Minor Damage Recognition from Postearthquake Buildings with an Improved Generative Adversarial Semantic Segmentation Network

Abstract

For postearthquake reconstruction and insurance claims, accurately extracting comprehensive postdisaster building damage is essential. Traditional satellite remote sensing techniques are insufficient for detecting small and minor damage targets. Instead, airborne and terrestrial images have emerged as promising alternatives for detailed damage detection. This study describes the creation of the Minor Damage Segmentation for Post-Earthquake Buildings data set, which includes roof holes, deformation, debris, facade cracks, and spalling, based on detailed field investigations and data collection. Given the limited sample size, a deep generative adversarial network (GAN) is employed for semantic segmentation to achieve detailed segmentation of damaged targets. The generative network of the GAN incorporates a transformer and coordinate attention mechanism to enhance the extraction of local damage features, while two PatchGANs in the discriminative network improve the model’s adaptability to various damage types common in roofs and facades. Additionally, improved penalty functions are incorporated to ensure stability within the adversarial network. The proposed method outperforms conventional approaches, achieving an intersection-over-union accuracy exceeding 90% and an overall accuracy greater than 80%.