Optimizing Soil Moisture Sensor Depth for Irrigation Management Using Universal Multiple Linear Regression

Abstract

HighlightsUniversal multiple linear regression (uMLR) was used to optimize soil moisture sensor depth.Eight years of soil moisture data from various water treatments were tested, then validated.The best single observation depth for irrigation management is at 30 cm below soil surface.A single calibrated soil moisture sensor may be used to aggregate soil water deficit in the root zone.Abstract. Soil moisture sensors are valuable tools in irrigation management, but due to soil spatial variability it is challenging for point-based sensors to characterize the entire soil water profile. Universal multiple linear regression (uMLR) is a simple machine learning tool that analyzes the relationship between all possible subsets of observations and the prediction of interest. This study used uMLR and volumetric water content data (VWC) taken from maize field experiments at the Limited Irrigation Research Farm near Greeley, Colorado, to optimize the number and placement of soil moisture sensors to characterize the entire soil water deficit (SWD) in the root zone. The uMLR analysis quantified the accuracy of using one or two VWC measurements to represent SWD. We evaluated seven field seasons, with 12 treatments and four replicates of each, where volumetric VWC was typically measured one to three times weekly. Results indicate that the best single sensor depth is at 30 cm, and the best combination of two sensors is at 0 to 15 cm and 60 cm. To validate the experiment, a continuous VWC sensor was used at 30 cm to predict SWD results with acceptable accuracy (RMSE = 4.60 mm). This study suggests that in the sandy loam site evaluated and with varying irrigation levels, soil moisture sensor placement near the soil surface along with knowledge of soil texture can be used to characterize SWD. This technique may be transferable to other locations and soils and shows potential for optimizing irrigation management via soil moisture sensor networks with continuous SWD data. Keywords: Deficit irrigation, Machine learning, Soil moisture sensor, Soil water deficit, Universal multiple linear regression, Volumetric water content.