Geospatial Modeling of Soil Properties Using Satellite Embeddings and Machine Learning

Machine learning is increasingly used in environmental monitoring to fill the gaps between sparse field data and satellite observations. This study evaluates multiple machine learning models using satellite embedding features for predicting soil properties and land characteristics across large geographic areas. We combine georeferenced data from the LUCAS 2018 Soil Module with 64-dimensional embeddings from Google’s Satellite Embedding Dataset (SED), which capture spectral and spatial characteristics of the Earth’s surface. These embeddings are used as input to supervised machine learning models for predicting key soil properties, such as organic carbon and pH, and land cover classes. All models are validated using field measurements from the LUCAS dataset. The results demonstrate the ability of satellite-based feature extraction to be used for predicting soil properties that cannot be observed directly from space. By generalizing from point-based measurements to continuous maps, this approach enables high-resolution modeling of soil and environmental variables. The system is designed to be scalable, and future work will include more parameters, models, and continental coverage to support soil monitoring, carbon estimation, and environmental policy planning.