Predicting Daily Aerobiological Risk Level of Potato Late Blight Using C5.0 and Random Forest Algorithms under Field Conditions

Abstract

Late blight, caused by Phytophthora infestans, is a major disease of the potato crop with a strong negative impact on tuber yield and tuber quality. The control of late blight in conventional potato production systems is often through weekly application of prophylactic fungicides, moving away from a sustainable production system. In support of integrated pest management practices, machine learning algorithms were proposed as tools to forecast aerobiological risk level (ARL) of Phytophthora infestans (>10 sporangia/m³) as inoculum to new infections. For this, meteorological and aerobiological data were monitored during five potato crop seasons in Galicia (northwest Spain). Mild temperatures (T) and high relative humidity (RH) were predominant during the foliar development (FD), coinciding with higher presence of sporangia in this phenological stage. The infection pressure (IP), wind, escape or leaf wetness (LW) of the same day also were significantly correlated with sporangia according to Spearman's correlation test. ML algorithms such as random forest (RF) and C5.0 decision tree (C5.0) were successfully used to predict daily sporangia levels, with an accuracy of the models of 87% and 85%, respectively. Currently, existing late blight forecasting systems assume a constant presence of critical inoculum. Therefore, ML algorithms offer the possibility of predicting critical levels of Phytophthora infestans concentration. The inclusion of this type of information in forecasting systems would increase the exactitude in the estimation of the sporangia of this potato pathogen.

Keywords: Phytophthora infestans; Solanum tuberosum L.; aerobiology; infection pressure; machine learning; weather factors.

Figure 1

Visual abstract about the inputs (weather and aerobiological variables) used in the machine learning (ML) algorithms. ARL: daily aerobiological risk level.

Figure 2

Receiver operating curve (ROC) for comparing the machine learning models random forest (RF) and C5.0 decision tree (C5.0). The broken lines in the ROC curve represent a classifier with no predictive value (diagonal line) and a perfect classifier (horizontal and vertical lines). ROC curves that are closer to the perfect classifier show a better model.

Figure 3

Comparison between prediction data by ML algorithms C5.0 and random forest (RF) and observed aerobiological risk level (ARL) by growing season. DAE: days after emergence.