AI-Based Prediction of Urban Air Temperature in Baghdad Using Long-Term Climate Reanalysis and Urbanization Proxy Features (1940–2025)

Abstract

This study develops a machine-learning framework for daily urban air-temperature prediction in Baghdad by combining long-term meteorological reanalysis with transparent, engineered proxy variables representing progressive urban transformation. Daily weather data for 1940–2024 were obtained for central Baghdad from the Open-Meteo Historical Weather API, while 2025 was reserved as a temporal holdout period. The predictor set includes mean, maximum, and minimum air temperature, precipitation, wind speed, cyclic seasonal encodings, and heuristic indices for cement expansion, vegetation decline, asphalt growth, anthropogenic heat, and drought. Missing values are handled by median imputation and invalid observations are removed before model fitting. A Random Forest regressor with 400 trees and a maximum depth of 18 is used because of its ability to model nonlinear interactions among heterogeneous predictors. The source implementation reports approximate 2025 performance ranges of 1.5–2.5 °C for MAE, 2–3 °C for RMSE, and 0.85–0.95 for R². Feature-importance results indicate that seasonal terms dominate predictive skill, whereas urbanization proxies provide comparatively smaller marginal contributions. The framework is therefore best interpreted as a climate-aware predictive prototype rather than a causal urbanization model. The study contributes a reproducible structure for long-horizon urban temperature modelling in data-scarce hot-arid cities and identifies clear directions for improvement using satellite-derived land-cover, humidity, radiation, and independently measured urban indicators.