Data-driven approach to weekly forecast of the western flower thrips (Frankliniella occidentalis Pergande) population in a pepper greenhouse with an ensemble model

Abstract

Integrated pest management (IPM) in European glasshouses has substantially advanced in automated insect pest detection systems lately. However, transforming such an enormous data influx into optimal biological control strategies remains challenging. In addition, most biological control forecast studies relied on the single-best model approach, which is susceptible to overconfidence, and they lack validation over sufficient sampling repetitions where robustness remains questionable. Here we propose employing an unweighted ensemble model, by combining multiple forecasting models ranging from simple models (linear regressions and Lotka–Volterra model) to machine learning models (Gaussian process, Random Forest, XGBoost, Multi-Layer Perceptron), to predict 1-week-ahead population of western flower thrips (Frankliniella occidentalis), a notorious pest in glasshouses, under the influence of its biological control agent Macrolophus pygmaeus in pepper-growing glasshouses.