Predicting urban cold-air paths using boosted regression trees

Abstract

Ensuring the transport of colder air masses from rural surroundings or inner urban open spaces into the dense urban center is a vital approach for reducing nocturnal warming of urban areas in summer, i.e. the urban heat island (UHI) effect. Hence, cold-air paths are an important mitigation aspect in urban planning. The identification of cold-air paths by field measurements, however, is time-consuming and it is difficult to cover larger spatial domains (e.g. entire cities). Therefore, we used a machine learning technique, boosted regression trees (BRT), for estimating the spatial distribution of cold-air paths in three German cities (Braunschweig, Freiburg, Stuttgart). We identified the most important predictor variables for cold-air path occurrence and tested the spatial transferability of BRT models from one training city to the other study cities.Three variables explaining differences in surface elevation indicated strongest influence on spatial distribution of cold-air paths: relative surface elevation, topographic position index, and topographic wind index. We achieved predictions for the spatial distribution of cold-air paths for cities the model was trained on, resulting in model performances of 0.85 < AUC < 0.96 (AUC: area under the receiver operating characteristic curve). Model transfers to other cities, however, achieved only poor to moderate results (0.5 < AUC < 0.8). The city of Freiburg tended to be the most suitable basis for a general model, since it covers the largest range of conditions. Transfer works best for ‘similar’ cities, i.e. cities with a comparable range of predictor variables and morphological structure. Cold-air path identification via BRT modelling appeared to be a promising method delivering helpful information for urban planning.