Back to search

Georgia Institute of Technology

Post-Disaster Damage Assessment Using Deep Learning with Surrogate Data

Abstract

dc:description.abstract

Natural hazards such as earthquakes, hurricanes, floods, wildfires, and landslides continue to impose substantial human and economic losses worldwide. Effective response and recovery depend on rapid, reliable damage assessment in the first hours to days after an event, when decision-makers must identify the hardest-hit areas, prioritize inspections, and allocate scarce resources. In practice, this early period is constrained by incomplete situational awareness, limited access to affected zones, environmental constraints, and the lack of event-specific labeled data and timely expert-verified assessments. This gap between urgent information needs and delayed labeled ground truth constitutes a persistent cold-start problem in post-disaster damage assessment. Surrogate data, such as social media information, provide a practical means of addressing the cold-start problem. In this work, surrogate data refers to timely, opportunistic, or non-traditional observations that provide proxy evidence of damage, disruption, or recovery when authoritative assessments are delayed. While conventional sensing sources, such as seismic products, SAR-based deformation indicators, optical satellite imagery, and drone surveys, remain foundational, their use can be constrained by acquisition latency, interpretation complexity, and environmental conditions. Surrogate data complement these sources by providing earlier or continuous signals that can support operational decisions under real-world constraints. Two surrogate data streams are examined in depth. First, nighttime lights (NTL) are studied as a regional proxy for disruption and recovery, using NASA Black Marble radiance time series to extract decision-relevant signals amid substantial temporal variability. Second, ground-view imagery from affected areas, primarily social media images and expert reconnaissance collections, provides fine-grained visual cues at the asset level but also introduces substantial heterogeneity in capture conditions. This dissertation contributes: (i) empirical analyses that clarify when and how NTL change ratios correlate with disaster impacts; (ii) the construction of expert-aligned benchmark datasets for social media-based earthquake damage classification and semantic segmentation from ground-view imagery; and (iii) scalable annotation protocols that improve consistency and reproducibility while reducing dependence on scarce expert effort. Together, these two data streams address both macro-scale and micro-scale assessment needs and motivate the development of computational methods that are robust to domain shift and limited supervision. Building on these surrogate data streams, this dissertation further develops computational methods that convert such observations into actionable damage information under the same cold-start constraints, including scarce event-specific labels, heterogeneous and often low-quality observations, time-intensive annotation, and limited computational resources during early response. Specifically, it introduces: (i) ScaleNet, a multi-scale representation learning approach that improves feature learning when supervision is limited; (ii) DASeg, a robust and domain-adaptable segmentation pipeline that leverages vision foundation models to enable weakly supervised infrastructure damage localization from noisy ground-view imagery; and (iii) SeismoMind, a multimodal large language model framework for zero-shot damage assessment that achieves performance comparable to state-of-the-art supervised baselines without task-specific training. Across these methods,the emphasis is not only on detecting damage, but also on maintaining reliable performance across events and capture conditions typical of early response, where domain shift is unavoidable and outputs must remain stable enough to support downstream prioritization and resource allocation. In the final chapter, the dissertation outlines future directions for extending these results toward operational deployment, with particular attention to the datasets to be collected, the features to be labeled, and the models to be developed. In summary, the dissertation consolidates these contributions through a set of published and under-review studies spanning surrogate-data analytics, dataset construction, representation learning, weakly supervised and foundation-model-based segmentation, and multimodal large language model reasoning for damage understanding. Collectively, the findings support the overarching conclusion that surrogate data, combined with robust deep learning frameworks, can substantially mitigate cold-start constraints and enable more timely, scalable, and actionable post-disaster damage assessment.

Degree

thesis:*
Name thesis:degree_name
Computational Science and Engineering, PhD
Department thesis:degree_department
Computational Science and Engineering
Grantor
Georgia Institute of Technology
Year dc:date.issued
2026

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Huang, Huili
Advisor dc:contributor.advisor
  • Frost, J. David
Committee members dc:contributor.committeemember
  • Roozbahani, Max
  • Chau, Duen Horng (Polo)
  • Roy, Nimisha
  • Prakash, B. Aditya
ORCID person:identifier.orcid

Rights

Language dc:language.iso
English

Identifiers

dc:identifier.*
Handle dc:identifier.uri
https://hdl.handle.net/1853/82184

Chain of custody

source
Harvested from
Georgia Tech
Base URL
repository.gatech.edu/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
related terms
citation

Huang, Huili. Post-Disaster Damage Assessment Using Deep Learning with Surrogate Data. Georgia Institute of Technology, 2026. https://hdl.handle.net/1853/82184