A Random Forest Turbulence Prediction Algorithm

Unlike traditional pilot reports, in-situ EDR reports of atmospheric turbulence from commercial aircraft contain both positive and negative instances, are reported regularly, and have relatively accurate positions and timestamps. These data therefore make it feasible to perform more sophisticated analyses of the causes of atmospheric turbulence than were formerly possible. Several real-time gridded products derived from satellite, radar and numerical weather model data that represent storm location and intensity currently exist. These include quantities such as vertically integrated liquid (VIL), echo tops, and wind direction and velocity. In this paper, the authors present a machine-learning algorithm that predicts in-situ peak EDR based primarily on the values of VIL and echo tops in a spatial neighborhood extending approximately 300km around the measurement point. To summarize the values of the gridded products associated with each in-situ EDR measurement, a set of quantities including distances to grid points with data over certain thresholds, maximum data values within each sub-region and the proportion of grid points over various thresholds within each sub-region were computed. A set of the most useful features for turbulence prediction was then determined using a large-scale automated feature selection algorithm. First, an estimate of the "quality” of each candidate feature was calculated by training a large number of decision trees on small random subsets of candidate features and comparing their performance on a testing set both with and without the feature in question. Then, a linear programming problem was formulated in which the "best" subset of features was chosen under the constraint that no two selected features for a given data source could overlap. The selected features and a large training ∗Corresponding author address: Andrew Cotter, Toyota Technological Institute at Chicago, University Press Building, Second Floor, 1427 East 60th Street, Chicago, IL 60637; e-mail: [email protected] set were then used to train a random forest as a predictive algorithm. Finally, the performance of the random forest on an independent testing set was evaluated and compared to another turbulence-prediction product.