Machine learning approach to estimate soil matric potential in the plant root zone based on remote sensing data

Area of Science:

• Agricultural Engineering
• Remote Sensing
• Machine Learning

Background:

• Internet of Things (IoT) sensors are increasingly used in agriculture for efficient farm management.
• High costs associated with deploying numerous soil moisture sensors in large fields limit their adoption.
• There is a need for methodologies to maintain monitoring intensity with fewer in-field sensors.

Purpose of the Study:

• To investigate the relationship between soil matric potential and satellite-derived crop evapotranspiration (ETcn).
• To evaluate machine learning models (DMP and SVR) for estimating soil moisture using this relationship.
• To assess the feasibility of reducing in-field sensor numbers for irrigation management.

Main Methods:

• Analysis of sensor data across two cotton growing seasons in Australia.
• Development and application of Dense Multilayer Perceptron (DMP) and Support Vector Regression (SVR) models.
• Utilized satellite-derived crop evapotranspiration (ETc) to estimate soil matric potential.

Main Results:

• Identified a second-degree function relationship between soil matric potential and cumulative ETcn.
• Both DMP and SVR models accurately estimated soil matric potential.
• Individual datasets achieved up to 90% accuracy (within ±10 kPa), combined datasets achieved 73% accuracy.

Conclusions:

• The developed technique accurately monitors root zone soil matric potential in cotton.
• Reduced in-field sensor requirements offer a cost-effective solution for broadacre agriculture.
• Promising applications for irrigation-decision support systems are highlighted.