Master's Thesis Defense - Caitlin LaNeve

Candidate: Caitlin LaNeve

Committee Chair: 
Ed Oughton, PhD, Geography and Geoinformation Science

Committee Members: 
Matt Rice, PhD, Geography, and Geoinformation Science
Konrad Wessels, PhD, Geography and Geoinformation Science

Title: Hazardous Material Detection Using Multi-Method Hyperspectral Analysis

Abstract: Wildfires present complex challenges for disaster response, particularly in the rapid identification of airborne and structural hazards. This study utilizes Airborne Visible/Infrared Imaging Spectrometer-3 hyperspectral imagery from the 2025 Palisades, California wildfires to evaluate material transformations and hazardous signatures in support of the National Aeronautics and Space Administration Disaster Response and Coordination System. A multi-method workflow (Spectral Angle Mapper, standard and iterative matched filter, whitened matched filter, sparse unmixing via Least Absolute Shrinkage and Selection Operator and Sparse Unmixing by Variable Splitting and Augmented Lagrangian was applied across pre- and post-fire temporal pairs targeting chrysotile and amphibole asbestos, asphalt, concrete, fire-affected vegetation, and forsterite, the thermally transformed product of chrysotile. Vegetation removal exposed built surfaces: asphalt detections increased by 36.3 percentage points under standard matched filter, while coarse chrysotile increased nearly ninefold under iterative matched filter, which suppresses the dominant chrysotile background before computing final scores. Spectral library verification confirmed only 24.2% of matched filter-flagged chrysotile pixels as spectrally consistent with chrysotile, with the remainder matched to asphalt, concrete, and tremolite, underscoring false-positive risk in complex post-fire scenes. Method sensitivity diverged substantially, with fine chrysotile detection rates ranging from 0.2% to 14.1% depending on approach. Five chrysotile and four forsterite hotspots were identified; the largest chrysotile hotspot overlapped 1,944 residential parcels of which 74.9% predate 1980, and both forsterite hotspots within the co-registered scene extent co-located with pre-1980 majority parcels (69.1% and 75.5%, respectively). These findings demonstrate that multi-method hyperspectral analysis can deliver hazard maps for disaster response while exposing the limitations of single-method detection in mixed post-fire environments.