Generalizable Disaster Damage Assessment via Change Detection with Vision Foundation Model

TL;DR

DAVI tackles the challenge of rapid, fine-grained building-level disaster damage assessment under severe cross-domain shifts by enabling test-time adaptation without target labels. It blends a source-domain change detector with a Vision Foundation Model (SAM) to generate pseudo labels, then refines them through pixel- and image-level processes, while minimizing entropy to focus on confident target predictions. Across diverse regions and disaster types, DAVI outperforms unsupervised change detection and domain adaptation baselines, with a real-world Türkiye earthquake case illustrating practical impact and robustness. The approach reduces reliance on ground-truth labels, enabling timely, scalable disaster response and resource allocation in unseen environments.

Abstract

The increasing frequency and intensity of natural disasters call for rapid and accurate damage assessment. In response, disaster benchmark datasets from high-resolution satellite imagery have been constructed to develop methods for detecting damaged areas. However, these methods face significant challenges when applied to previously unseen regions due to the limited geographical and disaster-type diversity in the existing datasets. We introduce DAVI (Disaster Assessment with VIsion foundation model), a novel approach that addresses domain disparities and detects structural damage at the building level without requiring ground-truth labels for target regions. DAVI combines task-specific knowledge from a model trained on source regions with task-agnostic knowledge from an image segmentation model to generate pseudo labels indicating potential damage in target regions. It then utilizes a two-stage refinement process, which operate at both pixel and image levels, to accurately identify changes in disaster-affected areas. Our evaluation, including a case study on the 2023 Türkiye earthquake, demonstrates that our model achieves exceptional performance across diverse terrains (e.g., North America, Asia, and the Middle East) and disaster types (e.g., wildfires, hurricanes, and tsunamis). This confirms its robustness in disaster assessment without dependence on ground-truth labels and highlights its practical applicability.

Figures (5)

• Figure 1: Collaborative efforts are underway to collect datasets on natural disasters, yet no single benchmark dataset covers the entire world. The yellow areas represent regions with recent disasters not included in the dataset.
• Figure 2: Model architecture. This figure illustrates how pre- and post-disaster images and their augmented versions are processed. The system has two phases: (Step 1) pseudo label generation, which leverages the source model and image segmentation foundation model (SAM) with a text prompt for generalizability across various disaster scenarios, and (Step 2) pseudo label refinement, which involves pixel- and image-level refinement to reduce noise in the fine-grained pseudo label. The final output is a fine-grained pseudo label indicating potential damage.
• Figure 3: Step-by-step visualization of DAVI. From left to right, images represent pre- and post-disasters, corresponding confidence maps, confidence difference maps discretized by the optimal threshold, binary change maps from the source model, fine-grained pseudo labels without refinement, our pseudo labels, and ground-truth labels. Structures accurately identified through image segmentation appear with a cyan box, while those recognized through refinement appear with a yellow box.
• Figure 4: Hyperparameter analysis of $\tau_{v}$. Solid line shows F1-score of DAVI, while dashed line shows F1-score between the source model's and SAM's binary change maps across thresholds.
• Figure 5: Evaluation on the 2023 Türkiye earthquake. Columns show pre- and post-disaster images, predictions from baselines and our method. The top row shows damaged cases, while the bottom row shows undamaged cases.