Wildfire incidents have surged in frequency and severity in recent years highlighting the need for advanced technologies to predict wildfire behavior early and mitigate its impact. Recent strides in machine learning research, the increased availability of wildfire data, and computational resources have fueled the rise of data-driven approaches in wildfire management. This study aims to advance data-driven methods for predicting wildfire behavior and aid in timely decision-making and resource allocation efforts by adopting a Graph Neural Network (GNN)-based framework for predicting the burned area resulting from a wildfire ignition. GNNs have shown success in handling irregular-sized inputs and capturing the long-range dependencies inherent in geospatial data, such as wildfires, making them a viable alternative to CNNs which impose limitations on geospatial data due to their reliance on fixed-size inputs and local receptive fields. A framework is developed to represent spatial wildfire data and its influencing factors as graphs followed by the development of three distinct GNN models based on different message-passing mechanisms to process the graph-structured data. GNN models outperform CNN-based segmentation models in wildfire prediction, achieving higher AUPRC (0.4787), precision (0.4536), and AUROC (0.9377), and illustrating the efficacy of GNNs in modeling wildfire behavior by effectively capturing spatial dependencies.

In response to the growing incidence and severity of wildfires, this paper presents Vi-Net, a novel hybrid deep learning framework for next-day wildfire spread prediction. By integrating U-Net’s fine-grained spatial segmentation with the global contextual modeling of Vision Transformers (ViT), Vi-Net formulates wildfire spread prediction as a semantic segmentation task. The model is trained on a decade-long (2012–2020) multimodal wildfire dataset that integrates meteorological, topographical, and vegetation features. To address the severe class imbalance inherent in wildfire data, Vi-Net employs a Focal Tversky loss function. Experimental results show that Vi-Net achieves an F1-score of ∼97% and an Intersection over Union (IoU) of ∼94% on test data, significantly outperforming standalone U-Net and ViT models. These findings underscore Vi-Net’s potential to improve wildfire mitigation planning, resource allocation, and emergency response.

Forest fires or wildfires are becoming more prevalent across Canada. They are both beneficial and harmful. They promote forest health and aid ecological processes. However, they play a devastating role in impacting the economy of a nation and also impact the health of humans. Hence, it is important to consider all data sources relevant to forest fires or wildfires. The Canadian Wildland Fire Information System (CWFIS) calculates the danger of forest fires. The Canadian Forest Fire Weather Index (FWI) System is a critical part of CWFIS, which does not consider land vegetation in its calculations. Considering it is the vegetation that burns in a forest fire, it is important to have an insight into what types of vegetation are more prone to fires. Earth observation data for vegetation over land is now available across North America. This research primarily provides an analytical data visualization of the vegetation land cover impacted by and impacting forest fires. We look into open-source vegetation land cover data and provide insights into forest fires or wildfires. A look into the change of vegetation is also provided.

Prevention of forest fires increasingly impacted by climate change is essential to maintain ecological balance, preserve natural resources, prevent economic loss, and protect human and animal life. Data for forest fires is available from multiple sources and is huge. Federated learning can be implemented to distribute the computing across multiple edge devices by saving transmission costs, protecting data privacy, and maintaining security with no single point of failure as local models exist across multiple resources in different geographic regions. The proposed framework extends the Spatio-Temporal Agnostic Subsampling (STAS) technique by distributing the data into multiple computation nodes to leverage federated learning. It was found that the models trained using federated learning on weather data gained on average 0.3 in F1 for classifying the occurrence of fire. This study also demonstrates how to optimally choose the sources of data for either predicting the occurrence of fire or the severity of fire.