Prediction of pasture condition in the Kimberley rangelands of Western Australia using simple classification tree models

Pasture condition assessments assist pastoralists, regulators, and policy makers in making informed decisions around livestock production and the preservation of natural resources in the Kimberley rangelands of Western Australia (WA). Reliable qualitative assessments of pasture condition require assessors with extensive expertise, which means frequent and reproducible assessments can be difficult to achieve. To develop a quantitative approach that can complement existing qualitative assessment approaches in the Kimberley, we investigated the use of simple classification tree models to predict pasture condition using assessment data from the Western Australian Rangeland Monitoring System (WARMS). Quantitative traits were derived from WARMS observation data for three major pasture groups in the Kimberley region and assessed for variable importance using random forest models. The traits with highest variable importance were used to train classification tree models for each pasture group. Important variables identified from random forests and classification trees were the percentage of quadrats containing at least one of a set of desirable perennial species and the percentage of quadrats without perennials. Classification trees selected to represent each pasture group had prediction accuracies between 0.71 and 0.81 when applied on reserved WARMS testing data, and between 0.65 and 0.74 when applied on independently collected validation data, indicating they generally performed well when used on data from the Kimberley rangelands. The classification tree models generated here can be used to complement qualitative assessments of pasture condition and to provide paddock-scale assessments, helping to inform decision making in the management of important WA pastoral resources.