Machine-learning regression applied to diagnose horizontal visibility from mesoscale NWP model forecasts

Low-visibility conditions are a weather hazard that affects all forms of transport, and accurate forecasting of their spatial coverage is still a challenge for meteorologists, particularly over a large domain. Current predictions of visibility are based on physical parametrizations in mesoscale models and are thus limited with respect to accuracy. This paper examines the use of supervised machine-learning regression techniques (tree-based ensemble, feed-forward neural network and generalized linear methods) to diagnose visibility from operational mesoscale model forecasts over a large domain. To achieve this, hourly forecasts of meteorological parameters in the lower levels of the atmosphere have been used. In the short-range forecasting framework, the machine-learning algorithms were developed to provide hourly forecasts up to 24 h. To assess the performance of the developed models, hourly observed data, collected at 36 synoptic land stations over the northern part of Morocco, have been used. This region is characterized by a heterogeneous topography. The tree-based ensemble methods have shown some improvement in visibility forecasting in comparison with the operational visibility diagnostic scheme based on Kunkel's formula and also with persistence. It is also found that this machine-learning technique performs better when the forecast depends on multiple predictors instead of only a few with very high importance. In addition, their performance is very sensitive to the disproportionality of data availability between daytime and night-time. Furthermore, it is found that the performance decreases when principal components are used instead of raw correlated data.