Cloud-Based Prediction of Soil Properties Using Remote Sensing Data and Machine Learning for Smart Farming Applications

Smart Farming (SF) practices are essential for reducing agricultural impacts on ecosystems while maintaining food security. However, the implementation of SF is often hindered by the lack of high-resolution soil property data. This study addresses this challenge by developing a cloud-based approach to predict soil texture (clay, silt, and sand) using a random forest machine learning model within Google Earth Engine (GEE) at three spatial scales: farm, regional, and national.

The analysis was conducted at four farm sites in the German states of Brandenburg and Mecklenburg-Vorpommern (with 355, 321, 392, and 151 topsoil samples), across the eastern part of Brandenburg at the regional scale (1,080 samples), and nationwide across Germany (2,199 samples). Soil data were sourced from smart farming projects and the LUCAS soil database. The datasets were split into 70% for model training and 30% for validation.

The input earth observation (EO) data included optical and radar remote sensing information from Sentinel-1 (S1) and Sentinel-2 (S2) satellites. Vegetation indices, soil indices, and bare soil pixels were calculated from S2 data, while S1 provided radar backscatter values (VV and VH polarizations). Temporal patterns were captured through statistical metrics such as mean, standard deviation, and coefficient of variation. Finally, the 71 raster datasets at farm scale and 55 raster datasets at regional and national scale were extracted at soil sampling locations and used as covariates in the random forest models. Model performance was evaluated using root mean square error (RMSE). At the farm scale, RMSE values ranged from 4.1% to 5.8% (R² 0.36 to 0.76) for clay, 5.3% to 8.7% (R² 0.35 to 0.51) for silt and 8.9% to 10.9% (R² 0.4 to 0.72) for sand. At the regional scale, RMSE values were 6.4% (R² 0.58) for clay, 6.5% (R² 0.4) for silt, and 10.7% (R² 0.46) for sand. At the national scale, clay predictions remained consistent with an RMSE of 6.9% (R² 0.49), while RMSE values for silt and sand increased to 11.1% (R² 0.51) and 14.8% (R² 0.56), respectively.

Key predictors across scales were S2 bands 11 and 12 (under bare soil conditions), S1 VV and VH backscatter, the VV-VH ratio, and elevation data from the Copernicus Digital Elevation Model. The influence of EO data was highest at farm and regional scales but diminished at the national level.

The developed models, implemented in GEE, can predict topsoil texture (0–30 cm depth) at a resolution of 10 m × 10 m for any arable field in Germany. This approach could help to increase the availability of high-resolution soil data for smart farming applications.