Automatic Classifier of Road Condition and Early Warning System for Potholes

Damaged roads can have a negative impact on road users and can fatally cause accidents. One sign of a damaged road is the presence of holes in the road. This research aims to develop an Android application that can display the location of potholes and provide early warning to driver in Simalungun Regency - North Sumatra. This research implements the Convolutional Neural Network (CNN) algorithm using the transfer learning techniques on the pre-trained MobileNetV3 model for automatic classification of road conditions. The dataset used in the research consisted of 22.538 images which were divided into two classes, namely pothole and normal. This research uses dataset with a ratio of 60:20:20, 70:20:10 and 80:10:10. MobileNetV3 large variant with a dataset ratio of 60:20:20 shows the best value with an F1-Score of 0,9035. The model was further converted to Tensorflow Lite with an F1-Score of 0.8985. This research succeeded in implementing the trained and evaluated model along with early warning of potholes via audiovisual in Android application. Application functionality testing that is carried out using black box testing, showing that the application can run well.