A deep transfer learning framework for segment-specific pavement condition forecasting using family-level priors and sparse section histories

This study proposes a two-stage transfer-learning framework for forecasting the Pavement Condition Index (PCI) of individual pavement sections with limited inspection histories. Transportation agencies commonly rely on family-level performance models because section histories are too sparse to support reliable individual models, although maintenance and rehabilitation decisions are made at the section level. The proposed framework bridges this gap by combining a pretrained deterioration prior with each section’s observed PCI history. Attention-based sequence-to-sequence models with Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) backbones are first pretrained on simulated PCI trajectories. The resulting prior is then transferred to field sections through either zero-shot inference or local adaptation. The framework uses pavement age and PCI as sequential inputs, handles missing future targets through masking, and applies PCI-range clipping, optional monotonicity enforcement, and adjustment diagnostics. Simulated-data results show that adaptation performance depends on the method and recurrent backbone. On a U.S. State DOT dataset, zero-shot transfer provides the most stable deep-learning configuration and remains competitive with Random Forest (RF) and XGBoost (XGB). Performance is strongest at the 2-year horizon and decreases substantially at the 6-year horizon. Overall, simulated pretraining provides a practical basis for section-level PCI forecasting when field histories are insufficient for training deep sequence models from scratch.