Deep learning-based mapping of urban heat islands

Abstract

Urban heat islands pose a serious problem for urban populations worldwide. In the view of global warming, cities face the challenge of counteracting the increasing overheating of their densely built centres during summer heat waves. In order to be able to take efficient countermeasures, city administrations and urban planners need to know the cooling effect of individual measures. However, empirical data on the effects or possibilities of ex-ante simulations of planned actions are often lacking. To support urban planners with this task, we propose a deep learning-based approach to high resolution mapping and prediction of local UHIs. For this, we employ a dense medium-cost sensor network distributed throughout the city of Dresden, Germany. With the gained temperature sensor measurements, we train a DL model against various data from the environment of the sensors, such as land use and cover, built-up density, building heights, and urban greenery. The trained model is subsequently applied to city-wide available land use data to enable spatially high-resolution mapping and prediction of local UHIs. We test the prediction accuracy of the model against different sensor network layouts in terms of the spatial distribution, the number, the location, and the random failure of individual sensors to provide guidance for optimal sensor network configuration and the transfer to other cities. Finally, we demonstrate the possibilities of simulating the effects of local countermeasures by feeding the trained model with alternative local urban configurations.