A Class Distance Penalty Deep Learning Method for Post-disaster Building Damage Assessment
Journal
KSCE Journal of Civil Engineering
Date Issued
2024-01-01
Author(s)
Tsai, Fang Jung
Abstract
Automatic building damage assessment can significantly aid rescue operations, attributed to booming deep learning and remote sensing technologies. However, the class imbalance of the dataset often skews prediction models towards the majority class in the segmentation of damaged buildings. This issue is further exacerbated when damaged buildings are categorized into multiple scales, intensifying biases within the models. Hence, this research adopts an algorithm-level method to improve the reliability of post-disaster damage assessment. It proposes a novel loss function named Ordinal Class Distance Penalty Loss (OCDPL), considering the ordinal relationship between classes and penalizing the misclassifications according to the class error distance. Two hyperparameters are also introduced to enable the model to fine-tune the contribution of ordinal relationships on the loss function. The satellite images of hurricane disasters in the xBD dataset were adopted as the case study. The results show that the proposed approach can improve F1 scores and Mean Absolute Error of overall damage level classes. Notably, the findings underscore the value of leveraging information on ordinal classes to facilitate the learning of minority classes and diminish class error distances. This aspect holds particular significance for emergency responses to widespread and severe disasters.
Subjects
Building damage assessment | Class imbalance | Image segmentation | Ordinal multi-class segmentation | Penalty loss function | Remote sensing imagery
Type
journal article