Deep Learning for Disaster Response

QC 20261002

Abstract

With climate change increasing the frequency and severity of extreme weather events, it becomes increasingly important to quickly assess the damages and respond appropriately. Remote sensing methods like satellite and drone observations allow us to survey these large-scale events with observations at matching scales. This thesis addresses mismatches between such applications in disaster response and standard computer vision practice: which data is used or left unused, how generalization is evaluated, and how predictive uncertainty is modeled. For building damage detection, we modify the commonly-used dataset split on the xBD dataset to measure generalization to completely unseen disaster events, instead of just unseen images within known events. We further condition the model predictions on the type of natural hazard involved, information that is always readily available in the application context and helps the model to generalize to unseen events. For wildfire spread prediction, we first introduce a new dataset which provides access to a fire's recent history instead of only the current day's observation. Finally, motivated by the limited performance of wildfire spread prediction throughout the literature, we rethink how to model the prediction targets. Instead of predicting only one segmentation mask or independent pixel-wise probabilities, we produce a set of distinct outputs that aim to capture the underlying uncertainty about possible outcomes.

Link to DiVA