Classes & TalksWednesday, November 18
Dissertation Proposal: Amirhossein Montazeri
Wednesday, November 18 · 12 p.m. · City Center Plaza (CCP)
- When
- Wednesday, November 18 · 12 p.m.
- Where
- City Center Plaza (CCP)777 West Main Street · Boise, ID · 83702
About this event
Title: Soil Burn Severity Prediction Using Foundation Models, Deep Learning, Remote Sensing, and Environmental Data
Program: Computing Ph.D.
Committee Chair: Mojtaba Sadegh
Committee: Mojtaba Sadegh, Yu Zhang, Xinyi Zhou
Abstract:
Burn severity is fundamental to post-fire impact assessment and emergency response. Burn
severity is typically divided into two components: vegetation burn severity (VBS), which
concerns fire impacts on vegetation and plant community structure, and soil burn severity
(SBS), which governs the hydrologic and geomorphic response of burned watersheds.
Operational SBS maps still rely on field recalibration of satellite-derived VBS products by
Burned Area Emergency Response (BAER) teams. This dependency introduces a latency
that frequently allows post-fire storms to trigger floods and debris flows before SBS maps
are available. This dissertation develops a data-driven, physically interpretable framework
for predicting SBS directly from multispectral remote sensing and environmental
covariates across the western US. The framework is built on the BurnSAVe dataset, a
harmonized benchmark of 532 large fires (2012–2024) that include 456 fire, climate,
terrain, soil, and vegetation attributes. On this foundation, we propose to (1) evaluate
whether pre-trained and foundation models outperform conventional baseline models, (2)
quantify sensitivity to loss functions, input modalities, handling of missing data, and fine-
tuning strategy, (3) test whether SBS models successfully transfer to unseen fires, and (4)
apply explainable and causal inference methods to identify the physical drivers that couple
or decouple VBS and SBS. Preliminary results show that BurnSAVe SBS maps agree with
independent debris-flow hazard estimates and capture contrasting VBS–SBS relationships
under fast and slow fire spread, and demonstrate that a foundation model outperforms a
baseline model in mean intersection over union (IoU) while also improving recall and
moderate-severity accuracy. The expected outcomes are an open dataset, an open modeling
toolkit, and a rapid SBS mapping workflow that can support post-fire emergency response.
Listed by Boise State University (Localist), gathered by the Caldwell events calendar.
Does something look wrong?
Location
City Center Plaza (CCP)777 West Main Street · Boise, ID 83702
Get directions