Machine and deep learning-based forecast of Tehran’s accelerating nighttime surface heat island crisis

Abstract

The nighttime Surface Urban Heat Island (SUHI) phenomenon in the metropolis of Tehran poses severe risks to urban climate and sustainability, representing a potential crisis for the entire Middle East and North Africa (MENA) region. A critical research gap persists in long-term predictive modeling for rapidly urbanizing arid regions like Tehran, particularly using advanced hybrid artificial intelligence (AI) architectures that integrate spatial and temporal learning. This study addresses this gap by developing and benchmarking a novel ensemble of six hybrid models, including three machine learning (ML) and three deep learning (DL) models to forecast nocturnal SUHI in Tehran through 2030. The methodology leverages a 22-year MODIS Aqua LST time series, processed with a rigorous spatiotemporal gap-filling process, to train models including innovative combinations such as U-Net with Long Short-Term Memory (LSTM) networks. The primary objective is to provide a robust, actionable forecasting framework aligned with Sustainable Development Goals (SDGs). Key findings indicate that parametric normal distribution models suggest temporal land surface temperature (LST) will rise by more than 2°C across four study areas (Tehran, Tehran Downtown, the suburbs, and the combined Tehran–suburbs), a trend further supported by non-parametric approaches, which also show that Tehran Downtown has the highest spatial LST. R-squared (R2), root mean square error (RMSE), and mean absolute error (MAE) were used to assess and compare the three ML and three DL models. The optimal model (U-Net + LSTM) projects a sustained annual SUHI intensity (SUHII) increase of 0.08 °C, potentially reaching a mean of 5.96 °C by 2030—a rise of 0.11–0.39 °C above historical levels. Spatial analyses of SUHI, supported by longitudinal and latitudinal profiles, reveal that Tehran is transitioning toward a “super-nighttime SUHI” state. Moreover, statistical and practical analyses confirm that warming is accelerating and will be disproportionately concentrated in the extreme upper percentiles, significantly amplifying the frequency and severity of hot nights. The broader significance of this work lies in its proof-of-concept for an AI-driven urban climate forecasting framework, which provides urban planners and policymakers with evidence-based insights for targeted mitigation and climate adaptation strategies, thereby supporting the development of resilient urban systems in Tehran and similar metropolitan areas worldwide.