Beyond Standalone Geo-Embeddings: Weighted Multi-Model Ensemble Prediction for Tropical Land-Cover Mapping

Abstract

Accurate land-cover mapping in tropical regions remains challenging because of high environmental heterogeneity, complex vegetation structure, and persistent cloud cover. Recent geospatial foundation models provide pre-computed Geo-embeddings that offer a new representation of Earth Observation data (EO), yet their potential for detailed tropical land-cover mapping and their performance relative to, and in combination with, conventional Sentinel-1 and Sentinel-2 satellite image time series remains largely unexplored. This study evaluated AlphaEarth Foundation Geo-embeddings for detailed vegetation mapping in northern Madagascar and investigated whether combining them with satellite image time series processed using the Satellite Image Time Series (SITS) R package could improve classification performance. Geo-embeddings were classified using a Random Forest model (GEO), while four supervised classifiers were trained on Sentinel-1 and Sentinel-2 time series. GEO achieved the highest standalone performance (OA = 0.74), outperforming all classifiers trained on the satellite image time series. Probability-level ensemble models were then used to assess whether both data representations could be beneficially combined. The best-performing ensemble, combining GEO (75%) with Temporal Convolutional Neural Network (TempCNN; 25%), increased overall accuracy from 0.74 to 0.82. These results suggest that Geo-embeddings provide an effective standalone representation for detailed tropical land-cover mapping and indicate that combining them with conventional satellite image time series may exploit complementary information. As geospatial foundation models continue to evolve, understanding how they can be integrated with established Earth observation workflows may support future operational land-cover mapping.