Semantic Segmentation in Action: Cross Domain Strategies for Wildfire Risk Reduction
Loading...
Date
2024-05
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
The Ohio State University
Abstract
The increasing severity and frequency of wildfires globally necessitate innovative and effective risk reduction strategies. This thesis presents SHINE (Susceptibility and Human Impact of Natural Emergencies), a groundbreaking tool developed to enhance wildfire risk mitigation and adaptation planning. By harnessing national data from the Centers for Disease Control and Prevention, Environmental Public Health Tracking Network, and the CDC/ATSDR Social Vulnerability Index (SVI), SHINE aims to refine disaster risk reduction strategies with a focus on enabling targeted, community- specific interventions. A cornerstone of this research is the creation of a novel dataset for semantic segmentation, derived from images captured by the HPWREN camera network. The dataset was developed employing a detection guided segmentation technique, utilizing the Segment Anything Model (SAM). Furthermore, the study explores the advanced technique of model distillation to enhance a yolov8 segmentation model. The distillation process leverages foundational models like DINO and SAM, the FLAME dataset,which encompasses drone-captured fire imagery, alongside self-supervised learning techniques, to imbue the yolov8 model with enhanced capabilities in fire detection. By integrating cross-domain strategies and cutting-edge deep learning methodologies, this study offers promising directions for building resilience against wildfire threats.
Description
Keywords
artificial intelligence, computer vision, deep learning, semantic segmentation, social vulnerability, wildfires