Approximation of Spatial and Temporal Variability of the Urban Heat Island in Moscow Using Machine Learning

This study is devoted to the application of machine learning (ML) methods for statistical approximation of the urban-induced temperature anomaly, known as the urban heat island (UHI), and its spatiotemporal dynamics, using the example of the Moscow megacity. This task is considered as part of a more general problem of statistical downscaling of meteorological fields for urban conditions. Therefore, we aim to approximate a high-resolution field of urban temperature anomalies based on predictors characterizing low-resolution meteorological data and high-resolution surface properties. As the input data for training ML models, we use the results of high-resolution hydrodynamic simulations of the meteorological regime in the Moscow region conducted with the COSMO regional atmospheric model coupled with the TERRA_URB urban canopy parameterization. For the ML model, we use the gradient boosting method implemented by the CatBoost algorithm with GPU support. To account for nonlocal dependences between UHI and surface properties, we use an original quasi-local approach to define the feature vectors. This approach consists of using data localized at individual points (nodes of the computational mesh of the COSMO model) as feature descriptions and generating additional features based on the predictors’ values for neighboring points using different types of convolution filters. As such filters, we use a moving average with a circular kernel of different radii and more advanced self-adjusting directional filters formed by taking into account large-scale data on wind speed and direction. We show that such nonlocal features are important for correctly reproducing the key patterns of the UHI spatial structure, in particular the smoother structure of seasonally-averaged temperature anomalies in comparison to surface properties, and the shift of temperature anomalies to the leeward side of the city for specific cases with different wind directions.