Application of statistical and neural network model for oil palm yield study

Abstract

This thesis presents an exploratory study on modelling of oil palm (OP) yield using statistical and artificial neural network approach. Even though Malaysia is one of the largest producers of palm oil, research on modelling of OP yield is still at its infancy. This study began by exploring the commonly used statistical models for plant growth such as nonlinear growth model, multiple linear regression models and robust M regression model. Data used were OP yield growth data, foliar composition data and fertiliser treatments data, collected from seven stations in the inland and coastal areas provided by Malaysian Palm Oil Board (MPOB). Twelve nonlinear growth models were used. Initial study shows that logistic growth model gave the best fit for modelling OP yield. This study then explores the causality relationship between OP yield and foliar composition and the effect of nutrient balance ratio to OP yield. In improving the model, this study explores the use of neural network. The architecture of the neural network such as the combination activation functions, the learning rate, the number of hidden nodes, the momentum terms, the number of runs and outliers data on the neural network’s performance were also studied. Comparative studies between various models were carried out. The response surface analysis was used to determine the optimum combination of fertiliser in order to maximise OP yield. Saddle points occurred in the analysis and ridge analysis technique was used to overcome the saddle point problem with several alternative combinations fertiliser levels considered. Finally, profit analysis was performed to select and identify the fertiliser combination that may generate maximum yield