Mapping urban temperature using crowd-sensing data and machine learning

Abstract

Understanding the patterns of urban temperature a high spatial and temporal resolution is of large importance for urban heat adaptation and mitigation. Machine learning offers promising tools for high-resolution modeling of urban heat, but it requires large amounts of data. Measurements from official weather stations are too sparse but could be complemented by crowd-sensed measurements from citizen weather stations (CWS). Here we present an approach to model urban temperature using the quantile regression forest algorithm and CWS, open government and remote sensing data. The analysis is based on data from 691 sensors in the city of Zurich (Switzerland) during a heat wave using data from for 25-30th June 2019. We trained the model using hourly data from for 25-29th June (n = 71,837) and evaluate the model using data from June 30th (n = 14,105). Based on the model, spatiotemporal temperature maps of 10 × 10 m resolution were produced. We demonstrate that our approach can accurately map urban heat at high spatial and temporal resolution without additional measurement infrastructure. We furthermore critically discuss and spatially map estimated prediction and extrapolation uncertainty. Our approach is able to inform highly localized urban policy and decision-making.