Machine Learning Methods in Site‐Specific Management Research: An Australian Case Study
Matthew L. Adams, Simón Cook, Peter A. Caccetta, Matthew J. Pringle · ASSA, CSSA and SSSA · 1999
Two machine learning methods based on the induction of regression trees and Bayesian networks from data were used to predict wheat yield in 1996 from fourteen soil, remotely sensed, and other variables for a field located near Wyalkatchem, Western Australia. The regression tree model explained 39.7% of the variation in the data, while the Bayesian network model explained 69.1% of the variation. Both models predicted similar spatial trends in yield, and both models predicted yield to within ±0.25 t ha-1 of measured yield >50% of the area of the field. In comparing model types, both models suffered from an inability to predict yields >0.33 t ha-1 or yields >1.89 t ha-1. We suggest that Bayesian network models are more suitable frameworks for site specific management research than regression trees, or other machine learning methods similar to regression trees.