Machine learning for supporting irrigation decisions based on climatic water balance

Abstract

A machine learning model was developed to support irrigation decisions. The field research was conducted on ‘Gala’ apple trees. For each week during the growing seasons (2009–2013), the following parameters were determined: precipitation, evapotranspiration (Penman–Monteith formula), crop (apple) evapotranspiration, climatic water balance, crop (apple) water balance (AWB), cumulative climatic water balance (determined weekly, ΣCWB), cumulative apple water balance (ΣAWB), week number from full bloom, and nominal classification variable: irrigation, no irrigation. Statistical analyses were performed with the use of the WEKA 3.9 application software. The attribute evaluator was performed using Correlation Attribute Eval with the Ranker Search Method. Due to its highest accuracy, the final analyses were performed using the WEKA classifier package with the J48graft algorithm. For each of the analysed growing seasons, different correlations were found between the water balance determined for apple trees and the actual water balance of the soil layer (10–30 cm). The model made correct decisions in 76.7% of the instances when watering was needed and in 87.7% of the instances when watering was not needed. The root of the classification tree was the AWB determined for individual weeks of the growing season. The high places in the tree hierarchy were occupied by the nodes defining the elapsed time of the growing season, the values of ΣCWB and ΣAWB.

Keywords

• apple trees
• evapotranspiration
• irrigation scheduling
• machine learning
• precipitation
• weka software