Widyan Hirzi Wibowo, Fitri Utaminingrum, Barlian Henryranu Prasetio
Smart city infrastructure management demands an automated, fast, and accurate road condition monitoring system. These systems often rely on edge devices such as Unmanned Aerial Vehicles that have limited computational resources. This study aims to find the most optimal deep learning architecture for road damage classification, focusing on the balance between accuracy and computational efficiency. This study conducted a comparative evaluation of four models: MobileNetV3 Small, MobileNetV3 Large, MobileNetV3 Small + CBAM 16, and MobileNetV3 Small + CBAM 32. These models were trained and tested using the UAV-PDD2023 dataset, which contains 11,158 preprocessed road damage images. The evaluation was conducted based on performance metrics and efficiency metrics. The results show that MobileNetV3 Small CBAM 16 consistently achieves the highest classification performance, with an accuracy of 94.17% and a weighted F1-score of 94.17%. This performance outperforms the much heavier baselines MobileNetV3 Small with 93.58% accuracy and MobileNetV3 Large with 93.36% accuracy. Furthermore, the MobileNetV3 Small CBAM 16 model remains highly efficient with 4.919 MB in size, 63.82 MMac FLOPs, making it an ideal choice for real-time implementation on resource-constrained Unmanned Aerial Vehicles devices. © 2026 SPIE.
Faculty of Computer Science, Universitas Brawijaya, Jl. Veteran No.8, East Java, Malang, 65145, Indonesia
