Comparison of soil property predictions in Lithuanian croplands using UAV, satellite, EMI data and machine learning

Combining remote and proximal sensing provides a cost-effective solution for mapping soil properties in croplands. This study assessed the potential of remote sensing based on high resolution multispectral UAV imagery (2.6 cm), satellite (Sentinel-2), and in-field measured electromagnetic induction (EMI) data for predicting six soil properties − soil organic carbon content (SOC), clay, sand, silt contents, pH, and soil water content (SWC) − across five Lithuanian agroclimatic zones. Seven modelling scenarios, using individual and combined sources of sensor data, employing a random forest model, were evaluated. To assess real-world applicability, sampling-reduction simulation were additionally performed. SOC and clay predictions achieved the highest accuracy, while silt, sand, and SWC showed acceptable accuracy only in a few sites or specific modelling scenarios. Soil pH predictions were poor across all scenarios. Prediction accuracy varied across study sites, likely influenced by climate, soil parent material, topography, and agricultural management. Sensor data resolutions (2.6 cm, 1.6 m, 10 m per pixel) significantly affected prediction accuracy. For SOC predictions, UAV and Sentinel-2 data performed best, while EMI alone was less effective. In contrast, for clay predictions, EMI data yielded the highest accuracy, emphasizing its role for soil texture assessment. Multi-sensor fusion improved model performance during training but did not consistently enhance validation accuracy across sites, highlighting important cost–accuracy trade-offs and the need for realistic performance evaluation. Overall, the results demonstrate that the benefits of multi-sensor soil mapping are property-specific and site-dependent, providing guidance for scalable and economically viable field-scale soil mapping strategies.