From Sparse Stations to Road Networks: Hourly Visibility Prediction Using a Machine Learning-Based Stacked Ensemble Model

Research on surface visibility and road-network-level visibility is crucial for transportation. However, complex terrain, frequent dust activities, and sparse observations increase the difficulty of continuous visibility monitoring along roads. This study develops a heterogeneous evidence fusion framework, termed VisCET, which links site-level supervision, seamless grid inference, and road-network-level representation. The framework integrates station observations, atmospheric composition, remote sensing, land-surface information, and other auxiliary data, and employs a two-layer stacked ensemble model composed of multiple machine learning algorithms and a deep learning module to generate seamless hourly surface visibility (SV) fields at a spatial resolution of 0.1° × 0.1° and long-distance road-network visibility (RV) fields. Sample-based 10-fold cross-validation of the VisCET model using approximately 94,131 data pairs demonstrated good robustness and accuracy, with a Pearson correlation coefficient (R) of 0.94 and a root mean square error (RMSE) of 3.31 km. Temporal and site-based cross-validation further confirmed the strong performance of the model in the temporal and spatial domains, with R values of 0.916 and 0.886 and RMSE values of 3.950 and 4.315 km, respectively. The seamless hourly SV fields capture the formation, development, and dissipation of representative low-visibility events. The generated hourly, continuous, gridded visibility fields and road-unit products extend atmospheric environmental monitoring from regular grids to road corridors, providing spatial environmental information for road weather services, route planning, and the identification of low-visibility road segments.