Systematic Evaluation of U-Net Variants for Multiclass Concrete Damage Segmentation: Addressing Class Imbalance with Resource Constraints

Abstract

Postdisaster inspection of concrete structures demands rapid, reliable assessment tools, yet conventional manual surveys are slow, subjective, and impractical under emergency conditions. Although deep learning has advanced crack detection, most prior work is limited to binary segmentation and neglects other critical surface damages such as spalling. In addition, postdisaster environments present two major challenges: (1) severe class imbalance at the pixel level; and (2) constrained computing resources that hinder deployment of computationally intensive models. The objective of this study is to systematically evaluate U-Net variants for multiclass concrete damage segmentation under class imbalance and computing-resource constraints. This study systematically evaluates four U-Net variants, the original U-Net, dilated U-Net, attention U-Net, and spatial and channel squeeze-and-excitation (SCSE) U-Net, alongside a well-established external benchmark, DeepLabv3 with a ResNet-101 backbone. All models were trained, validated, and tested on an annotated dataset of cracks and spalling and assessed using both quantitative metrics and qualitative inspections of segmentation outputs. Computational efficiency, including training time, inference time, and memory usage, was also monitored to gauge practical feasibility for on-site deployment. Results show that all U-Net variants achieved strong segmentation performance with only marginal differences between them. Attention U-Net delivered the most balanced class-specific performance, while the original U-Net combined balanced accuracy with the lightest computational footprint. Compared to the benchmark, U-Net variants achieved 5.9% faster training, 49.8% faster inference, and overall reduced memory usage, underscoring their suitability for practical deployment in postdisaster environments. This study provides a systematic evaluation of U-Net variants for multiclass concrete damage segmentation. It offers empirical insights into the persistent challenge of class imbalance and emphasizes computational efficiency as a core evaluation dimension. These findings lay the groundwork for future frameworks that link automated damage maps with structural models to accelerate disaster reconnaissance.