Recursive partitioning framework to manage high variance and build decision trees: a study to model the effects of temperature and leaf wetness on Sclerotinia Stem Rot in soybeans

Abstract Decision trees can classify plant disease records even under high variance, such as when incidence and severity shift from unfavourable to conducive environmental conditions. This hypothesis was tested in greenhouse trials to model the interaction between leaf wetness period (LWP) and temperature and their effects on the infection and development of Sclerotinia stem rot (SSR) caused by Sclerotinia sclerotiorum in four soybean cultivars. Plants inoculated with pathogen-infected flowers were exposed to temperatures of 10, 15, 20, 25 and 30 °C in moist chambers for 4, 8, 12, 16, 20, 24, 28 or 32 h. Disease severity was categorised using k-means clustering into three classes (low, average, high) and multi-class (non-dichotomous) decision trees were constructed, typically using LWP as the root node (starting point), to partition severity. In contrast, temperature served as the root node in most dichotomous decision trees used to classify the presence or absence of SSR. Both tree types required fewer splits to predict SSR in susceptible cultivars than in partially resistant ones. The high accuracy of the dichotomous trees was confirmed by a confusion matrix, showing all predictions were correct (accuracy = 1.0). SSR incidence was observed at temperatures in the 10–30 ºC range when LWP lasted at least 16 h. Across all cultivars, LWP was more important than temperature for identifying SSR severity classes in the three class (non-dichotomous models), whereas the binary (dichotomous) models more accurately predicted disease incidence.