Advancing Built-Up Area Monitoring Through Multi-Temporal Satellite Data Fusion and Machine Learning-Based Geospatial Analysis

Monitoring built-up dynamics is essential for sustainable urban and territorial planning. This study presents an innovative geospatial methodology integrating multi-temporal satellite data fusion, transfer learning, machine learning classification, and open-access cloud computing to systematically identify, quantify, and map the spatiotemporal evolution of built-up areas. The methodology was applied at a territorial scale in southern Italy using Landsat multispectral imagery acquired and elaborated through Google Earth Engine. Compared to more conventional classification methods, the proposed integrated approach ensures scalability, reproducibility, and computational efficiency. Landsat multispectral imagery from 2006 to 2024 was classified using a Random Forest (RF) algorithm, trained and validated with CORINE Land Cover maps for 2006, 2012, and 2018. For 2024, a transfer learning strategy was adopted, enabling classification through a model fine-tuned with historical data and validated independently. Accuracy assessment returned an Overall Accuracy (OA) of 0.890 and F1-scores between 0.803 and 0.811 for 2006–2018. For 2024, the OA reached 0.926 with an F1-score of 0.926, confirming the effectiveness of the proposed framework. This integrated methodology not only allows for determining the extent of urban expansion over the considered timelines, but, by introducing two spatial metrics, Urban Density and the Urban Dispersion Index (UDI), also enables the characterization of the morphological evolution of urban growth. The methodology ensures spatial and temporal consistency, offering a scalable and automated framework for long-term monitoring that provides a decision support tool for urban growth management and environmental planning, especially in data-limited contexts.