<![CDATA[Road authorities in developing countries lack cost-effective pavement monitoring tools. This study optimizes smartphone vibration sensing and crowdsourced data for automatic road damage prediction. Accelerometer and GPS data were collected from motorcycles on 42 urban road segments (86.4 km) in Padang, Indonesia, producing 12,480 labelled windows across four pavement condition classes. Four machine-learning classifiers were trained and compared: SVM, ANN, XGBoost, and Random Forest (RF). Sampling rate, mounting position, and contributor aggregation effects were systematically evaluated. RF achieved the highest performance (92.4% accuracy, 91.4% F1-score). A 50 Hz sampling rate with dashboard mounting provided optimal results, while pocket mounting degraded accuracy substantially. Aggregating data from five or more contributors reduced IRI estimation RMSE from 2.31 to 1.21 m/km, with strong correlation to ground-truth measurements (r = 0.91). Optimized smartphone sensing offers a scalable, low-cost alternative for pavement assessment in resource-constrained regions, supporting evidence-based maintenance prioritization.]]>
Copyrights © 2026
