Pavement condition forecasting from incomplete and irregular inspection data using sliding mask augmentation

Pavement Management Systems (PMS) require accurate deterioration forecasting to support maintenance and rehabilitation decisions, yet traditional predictive models often struggle with the irregular, noisy, and incomplete nature of real-world inspection data. While machine learning (ML) offers superior predictive capabilities over standard empirical regression, managing missing data without introducing bias through denoising or data leakage remains a significant challenge. This study introduces an ML framework of a sliding-window mask technique that utilizes all available historical condition records and explicitly incorporates irregular missing patterns into the learning process without requiring data exclusion or denoising procedures. Consequently, the proposed ML framework delivers consistent pavement condition forecasts regardless of temporal data fragmentations and data limitations at later ages of pavement. The predictive capabilities of nine ML algorithms were extensively evaluated across two datasets: a synthetically generated dataset simulating 30 years of Pavement Condition Index (PCI) observations for 500 segments with noise and missingness; and a real-world dataset containing Pavement Distress Score (PDS) observations for 347 segments with irregular temporal spacing. Results demonstrate that advanced non-linear algorithms consistently outperform traditional linear approaches across both datasets. For near-term predictions, CatBoost, LightGBM, and Artificial Neural Networks (ANN) achieved the highest accuracy on two different datasets. Conversely, linear models exhibited a pronounced predictive bias, restricted prediction ranges, and systematic underfitting. For extended forecasting horizons, prediction uncertainty naturally increased across all models; however, the advanced machine learning algorithms maintained the most reliable performance by yielding unbiased mean error profiles. Methodologically, sliding mask augmentation proved highly effective in enabling models to learn deterioration patterns from irregularly sampled data while strictly preventing data leakage between the training and testing phases. The findings suggest that advanced ensemble methods and ANNs can capture complex, nonlinear pavement deterioration directly from incomplete datasets, which provide transportation agencies with a data-driven approach for comprehensive infrastructure management.