Automated Land Surface Temperature Estimation and Prediction using Multispectral and Multi-Temporal Indices in Google Earth Engine: Bizerte study Case, Northern Tunisia

Abstract

This study analyzes the spatiotemporal evolution and near-future trajectory of land surface temperature (LST) in Bizerte, a rapidly urbanizing Tunisian coastal governorate. Using multi-temporal Landsat imagery (2009–2024) processed in Google Earth Engine, we quantify bi-seasonal warming and project thermal conditions to 2029. Over 2009–2024, mean summer LST rose by 2.8 °C (to 41.7 °C) and winter LST by 2.5 °C (to 18.9 °C), with urban and bare-soil areas warming most intensely (+5.3 °C). This coincided with increased built-up density and marked declines in vegetation and moisture indices. A Random Forest regression model, trained on 2009–2019 data and validated on 2024 observations, achieved robust performance (R2 = 0.875). SHAP analysis identified the Urban Thermal Field Variance Index (UTFVI) and built-up intensity as the strongest warming drivers, while vegetation and water indices exerted the greatest cooling influence. Projections to 2029 under a Business-as-Usual scenario indicate extensive extreme summer heat zones (39.9–46.9 °C), covering over 44% of the governorate, with intense hotspots along the Bizerte-Menzel Bourguiba urban-industrial corridor. Conversely, an ambitious Green-City scenario confines extreme heat (≥ 37.9 °C) to only ∼20% of the area, primarily within dense urban cores. Forests and wetlands consistently provide 10–18 °C of localized cooling. We conclude that local land-use change is the primary driver of near-term thermal amplification in this Mediterranean coastal city, and that proactive green-blue infrastructure expansion is a critical and immediately deployable mitigation pathway. The developed open-source workflow offers a transferable framework for thermal risk assessment in similar urbanizing regions.