Metrics
1 Record Views
Dissertation
Building Labeled Satellite Imagery Datasets and Benchmarking Deep Learning Models for Wildfire Detection and Mapping
University of West Florida Libraries
Doctor of Philosophy (PHD), University of West Florida
2026
Abstract
Detecting and mapping wildfire damage from satellite imagery using deep learning is constrained by the scarcity of large, labeled, geographically diverse training datasets. This dissertation addresses that bottleneck through three integrated lines of work.First, we develop two fully programmatic pipelines that convert open-access Sentinel-2 imagery and public fire perimeter records into model-ready datasets. The California Wildfire GeoImaging Dataset provides large-scale before-and-after tile pairs spanning hundreds of California wildfires over multiple years. CalFireSeg-50 covers dozens of large fires across four temporal stages, using conditional labeling to produce spatially precise pixel-level annotations.
Second, we benchmark deep learning architectures for wildfire detection at two levels of spatial detail. For image-level classification, we show that early fusion of pre- and post-fire imagery outperforms both unitemporal and Siamese baselines. For pixel-level segmentation, we evaluate CNN, hybrid, and transformer architectures in unitemporal and bitemporal settings. Bitemporal approaches consistently improve segmentation quality, with the strongest models demonstrating substantial gains over their unitemporal counterparts.
Third, we explore whether diffusion-based generative models can synthesize realistic post-fire imagery from burn masks to augment limited training data. An inpainting pipeline conditioned on pre-fire context produces spatially and spectrally plausible outputs, and vision-language-model prompting proves competitive with the same pipeline.
Together, these contributions demonstrate that an integrated approach combining automated dataset construction, systematic temporal benchmarking, and targeted generative augmentation can meaningfully advance deep-learning-driven wildfire mapping.
Details
- Title
- Building Labeled Satellite Imagery Datasets and Benchmarking Deep Learning Models for Wildfire Detection and Mapping
- Resource Type
- Dissertation
- Contributors
- K. Brent Venable (Committee Chair)Derek Morgan (Committee Member)Arash Mahyari (Committee Member)Steve Norman (Committee Member)
- Publisher
- University of West Florida Libraries
- Format
- pdf
- Number of pages
- 219
- Copyright
- Permission granted to the University of West Florida Libraries by the author to digitize and/or display this information for non-profit research and educational purposes. Any reuse of this item in excess of fair use or other copyright exemptions requires the permission of the copyright holder.
- Identifiers
- 99381761451106600
- Academic Unit
- Intelligent Systems and Robotics; Hal Marcus College of Science and Engineering
- Language
- English
- Awarding Institution
- University of West Florida; Doctor of Philosophy (PHD)
- Theses and Dissertations
- Doctor of Philosophy (PHD), University of West Florida