Non-Invasive Soil Texture Prediction Using Machine Learning and Multi-Source Environmental Data

Accurate prediction of soil texture is essential for effective soil management, precision agriculture, and hydrological modeling. This study proposes a novel, data-driven approach for estimating soil texture without the need for laboratory-based analysis. High-frequency in situ soil moisture measurements from EnviroSCAN (Sentek Technologies, Stepney, Australia) sensors and satellite-derived vegetation indices (NDVI) from Sentinel-2 were collected across 25 sites in Hungary. Temporal soil moisture dynamics were encoded using a Long Short-Term Memory (LSTM) neural network, designed to capture soil-specific hydrological response behavior from time-series data. The resulting latent embeddings were subsequently used within an ordinal regression framework to predict ordered soil texture classes, explicitly enforcing physical consistency between classes. Model performance was evaluated using leave-one-soil-out cross-validation, achieving an overall classification accuracy of 0.54 and a mean absolute error (MAE) of 0.50, indicating predominantly adjacent-class errors. The proposed approach demonstrates that soil texture can be inferred from dynamic environmental responses alone, offering a transferable alternative to fraction-based regression models and supporting scalable sensor calibration and digital soil mapping in data-scarce regions.