Candidate:Manavjit Dhindsa
Date: Jan 13, 2025
Time: 1:30pm
Location: Teams
Supervisor: Dr. Sagar Naik
Abstract:
Forest fires are becoming more prevalent than ever, and their intensity and frequency are expected only to increase owing to climate change and environmental degradation. These fires severely threaten the economy, human lives, and infrastructure. Therefore, effective management of wildfires is of utmost importance, and accurately predicting the wildfire spread lies at the core of it. Reliable predictions of fire spread not only provide insights about the at-risk regions but also help in planning several mitigation activities including resource allocation and evacuation planning. This thesis introduces Vi-Net, a hybrid deep learning model, which integrates the localized precision of U-Net with the global contextual awareness of Vision Transformers (ViT) to predict next-day wildfire spread with remarkable accuracy.
This study utilizes an extensive multimodal dataset that accumulates data from different sources across the United States from 2012 to 2020 incorporating critical factors such as topographical, meteorological, anthropological, and vegetation indices. These elements are vital for modeling the complex dynamics of wildfire spread. A significant challenge in this domain is the class imbalance as the “fire” points are generally less than the “non-fire” points. The dataset used in this study had “fire” regions less than 5% of the total data. To address this issue, advanced loss functions are employed, tested, and compared, to prioritize accurate segmentation of fire areas in the regions while minimizing false negatives. Focal Tversky Loss (FTL) performs better out of the four loss functions and gives the best results for the task of wildfire spread prediction. It modifies the focus toward hard-to-predict regions and crucial boundaries, enhancing the model’s predictive accuracy and reliability in practical applicability.
Vi-Net addresses the complexities in modeling fire dynamics by synergizing the strengths of U-Net and ViT. U-Net excels in extracting fine-grained spatial details, crucial for segmenting localized fire areas accurately. In contrast, ViT captures global dependencies through its self-attention mechanism, enabling the model to recognize patterns and dependencies over larger geographical extents. The integration of U-Net and ViT in Vi-Net allows for the model to achieve high precision and recall, effectively balancing the sensitivity and specificity needed in wildfire predictions. The dual approach allows the model to process both detailed local information and global context, making it exceptionally capable of identifying and predicting fire spread.
Experimental results showcase the superiority of Vi-Net over traditional models, achieving an F1 Score of 97.25% and an IoU of 0.9415 on the test dataset. These metrics highlight its capability to accurately capture both localized fire patches and long-range dependencies while avoiding overprediction. It validates the capability of the model to capturethe intricate interdependency between different environmental factors and the dynamic complexities of wildfire spread, offering nuanced predictions.
In addition to predictive modeling, this research extends its practical applicability by integrating the predicted fire masks into an optimized A*(OA*) algorithm for safe path planning. The OA* algorithm generates a feasible path between two points considering all the safety constraints. This facilitates efficient evacuation routes and resource allocation planning by providing actionable insights for emergency response teams and wildfire management departments. Qualitative and quantitative analyses confirm the hybrid model’s efficacy, with visualizations demonstrating Vi-Net’s ability to preserve spatial detail while capturing broad environmental contexts, and path planning results illustrating the model’s robustness and reliability.
This research sets a new benchmark in wildfire prediction by addressing challenges such as class imbalance, loss function optimization, and the integration of prediction out- put into path planning algorithms. This establishes the hybrid Vi-Net model as a critical tool in understanding and accurately predicting the nature of fire spread under the influence of varying environmental factors. This work also highlights the potential of hybrid deep-learning architectures to enhance predictive capabilities and combat the limitations of individual models. The integration of prediction results with the OA* algorithm directly contributes to risk mitigation strategies in real-world applications and provides a comprehensive framework for proactive wildfire management.”