Classification of Road Damage in Sidoarjo Using CNN Based on Inception Resnet-V2 Architecture

Abstract

Road damage is a serious issue in Sidoarjo Regency, posing risks to road users' safety. This study aims to classify road surface conditions using a Convolutional Neural Network (CNN) model based on the Inception ResNet-V2 architecture. The research develops an image-based classification model by combining secondary data from Kaggle and primary data obtained through Google Street View API scraping, along with training strategies such as data augmentation, class balancing, early stopping, and model checkpointing. A total of 885 images were used, categorized into three classes: potholes, cracks, and undamaged roads. The model was trained over 20 epochs with early stopping triggered at epoch 15, when validation accuracy reached 95.95%. Evaluation on the test set showed a test accuracy of 83%. The undamaged road class achieved the highest performance with an F1-score of 0.89, while the pothole class recorded an F1-score of 0.79. The lowest performance was observed in the cracked road class, with an F1-score of 0.65, indicating the model's limited ability to detect fine crack features. This limitation is likely due to class imbalance and visual similarity between classes. Although the model demonstrated good generalization for the two majority classes, the performance gap between validation and test accuracy highlights the need to improve detection for minority classes. Future work is recommended to explore advanced augmentation techniques, increase the representation of minority class data, and consider alternative architectures or ensemble methods to enhance the model’s sensitivity to subtle road damage features.