Bridging Data Gaps in Soil Matric Potential for Enhanced Water Management

Handling unstructured and missing data (UMD) remains a significant challenge in environmental monitoring and precision agriculture. This study focuses on the imputation of UMD in soil matric potential (SMP) datasets, a critical parameter in assessing soil water availability and managing irrigation systems. Missing data can distort trends, complicate analysis, and hinder decision-making in critical areas such as water management and precision irrigation. Using Extreme Learning Machine (ELM) and Time Series Models with Exogenous Inputs (TSMX), the research reconstructs missing SMP records by integrating adjacent sensor datasets and explanatory environmental variables. This approach demonstrates the potential of advanced data-driven techniques to enhance the reliability of agricultural and hydrological datasets. The dataset encompasses hourly SMP measurements and explanatory variables, including meteorological inputs such as relative humidity, air temperature, and soil properties, collected across multiple sensors in a precision agriculture setup. Exploratory analysis revealed variations in data structure, including non-stationary trends and significant statistical differences between training and testing datasets. These insights guided the selection of inputs and model configurations, emphasizing the importance of autocorrelation analysis in determining the most significant predictors. The ELM model exhibited superior performance in imputing missing SMP values, achieving an R-value of 0.992, RMSE of 0.164 cm, and NSE of 0.983 using five key inputs. This robustness highlights ELM's capability to generalize across diverse input combinations effectively. Additionally, TSMX has also been explored for its potential to leverage temporal dependencies and explanatory variables for consistent imputation. The incorporation of adjacent sensor data in modeling efforts underscores the importance of spatial and temporal relationships in enhancing accuracy, particularly in heterogeneous environmental conditions. This research underscores the critical role of input selection and model tuning in addressing UMD in SMP datasets. The findings demonstrate the complementary strengths of ELM and TSMX, offering practical insights for improving data reliability in precision irrigation and environmental monitoring. Future studies could explore integrating additional explanatory variables and employing advanced machine learning architectures to optimize imputation performance under varying environmental conditions further.

Keywords: Missing Data Imputation; Soil Matric Potential; Extreme Learning Machine; Time Series Models; Exogenous Inputs; Precision Agriculture; Environmental Monitoring.