Simulation of arthropod abundance from plant composition

The relationship between arthropod abundance and plant composition is extremely complex. It is very hard to develop a mechanistic model to describe the relationship. This study aimed to simulate arthropod abundance from plant composition on grassland using an artificial neural network developed by the author, and to compare simulation performances between the neural network and conventional models. The results revealed that there were complex interactions between plants and arthropods, and the arthropod abundance on grassland was significantly determined of plant families and their cover-degrees rather than plant species and their cover-degrees. Neural network exhibited a better simulation performance than multivariate regression and response surface model. Cross validation indicated that prediction performance of neural network was also superior to these models. It was concluded that neural network is an effective tool to model arthropod abundance from plant composition on grassland. A moderate dimensionality for input space may be determined to produce a reasonably trained neural network. Such procedures for dimensionality reduction as PCE, etc., were suggested being used in the data treatment in neural network modeling. A high dimensionality for input space and a few samples in the input set would result in the deficient learning of neural network. Randomization procedure for sample submission would help to eliminate the sequence correlation but may result in a worse performance in simulation and prediction. It was suggested that randomization procedure could be used to the sample submission for these situations with a lot of samples and a lower dimensionality.