Processing AROME Vertical Profiles with Machine Learning Methods to Diagnose Aeronautical Ceiling in TAF messages

Accurate cloud-ceiling-height forecasts derived from numerical weather prediction model data has a great importance for aircraft operations. Recent improvements in microphysical schemes at high resolution provide useful information. At the same time, improvements in machine learning methods and calculation capacities allow to combine large sets of meteorological predictors issued from NWP models in order to estimate ceiling height. This study shows how mutiple machine learning methods can be useful to diagnose ceiling height by using vertical profiles of several meteorological parameters as predictors.

This work is part of Meteo-France’s IniTAF (TAF Initialization) innovative project. IniTAF aims at producing draft TAFs, thus optimizing the forecasters’ workload while maintaining high safety levels in take-off, taxi, and landing operations. In Terminal Aerodrome Forecasts (TAF), ceiling corresponds to the lowest reported occurrence of BKN (broken, sky cover > 50%) or OVC (overcast, sky cover = 100%) cloud layers, or to the vertical visibility when the sky is obscured. Following Annex 3 of the Convention on International Civil Aviation, forecasters have to mention when the base height of the lowest layer or mass of cloud of BKN or OVC extent is forecast to lift and change to or pass through one or more of the following values : 30, 60, 150, 300 m (100, 200, 500, 1 000 ft). Thus, we explore the ability of learning methods to construct a classification matching those thresholds and the optional threshold of 450 m (1500 ft).

A training and validation datasets are composed by hourly ground-based METAR Meteorological Aeronautical Reports (METAR) and data from Meteo-France’s network independently selected among inputs from January 2016 to mid-2017 (66 airports, i.e. 10⁵ records). Cloud layers and ceiling are computed from ceilometers (CL31) using the Automated Surface Observing System (ASOS) algorithm. Every METAR is matched to outputs of the French regional model AROME, extracted on a 20 km x 20 km horizontal sub-grid and centered around each airport. For each point in the horizontal grid, meteorological parameter profiles are extracted for 24 levels from 10 m to 3000 m. Machine learning techniques (including Random Forest with bagging or boosting, and deep neural networks) are implemented. The generated models are evaluated in terms of accuracy an independent set of METAR observations and ability to handle a large amount of predictors. Relevant scores are computed for each method and compared with ceiling forecasts from persitence as well as a benchmark ceiling diagnosis based on three-dimensional AROME nebulosity outputs. This validation is completed through the analysis of several case studies.