Reliable road condition monitoring is fundamental to maintenance decision-making and transportation safety, particularly in regional contexts where data resources are often scarce. This study presents a comparative evaluation of convolutional neural network (CNN) architectures for classifying road damage types using digital images collected in Sleman Regency. Three widely used CNN architectures, VGGNet-16, InceptionV3, and Xception, were evaluated under a unified experimental framework employing transfer learning, consistent preprocessing, explicit hyperparameter tuning, and four-fold cross-validation. The dataset comprises three road damage categories, alligator crack, corrugation, and pothole, captured under heterogeneous pavement and lighting conditions. Image preprocessing includes resizing, augmentation, and contrast enhancement using Contrast Limited Adaptive Histogram Equalization (CLAHE). To assess the contribution of preprocessing choices, an ablation study was conducted by comparing model performance with and without CLAHE. Experimental results indicate that all evaluated architectures achieve high classification performance. Among them, Xception consistently demonstrates the best overall performance across validation and test sets, achieving the highest accuracy, precision, recall, and F1-score. The ablation analysis further shows that including CLAHE consistently improves performance, particularly in recall and F1-score, indicating enhanced robustness under uneven illumination conditions. Although the contribution of this study is incremental rather than algorithmically novel, the findings provide empirical insights into the comparative behavior of CNN architectures under region-specific road conditions. The results highlight the importance of systematic preprocessing and controlled evaluation in CNN-based road-damage classification and provide practical guidance for selecting suitable architectures for regional road maintenance decision-support systems.
Sari et al. (Wed,) studied this question.