Adaptive data collection for post-disaster rapid damage assessment: A multi-objective optimization approach

This study proposes a novel adaptive data collection framework for post-disaster rapid damage assessment. The framework aims to maximize information gain by effectively adapting to new observations when identifying and targeting the most informative locations, while accounting for data collection costs, measured by total travel distance, and promoting sample diversity to prevent biased estimations. More specifically, the method consists of two main components: (1) an ordinal probabilistic gradient boosting model that predicts the probabilities of damage severity levels for buildings affected by a disaster, and (2) an adaptive sampling mechanism that iteratively evaluates the prediction model's performance and enhances it by identifying and incorporating the most informative samples. The adaptive sampling process is guided by a multi-objective optimization problem designed to achieve three key goals: (1) reduce prediction uncertainty by prioritizing the selection of samples that are most informative to the prediction model, (2) minimize data collection costs based on total travel distance, and (3) maintain balanced sampling by ensuring that selected observations are representative of the underlying distribution across predefined clusters. The optimization problem is solved using the Non-Dominated Sorting Genetic Algorithm II (NSGA-II). The proposed framework was empirically validated using historical damage assessment data from Hurricanes Michael and Dorian. The results demonstrated the method's ability to continuously improve the predictive power of the model by identifying the most informative target observations while minimizing data collection costs and preserving sample diversity. Specifically, the adaptive strategy reduced the Ranked Probability Score (RPS) to 0.20 with only 10 % of local data, and achieved statistically significant improvements over stratified random sampling. Stable performance was reached after incorporating approximately 35 % of the data, highlighting the framework's efficiency. In addition, the adaptive process achieved shorter total travel distance and lower total variation distance (TVD).