Computer vision tools for early post-disaster assessment: Enhancing generalizability

Remote sensing data, particularly satellite imagery, have made early, post-hazard aerial damage assessment possible due to its fast availability and extensive coverage. Despite breakthroughs in using deep computer vision with satellite image inputs, achieving high generalizability across diverse hazards and locations remains the main obstacle to effectively deploying early assessment tools in real-world scenarios, as the same hazard can manifest differently across various landscapes and urban textures. This challenge may be overlooked when working with curated datasets with minimal (test-to-train) shifts in urban textures and hazard damage features (e. g. , due to undersized study regions), leaving models ill-prepared for real-world scenarios. The primary objective of this study was to understand non-trivial generalizability challenges by taking the 2023 Turkiye earthquake and the Maui wildfire incidents as "training" and "unseen test" events that simulated a demanding scenario. Subsequently, strategies such as augmenting image channels with damage proxy maps, data fusion, deep ensemble learning, and Test Time Augmentation were exclusively designed and implemented to address those challenges. These measures significantly improved the damage detection, with or without severity classification, with F1 scores increased from 0. 71 to 0. 82 and 0. 40 to 0. 87, respectively. Furthermore, and through data fusion, the proposed framework accommodates estimating socioeconomic loss metrics at the individual building level, supporting both response and recovery phases. This research has the potential to enhance the effectiveness of rapid aerial damage assessment models, ultimately aiding in more efficient and targeted disaster response and recovery efforts. Data, models, and codes are available at https: //github. com/TRG-AI4Good/LahainaGeneralizability.