A Comparative Analysis of Satellite Imagery for Land Cover Classification: Evaluating Google Satellite Embeddings, Sentinel-2, And Landsat-8 Data with XGBoost

Accurate land cover mapping is fundamental for sustainable urban planning and environmental monitoring, yet the optimal data source for classification in complex landscapes remains a subject of research. This study presents a comparative analysis of land cover classification performance using three distinct satellite-derived data sources: traditional multispectral imagery from Landsat 8 and Sentinel-2, and the novel Google Satellite Embeddings (V1), a deep learning-based feature representation. The research focused on the rapidly urbanizing Greater Accra Region of Ghana, a challenging coastal metropolitan area. A supervised classification approach was implemented using the Extreme Gradient Boosting (XGBoost) algorithm to ensure a consistent methodological framework. Results demonstrated a clear hierarchy in performance. The Google Embeddings achieved superior accuracy, with an Overall Accuracy (OA) of 98% and a Kappa coefficient of 0.973, significantly outperforming Sentinel-2 (OA: 94.75%, Kappa: 0.93) and Landsat 8 (OA: 93.8%, Kappa = 0.92). The embeddings excelled particularly in discriminating the most challenging classes, Urban and Bare Land, where spectral confusion typically occurs, attaining F-scores of 97.35% and 95.95%, respectively. This performance advantage is attributed to the embeddings' ability to encapsulate rich spatial and structural patterns beyond raw spectral information. The findings highlight a shift in remote sensing, showing that AI-derived feature spaces can significantly improve classification accuracy compared to traditional spectral data. This provides a valuable tool for precise land cover mapping, which is vital for informed decision-making in ever-changing urban areas. The discrepancies in land cover areas also emphasis how the choice of data sources can directly influence environmental monitoring and urban planning.