Towards ML-Based Weather-Dependent Diagnostic Index for Urban Heat Island Intensity in the Arctic Cities

The urban heat island (UHI) effect in Arctic cities remains poorly understood, and its dependence on weather conditions has never been systematically analyzed for this region. This study examines the extent to which large-scale meteorology can explain UHI variability and whether a universal, weather-driven index representing such dependencies can be developed. Using data from the UHIARC observational network and ERA5 reanalysis for seven Russian Arctic cities, we evaluated several classical semi-empirical parameterizations (Wind Factor, Weather Factor, Cloud Factor, Diurnal Factor) alongside a suite of machine learning models, namely linear regression, MLP, LSTM, and CatBoost (version 1.2.10). The empirical baselines showed near-zero predictive skill, whereas CatBoost with explicit temporal features representing delayed connections captured up to 64% of the local UHI variance. A universal index was then constructed by training CatBoost on standardized UHI data from four cities and deriving local statistics (mean and standard deviation) from a 30-year ERA5 climatology. This index had an average R2 = 0.49 on the training cities, generalized successfully to unseen cities, and substantially outperformed the empirical baselines. Our results demonstrate the feasibility of a universal UHI index that generalizes across diverse Arctic settings, enabling climatological assessments of urban heat potential and supporting territorial planning from regional to pan-Arctic scales, including future climate projections.