Soil penetration resistance prediction based on a comparative evaluation of individual and ensemble machine learning under varying tillage, fertilization and liming treatments

This study evaluated ensemble machine learning for soil penetration resistance prediction under multiple tillage, fertilization and liming treatments, potentially reducing time-consuming field sampling. Fieldwork was conducted between 2020 and 2023 at two locations in continental Croatia, resulting in a total of 1458 samples per location during 2020 and 2021, and 972 samples in 2023. Four individual machine learning methods, including Random Forest (RF), Cubist (CUB), Support Vector Regression (SVR) and Bayesian Regularized Neural Networks (BRNN), and their ensemble were evaluated using 10-fold cross-validation in 10 repetitions for each combination of locations and years. The ensemble machine learning model achieved superior prediction accuracy in comparison to the four individual machine learning models evaluated in the study, with R² values of 0.681–0.896. Among covariates examined, soil measurement depth, day of year (DOY) of sampling and tillage were the most impactful for the optimal ensemble model, while liming had a limited effect on the soil penetration resistance prediction. These results suggest that ensemble machine learning provided a stable and accurate soil penetration resistance prediction approach, which could reduce labor requirements of future fieldwork campaigns.