Assessment of deep learning models integrated with weather and environmental variables for wildfire spread prediction and a case study of the 2023 Maui fires

🔥 Research Overview

Predicting the spread of wildfires is essential for effective fire management and risk assessment. With the fast advancements of artificial intelligence (AI), various deep learning models have been developed and utilized for wildfire spread prediction. However, there is a lack of understanding of the advantages and limitations of different models, and it is also unclear how deep learning-based fire spread models can be compared with traditional physics-informed fire models. In this work, we assess the ability of five deep learning models integrated with weather and environmental variables for predicting fire spread based on over ten years of wildfire data in Hawaii. We further use the 2023 Maui fires as a case study to compare deep learning models with a widely-used and physics-informed and semi-empirical fire spread model, FARSITE.

🗂️ Repository Overview

This repository is designed to facilitate preprocessing, training, testing, and analysis of wildfire spread prediction models for the Hawaii state and Maui region.

🔍 Repository Organization

• 01_DataPreprocessing.ipynb: Jupyter Notebook for loading, cleaning, and filtering the Maui fire dataset using geospatial libraries and clustering algorithms.
• 02_Train.py: Python script for training the wildfire prediction model.
• 03_Test.ipynb: Jupyter Notebook demonstrating model testing and performance evaluation.
• 04_Train_Test_FARSITE_Input.ipynb: Notebook for training and testing with FARSITE input data integrations.
• 05_Visualization.ipynb: Notebook for visualizing model outputs and geospatial data.
• 06_FeatureImportance.ipynb: Notebook analyzing feature importance for model interpretability.
• UNets.py: Source code defining the U-Net architecture for image-based fire detection.
• LSTM_modules.py: Python module containing utility functions and classes for LSTM models.
• LSTM.py: Script implementing LSTM-based temporal prediction models.
• ConvLSTMs.py: Module combining convolutional and LSTM layers for spatiotemporal modeling.
• FARSITE.py: Wrapper and utility functions for integrating with FARSITE fire simulation software.
• DatasetBuilder.py: Script for constructing and preprocessing custom datasets.
• Data/: Directory containing raw and processed sample datasets.

📁 Note on Data Availability

Due to GitHub’s file size limitations, only a sample of the training dataset is provided in this repository. If you are interested in accessing the full dataset, please contact us via email at 📩jiyeonki@buffalo.edu.