Summer Nights in Berlin, Germany: Modeling Air Temperature Spatially with Remote Sensing, Crowdsourced Weather Data, and Machine Learning

Abstract

Urban areas tend to be warmer than their rural surroundings, well-known as the 'urban heat island' effect. Higher nocturnal air temperature (Tair) is associated with adverse effects on human health, higher mortality rates, and higher energy consumption. Prediction of the spatial distribution of Tair is a step toward the 'Smart City' concept, providing an early warning system for vulnerable populations. The study of the spatial distribution of urban Tair was thus far limited by the low spatial resolution of traditional data sources. Volunteered geographic information provides alternative data with higher spatial density, with citizen weather stations monitoring Tair continuously in hundreds or thousands of locations within a single city. In this article, the aim was to predict the spatial distribution of nocturnal Tair in Berlin, Germany, one day in advance at a 30-m resolution using open-source remote sensing and geodata from Landsat and Urban Atlas, crowdsourced Tair data, and machine learning (ML) methods. Results were tested with a 'leave-one-date-out' training scheme (testingcrowd) and reference Tair data (testingref). Three ML algorithms were compared - Random Forest (RF), Stochastic Gradient Boosting, and Model Averaged Neural Network. The optimal model based on accuracy and computational speed is RF, with an average root mean square error (RMSE) for testingcrowd of 1.16 °C (R2 = 0.512) and RMSE for testingref of 1.97 °C (R2 = 0.581). Overall, the most important geographic information system (GIS) predictors were morphometric parameters and albedo. The proposed method relies on open-source datasets and can, therefore, be adapted to many cities worldwide.