Visibility Prediction Based on Kilometric NWP Model Outputs Using Machine-Learning Regression

Abstract

Low visibility conditions have a strong impact on air and road traffics and their prediction is still a challenge for meteorologists, particularly its spatial coverage. In this study, an estimated visibility product over the north of Morocco, from the operational NWP model AROME outputs using the state-of-the art of Machine-learning regression, has been developed. The performance of the developed model has been assessed, over the continental part only, based on real data collected at 37 synoptic stations over 2 years. Results analysis points out that the developed model for estimating visibility has shown a strong ability to differentiate between visibilities occurring during daytime and nighttime. However, the KDD-developed model have shown low performance of generality across time. The performance evaluation indicates a bias of -9m, a mean absolute error of 1349m with 0.87 correlation and a root mean-square error of 2150m.