Improvement in the Forecasting of Low Visibility over Guizhou, China, Based on a Multi-Variable Deep Learning Model

High-quality visibility forecasting benefits traffic transportation safety, public services, and tourism. For a more accurate forecast of the visibility in the Guizhou region of China, we constructed several visibility forecasting models via progressive refinements in different compositions of input observational variables and the adoption of the Unet architecture to perform hourly visibility forecasts with lead times ranging from 0 to 72 h over Guizhou, China. Three Unet-based visibility forecasting models were constructed according to different inputs of meteorological variables. The model training via multiple observational variables and visibility forecasts of a high-spatiotemporal-resolution numerical weather prediction model (China Meteorological Administration, Guangdong, CMA-GD) produced a higher threat score (TS), which led to substantial improvements for different thresholds of visibility compared to CMA-GD. However, the Unet-based models had a larger bias score (BS) than the CMA-GD model. By introducing the U²net architecture, there was a further improvement in the TS of the model by approximately a factor of two compared to the Unet model, along with a significant reduction in the BS, which enhanced the stability of the model forecast. In particular, the U²net-based model performed the best in terms of the TS below the visibility threshold of 200 m, with a more than eightfold increase over the CMA-GD model. Furthermore, the U²net-based model had some improvements in the TS, BS, and RMSE (root-mean-square error) compared to the LSTM_Attention model. The spatial distribution of the TS showed that the U²net-based model performed better at the model grid scale of 3 km than at the scale of individual weather stations. In summary, the visibility forecasting model based on the U²net algorithm, multiple observational variables, and visibility data from the CMA-GD model performed the best. The compositions of input observational variables were the key factor in improving the deep learning model’s forecasting capability, and these improvements could improve the value of forecasts and support the socioeconomic needs of sectors reliant on visibility forecasting.