Building damage detection from satellite images after natural disasters on extremely imbalanced datasets.

Assessment of natural disasters caused damage(s) to buildings is important for rescue work coordination which, however, remains a difficult engineering task to be conducted effectively. To automatically detect building damages from satellite imagery, this paper presents a two-step solution approach, including building localization and damage classification. To handle the extremely imbalanced distributions of the building damages, where the minority class occupies less than 0.1%, the architecture is supplemented with a new learning strategy comprising normality-imposed data-subset generation and incremental training. The validity of the proposed architecture is evaluated on a recent open-source dataset named xBD. The experimental study achieves a testing accuracy of 0.9729 and an Intersection over Union (IoU) of 0.5378 on three historical disaster events (i.e., "Mexico-earthquake", "Midwest-flooding", "Palu-tsunami") for the localization analysis, and a testing accuracy of 0.9955 and a weighted F1-score of 0.9953 on the extracted building patches from "Mexico-earthquake", for the followed classification analysis • A two-step solution is provided to automatically detect building damages after natural disasters. • Data normality imposition and incremental learning can handle extremely imbalanced problems. • Additional feature reflecting spatial compactness is created in damage classification. • It achieves a weighted F1-score of 0.9953 on an extremely imbalanced dataset. • It is feasible to be applied in large-scale damage assessment and mapping. [ABSTRACT FROM AUTHOR]