A multitemporal Sentinel-2 framework leveraging machine learning for urban sealing detection across climatic regions

: Urban expansion represents one of the most pervasive anthropogenic pressures on terrestrial ecosystems, with soil sealing emerging as a key process affecting hydrological dynamics, carbon cycling, and biodiversity. Despite its environmental relevance, the accurate and scalable detection of impervious surfaces remains challenging due to spectral confusion with bare soil, landscape heterogeneity, and limitations in satellite data availability. To address these issues, this study develops a multitemporal classification framework for mapping soil sealing using harmonized Sentinel-2 imagery. The novelty of this work lies in the combined use of a newly developed Sealed Urban Index (SUI) and a multitemporal, multi-index feature space designed to enhance the discrimination of impervious surfaces in heterogeneous environments. Six spectral indices, including five established metrics and the proposed SUI, are used to characterize the spectral response of land surface materials. These features are integrated into two supervised machine learning classifiers, Classification and Regression Trees (CART) and Support Vector Machines (SVM), implemented within the Google Earth Engine cloud platform. The framework is applied to three urban regions with distinct climatic and morphological characteristics, Bari (Italy), Boston (USA), and Islamabad-KPK (Pakistan), to evaluate classification accuracy, spatial transferability, and seasonal robustness. A multitemporal dataset comprising 28 Sentinel-2 images for Bari (2018-2024) and representative seasonal imagery for Boston and Islamabad (2020) is analyzed. Individual spectral indices achieve overall accuracies above 86% in Islamabad, 85% in Boston, and 78% in Bari, with Kappa coefficients exceeding 0.68 in all study areas, although slight variability is observed across sites. When combined within the machine learning framework, classification performance improves substantially, with SVM consistently outperforming CART and achieving overall accuracies exceeding 90% across all study areas (Bari, Boston, and Islamabad), with Kappa values consistently above 0.85. The results demonstrate that the proposed SUI effectively reduces spectral confusion between impervious surfaces and bare soil, while the multitemporal approach mitigates seasonal variability, particularly in winter conditions. The resulting framework provides a scalable, transferable, and computationally efficient solution for high-resolution mapping of soil sealing, supporting data-driven urban planning and sustainable land management.