Spatiotemporal assessment and monitoring of urban heat islands in metropolitan areas using machine learning and downscaling

Urban heat islands (UHI) are a significant phenomenon that results from human modification of the land. Numerical models that are used to study the impacts of UHIs on air temperature require high resolution to estimate spatial fields accurately.This study employs both statistical and dynamic downscaling methods to estimate urban heat islands (UHI) from 1 km-resolution air temperature fields from outputs of the Weather Research and Forecasting (WRF) model, initially generated at a coarser resolution of 5 km, for the Metropolitan Area of Rio de Janeiro (MARJ). In the dynamic method, a grid with 1 km spatial resolution is obtained through a nesting system with three domains of the WRF model of 25–5–1 km. The latter domain couples to the Single-Layer Urban Canopy Model (SLUCM). The statistical approach introduces a novel methodology based on the extreme gradient boosting machine learning algorithm, which correlates air temperature with physiographic landscape variables on a scale of 1 km by 1 multiple nonlinear regression. Additionally, SHAP analysis is applied to assess individual feature contributions in the machine learning model. The performance of theses downscaling methods is evaluated using atmospheric temperature measured at meteorological stations and estimated from remote sensing data. Both methods satisfactorily simulate the temporal and spatial behavior of UHIs in the metropolitan region of Rio de Janeiro, however, the statistical approach demonstrates significantly lower computational costs. This result demonstrates the feasibility of this machine learning-based approach as an alternative for studying and monitoring UHI with limited computational resources.