High-resolution European soil property maps leveraging foundation-model Earth observation embeddings

High-resolution soil property maps (SPMs) are of high relevance on multiple spatial scales. At field-scale, highly resolved knowledge of soil properties can inform soil management zone delineation for precision agriculture applications like irrigation, fertilisation or compaction risk management. On regional to national scales, high-resolution SPMs can aid in formulating soil policy with accurate baselines and realistic improvement goals. They may assist in soil unit definition, aiding in the balancing act between scientific precision and geographic cohesion, i.e. administrative effectiveness. On a continental, European scale, these SPMs can inform on the adequacy of Soil Monitoring Law implementation and can increase the representativity of continental soil sampling campaigns.

Contributing to this effort, we produced 10 m-resolution maps of soil pH, bulk density, coarse fragments, and texture (sand, silt, clay). This represents a large increase in spatial resolution compared to previously published maps, whose cells cover at least 250 x 250 m. We generated our maps by training shallow property-specific artificial neural networks (ANNs) on soil sample data across the EU-27 + the United Kingdom. We used soil data from the Land Use/Coverage Area frame statistical Survey (LUCAS): texture and coarse fragments data from the 2009 campaign, and pH and bulk density data from 2018. Instead of collecting a large number of spatial covariate data layers, we used the AlphaEarth Foundations satellite embeddings, produced by Google DeepMind. The satellite embeddings are the result of a representation learning model trained to efficiently compress Earth observation data. In this way, data from various sources is represented as a 64-dimensional vector per 10 m pixel. We extracted the 2018 embeddings as predictors for all soil properties. After training the models, we produced the full maps on Google Earth Engine (GEE) using a hand-implemented ANN with five neurons in the hidden layer, followed by a dropout layer. The dropout layer gave us the opportunity to additionally provide a prediction uncertainty map per soil property. A single output model was used to produce pH, bulk density and coarse fragments, while a multioutput NN generated the three texture components. Model performances varied, with R² ranging from 0.21 (coarse fragments) to 0.69 (pH). We validated our maps using a variety of field- and regional-scale soil datasets, furthermore by comparison of the value distributions with LUCAS data on European and national scales, and by visually contrasting with published soil property maps. We look forward to testing the temporal stability of our models once LUCAS 2022 data are available. Our soil maps will be necessary and useful for the scientific community across scales, from the field to the continent.