Pavement performance spatio-temporal prediction based on multi-stage deep learning approach

Pavement deterioration tracking faces various challenges, including partial data shortage, short observation cycles and spatio-temporal randomness. This paper presents a multi-stage deep learning framework integrating a random forest neural network (RF-NN), a long short-term memory network (LSTM) and a temporal convolutional network (TCN) for pavement performance prediction. The RF-NN model quantifies the relationships among various pavement performance indicators, while the LSTM model performs time-series prediction of the pavement surface condition index (PCI), riding quality index (RQI) and rutting depth index (RDI) using single- and multi-indicator strategies. Incorporating spatial information, the TCN further improves prediction accuracy and analyses spatio-temporal sensitivity. The results show that the RF-NN model produces high-confidence and well-distributed PCI estimates. The multi-indicator LSTM model outperforms the single-indicator LSTM model. The TCN model yields a lower error than the LSTM model. The mean absolute percentage error of RQI in the TCN prediction is 0.757%.