Semantic Segmentation-Based Identification of Seismic-Induced Building Damage by Using UAV Orthophoto

Abstract

: Due to transportation disruptions after the earthquake, on-site surveys could not quickly assess house damage. Remote sensing can quickly acquire disaster-related information over large-scale affected areas and support rapid localization and morphological identification of damaged buildings from top-view observations. Nevertheless, on-site inspection by experienced engineers is still the most accurate method for judging real building structural damage, since UAV orthophotos only record surface features and cannot capture internal building damage. Field investigation data in our study served as ground-truth references for UAV image interpretation and labelling. Current remote sensing-based post-earthquake building assessment mostly adopts a binary classification system of collapsed and non-collapsed, which makes it difficult to distinguish between lightly damaged and intact buildings, and the imbalance between positive and negative sample sizes leads to low identification accuracy. This article created the orthophoto images from unmanned aerial vehicles in Shuanghe Town, Changning County, Sichuan Province and established a classification standard for building seismic damage based on visual interpretation. The earthquake-damaged buildings were ultimately classified into three damage levels: completely collapsed, partially damaged, and basically intact. Semantic segmentation can identify different objects in an image. This article compares the performance of different models, sample sizes, and loss functions. This study compares 50% overlapping cropping with non-overlapping cropping, and compares a Focal + Dice hybrid loss with cross-entropy, on UAV orthophoto patches of Shuanghe Town. Reported scores are mean metrics on the test block of this single mosaic. The results show that the MobileNetV3 model with Focal Loss and Dice Loss reported the highest mean mIoU, mPA, and F1 scores of 92.36%, 97.32%, and 96.77%, respectively, on the evaluated test block with 512 × 512 pixels and 50% overlapping sampling.