Climate change has been correlated with a significant increase in the occurrence of wildfires as well as global forest loss, resulting in a devastating $42$-$44$ percent loss in North America alone between 2001 and 2019 and in $2022$ alone over $6.6$ million hectares of tree cover was lost. Studies have found factors such as increased fuel acridity and flammability and elevated levels of Persistent Positive Anomalies (PPAs) to be contributing to fire ignitions with implications of a further upward trend as temperatures continue to rise globally. The detrimental impact is not limited to economic cost but also physical and mental health-related, disproportionately affecting elderly and low-income populations.
As a result of these often catastrophic events becoming more frequent, wildfire prediction and modeling have become crucial in risk mitigation and targeted resource allocation, providing key insights on fire behavior to allow for early planning efforts. Simulation of wildfire spread from remote sensing data in particular offers an advantage in real-world applications due to its accessibility and extensibility, with potential utility in the timely mapping of wildfire activity to aid in emergency response and management.
Responding to the call for action to leverage machine learning in combating climate change challenges, a plethora of tools, datasets, and techniques have emerged, enhancing innovation in environmental science. For fire detection and prediction, deep learning is a promising approach due to the abundance of high-performing network architectures and efficiency in solving complex problems with large feature spaces such as the remote sensing data from satellite imagery. While deep learning models have generally achieved great performance on fire detection tasks with large-scale data, results from spread prediction efforts, however, have been highly inconsistent across dimensions including temporality, region, and size of the data used, performance metrics, and formulation of the problem. This reflects the difficulty of the task as well as the need for a broader search for a generalized method capable of informing decisions across areas on a continental scale - we use vision transformers in a U-shaped formation combined with a more comprehensive data set to tackle this problem.
In this study, we tailor various state-of-the-art segmentation methodologies for large-scale wildfire spread modeling, utilizing remote sensing data. Our approach involves a comprehensive examination of input features, loss functions, deep network architectures, and mechanisms for attention and representation modulation. To evaluate the generalizability of our model, we have extended the scope of the Next Day Wildfire Spread (NDWS) dataset. Originally, NDWS encompassed data from 2012 to 2020 within the United States. We have expanded this dataset by including fire incidents from 2012 to 2023 across North America. This expansion has nearly doubled the dataset’s size. Building on our analyses, we introduce a wildfire prediction model based on a symmetrical encoder-decoder architecture, employing Swin-Unet with spatial attention and focal modulation. This model is designed to predict fire locations for t+1 day.
Download the full paper here