Date of Award

5-2026

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Mathematics and Systems Engineering

First Advisor

Nezamoddin Nezamoddini-Kachouie

Second Advisor

Robert van Woesik

Third Advisor

Jian Du

Fourth Advisor

Ryan White

Abstract

Wildfire ignition risk in the northern Los Angeles region is shaped by distinct ecological mechanisms that vary with ignition source and landscape context. However, most existing predictive frameworks operate at spatial and temporal scales that obscure the fine-scale structure of fuels, topography, and infrastructure that governs ignition processes locally. This thesis develops and applies a progression of machine-learning models to characterize both natural and anthropogenic ignition risk, moving from a coarse preliminary analysis to a fine-scale, spatially hierarchical investigation. The final model, a hierarchical pooling spatiotemporal convolutional neural network operating at approximately 1 km spatial and 16-day temporal resolution, substantially outperformed all baseline architectures. It achieved a test AP-lift of 69.70× for natural ignitions, compared to the next-best result of 45.78×. Model behavior was further interpreted using branch-level Shapley attribution, habitat-stratified predictor contrasts, iii and a data-driven classification of hydroclimate regimes. For natural ignitions, the analyses revealed a clear mechanistic typology of ignition environments. Conifer forest systems are consistently fire-prone during summer, driven by synoptic-scale atmospheric demand that removes moisture from an otherwise continuously available fuel bed. In contrast, desert shrub systems are limited by fuel availability and depend on episodic moisture inputs to generate sufficient vegetation to carry fire. The hydroclimate regime classification identified five spatially coherent coupling classes, revealing substantial within-habitat heterogeneity not captured by static land cover categories. Anthropogenic ignitions exhibited a markedly different spatial pattern. Predicted risk was concentrated along road networks and transmission corridors intersecting vegetated terrain, rather than within high-elevation montane interiors. This produced two largely distinct risk surfaces, effectively expanding the fire niche into habitats that are rarely affected by natural ignitions. The hydroclimate analysis further showed that anthropogenic ignitions disrupt the moisture-driven coupling structure characteristic of natural fire regimes, replacing it with an infrastructure-anchored signature across much of the study domain. Taken together, these results demonstrate that natural and anthropogenic wildfire risks in this landscape are governed by fundamentally different processes. As such, they require distinct monitoring and management strategies, a distinction that is likely to become increasingly important as infrastructure expansion and climate change continue to reshape ignition dynamics across Southern California.

Download

Available for download on Tuesday, November 09, 2027

Share

COinS