Remote sensing-based road damage recognition using EfficientNet with attention mechanism

Potholes pose a major challenge to road safety in South Africa. In this study, we developed a deep learning-based pothole detection system using a dataset of 4,026 labeled street images. After preprocessing and data augmentation, we trained an EfficientNet-B0 model with Adam optimization and early stopping. The model achieved 99.67% accuracy on the training set and maintained high generalization on the validation (98.51%) and test sets (98.39%, ROC-AUC ⪆ 0.998). Visualizations with bounding boxes, Grad-CAM heatmaps, and contour maps further confirmed model interpretability. These results demonstrate the robustness and practical value of EfficientNet-B0 for automated road infrastructure monitoring.