfires.
The BADMDataset_set.py script initializes the dataset required for training and testing. It deals with dual temporal images, each having dimensions of 256x256x3, and produces predictions of size 128x128x1.
If you require access to the dataset or have any related inquiries, please feel free to reach out to me via email:
Email: [email protected]