TITLE:
Boosting Algorithm: An Ensemble Learning Tool for Land Use Land Cover Classification Using Google Alpha Earth Foundations Satellite Embeddings
AUTHORS:
Saviour Mantey, Isaac Selasi Kojo Attipoe, Joel Kwabena Lawerteh
KEYWORDS:
LULC, XGBOOST, AdaBoost, Gradient Boosting Machine, AlphaEarth Foundations
JOURNAL NAME:
Advances in Remote Sensing,
Vol.15 No.1,
January
14,
2026
ABSTRACT: Accurate land use/land cover (LULC) classification remains a persistent challenge in rapidly urbanising regions especially, in the Global South, where cloud cover, seasonal variability, and limited ground truth data often undermine the reliability of traditional remote sensing approaches. While recent advances in satellite embedding models, such as AlphaEarth Foundations (AEF) offering, task-ready representations of Earth’s surface, there is limited empirical evidence on how different machine learning algorithms perform when applied to these embeddings, particularly in African urban contexts. This study addresses that gap by evaluating the performance of four leading boosting algorithms (Light Gradient Boosting Machine (LightGBM), Categorical Boost (CatBoost), AdaBoost, and XGBoost) used in conducting supervised classification for four dominant land cover types (Urban, Water, Bare Land, and Vegetation) in the Greater Accra Area, Ghana, using AEF embeddings. The findings indicate that LightGBM achieved the highest classification performance, with a superior F-score across all classes, an overall accuracy (OA) of 98.35%, and a Kappa coefficient of 0.978, reflecting excellent agreement with ground truth labels. XGBoost followed closely, with an OA of 98% and a Kappa value of 0.973. CatBoost ranked third, with an OA value of 97.5% and a Kappa value of 0.966. AdaBoost performed least effectively with an OA value of 96.7% and a Kappa value of 0.955. These results provide further clarity and insights into the comparative performance of boosting algorithms when applied to AEF embeddings for supervised land use/land cover (LULC) classification. Additionally, the results highlight the potential of embedding-based approaches for scalable and interpretable land cover mapping in complex, data-constrained urban environments.