Spatio-temporal prediction of soil moisture content at various depths in three soil types using machine learning algorithms

Tarek Alahmad^1*Miklós Neményi¹Adrienn Széles²Nour Ali³Omar Hijazi⁴Anikó Nyéki¹

• ¹Széchenyi István University, Albert Kázmér Mosonmagyaróvár Faculty of Agricultural and Food Sciences, Department of Biosystems Engineering and Precision Technology, 9200, Mosonmagyaróvár, Hungary
• ²Institute of Land Use, Engineering and Precision Farming Technology, Faculty of Agricultural and Food Sciences and Environmental Management, University of Debrecen, 138 Böszörményi street, 4032, Debrecen, Hungary
• ³Laboratory of Crop Production and Multiplication, Field Crops Research Department, Agricultural Engineering Faculty, Damascus University,, Damascus, Syria
• ⁴A Chair of Wood Science, Technical University of Munich, 85354,, Freising, Germany

Accurate prediction of soil moisture content (SMC) is crucial for agricultural systems as it affects hydrological cycles, crop growth, and resource management. Considering the challenges with prediction accuracy and determining the effect of soil texture, depth, and meteorological data on SMC variation and prediction capability of the used models this research has been conducted. Three machine learning (ML) models Random Forest Regression (RFR), eXtreme Gradient Boosting (XGBoost), and Long-Short Term Memory (LSTM) were developed to predict SMC in three soil types (loam, sandy loam, and silt loam) at five depths of 5, 20, 40, 60, and 80cm. The results showed variations in SMC in all studied soil types and depths, with silt loam exhibiting the highest variation in SMC. RFR demonstrated high accuracy at different depths and soil types, particularly, in loam soil, at a depth of 80 with root mean square error (RMSE) values of 0.89 and mean absolute error (MAE) of 0.74, and in silt loam at 40cm depth with RMSE of 0.498, and MAE of 0.416. LSTM performed effectively at shallower and moderate depths (60 and 20 cm) with RMSE values of 0.391 and 0.804 and MAE of 0.335 and 0.793, respectively. In sandy loam soil at 5cm depth, XGBoost displayed minimal errors and robust performance at the same depths with higher accuracy achieving RMSE of 0.025 and MAE of 0.159. Analysis of training and validation loss revealed that LSTM model stabilized and improved with more epochs, showing a more consistent decrease in MSE, while RFR and XGBoost exhibited higher performance with increased model complexity, shown in low MSE and RMSE values. Comparisons between measured and predicted SMC% values demonstrated the models' effectiveness in capturing soil moisture dynamics. Furthermore, feature importance analysis revealed that solar radiation and precipitation were the most influential predictors across all models, offering critical insights into dominant environmental drivers of soil moisture variability. By providing precise SMC predictions across different spatial and temporal scales, this study underscores the value of ML models for SMC prediction, which could have implications for improving irrigation scheduling, reducing water wastages, and enhancing sustainability