Neural network modeling of survival dynamics of holometabolous insects: A case study

Abstract

Survival process and mortality distribution of holometabolous insects were hard to be described by mechanistic models due to their distinctive development stages in the life cycle. Neural networks are flexible approximators for linear or nonlinear ecological systems. This study aimed to evaluate the effectiveness and performance of BP ANN (feed-forward backpropagation artificial neural network) and conventional models in modeling the survival process and mortality distribution of a holometabolous insect, Spodoptera litura F. (Lepidoptera: Noctuidae). Training data on survival process and mortality distribution of S. litura were recorded under different temperatures. BP ANN, three empirical models, five probabilistic density functions, a multi-stage based dynamic model, and a trend surface model were used to modeling the time changing and temperature dependent relationships of the insect. Overall performances were compared among these models. The results demonstrated that BP ANN could be effectively used to model the survival process and mortality distribution of S. litura. It exhibited the best performance in the one-dimensional and two-dimensional simulations. Some features in survival process of the holometabolous insect, like the static stages for egg and pupa, were better simulated by BP ANN. Cross validation and prediction analysis demonstrated that BP ANN was also a better and robust predictor for the survival process of S. litura. Among the probabilistic density functions such as normal function, logarithm-normal function, Cauchy function, and chi-squared function, the performance of Cauchy density function was similar to BP ANN in the simulation of mortality distribution, and chi-squared function was the worst. The multi-stage based dynamic model yielded the worst performance in survival process simulation. It was proved that the network settings, such as the number of hidden layers and hidden neurons influenced the performance of BP ANN.