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

There is an increasing interest in using the Internet of Things (IoT) in the agriculture sector to acquire soil- and crop-related parameters that provide helpful information to manage farms more efficiently. One example of this technology is using IoT soil moisture sensors for scheduling irrigation. Soil moisture sensors are usually deployed in nodes. A more significant number of sensors/nodes is recommended in larger fields, such as those found in broadacre agriculture, to better account for soil heterogeneity. However, this comes at a higher and often limiting cost for farmers (purchase, labour costs from installation and removal, and maintenance). Methodologies that enable maintaining the monitoring capability/intensity with a reduced number of in-field sensors would be valuable for the sector and of great interest. In this study, sensor data analysis conducted across two irrigation seasons in three cotton fields from two cotton-growing areas of Australia, identified a relationship between soil matric potential and cumulative satellite-derived crop evapotranspiration (ET) between irrigation events. A second-degree function represents this relationship, which is affected by the crop development stage, rainfall, irrigation events and the transition between saturated and non-saturated soil. Two machine learning models [a Dense Multilayer Perceptron (DMP) and Support Vector Regression (SVR) algorithms] were studied to explore these second-degree function properties and assess whether the models were capable of learning the pattern of the soil matric potential-ET relation to estimate soil moisture from satellite-derived ET measurements. The algorithms performance evaluation in predicting soil matric potential applied the k-fold method in each farm individually and combining data from all fields and seasons. The latter approach made it possible to avoid the influence of farm consultants' decisions regarding when to irrigate the crop in the training process. Both algorithms accurately estimated soil matric potential for individual (up to 90% of predicted values within ±10 kPa) and combined datasets (73% of predicted values within ±10 kPa). The technique presented here can accurately monitor soil matric potential in the root zone of cotton plants with reduced in-field sensor equipment and offers promising applications for its use in irrigation-decision systems.