Forecasting the seasonal dynamics of Trichoplusia ni (Lep.: Noctuidae) on three Brassica crops through neural networks.

The cabbage looper, Trichoplusia ni Hübner (Lep.: Noctuidae), is a destructive pest of Brassica crops. Their larvae defoliate plants, leading to reduced crop yield. Understanding and modeling pest seasonal dynamics is central to management programs because it allows one to set up sampling and control efforts. This study aimed to train, with field-collected data, artificial neural networks (ANN) for T. ni forecasting on Brassica crops. ANNs were used due to their suitability to fit complex models with multiple predictors. Three weather variables (air temperature, rainfall, and relative humidity lagged at different intervals from the day of pest assessment) and three host plants (broccoli, cabbage, and cauliflower) along with another plant-related variable (days after transplanting) were used as input variables to build ANNs with different topologies. Two outputs (T. ni eggs or larvae) were tested to verify which one would yield more precise models. ANNs forecasting T. ni eggs performed better, based on Pearson's correlation (r v ) of observed with fitted values. The winning ANN (r v  = 0.706) had weather data lagged by 15 days, 2 neurons in the hidden layer, hyperbolic tangent as the activation function, and resilient propagation as the learning algorithm. Broccoli and cauliflower were the hosts with major contributions for T. ni occurrence. Rainfall was the primary environmental predictor and affected T. ni negatively. Therefore, the winning ANN may be used to forecast T. ni egg densities 15 days in advance, allowing for timely management of this pest.