Improvements to NWP visibility forecasts using statistical post-processing

Accurate visibility forecasting is essential for aviation, road safety, and maritime operations as well as communicating the weather on a daily basis to the public. Despite advancements in Numerical Weather Prediction (NWP) models, it is well understood in the forecasting community that NWP visibility forecasts are inherently poor, often suffering from calibration issues and systematic biases. In post-processing we can enhance skill, however, it is very difficult to add skill when the input data are particularly poor, so this diagnostic remains a known problem. 

This study explores the application of different parametric and non-parametric statistical post-processing techniques to enhance the accuracy and reliability of visibility forecasts. The chosen method will build upon a new visibility scheme at the Met Office, VERA (Visibility Employing Realistic Aerosol), which uses a more physically realistic representation of the condensation nuclei required to form fog and therefore produces a better distribution of visibility for statistical post-processing to work with. 

The calibration methods included in this study include Quantile Regression Random Forests, Reliability Calibration, Bayesian Additive Regression Trees, and finally Distributional Regression Networks using truncated normal and log normal Continuous Ranked Probability Score loss functions, as well as threshold weighted variants of these loss functions. These methods are tailored, where appropriate, to better support the characteristics of visibility data. 

The methodology is tested on an extensive training dataset from the European Centre for Medium-Range Weather Forecasts (ECMWF), which spans 20 years of reforecasts and several European countries capturing a wide range of visibility conditions, including the rarer low visibility events which are most impactful. 

Initial results demonstrate that Quantile Regression Random Forests post-processed forecasts show a marked reduction in Root Mean Square Error compared to raw NWP outputs, and work is in progress to compare this to other methods. These improvements, so far, highlight the great potential of statistical post-processing in refining visibility predictions and supporting decision-making in weather-sensitive sectors.