The potential for using artificial neural networks to model radiata pine branch diameter distribution patterns

Understanding diameter distribution patterns of branches within whorls in radiata pine (pinus radiata D. Don) is crucial/important to the eventual development of automation designed to detect defects (knots) on logs, which aims to enhance the quality of sawn timber. This study evaluated the feasibility of using artificial neural networks (ANN) to model branch diameter distribution patterns. The approach was to use information on branch size and distribution on the top surface (e.g. from a scanned image) to predict the largest branch diameter on the bottom half log. The branch with the largest diameter determines the extent of the defect, and is therefore of greatest concern. Two data sets were used to train and evaluate the ANN’s. The first data set consisted of data collected in the field from 290 trees (that were cut into 314 logs), with 12800 branches measured in 1461 whorls. The second data set (the artificial data set) contained information for 30069 branches on 50 artificial stems. This data set was generated by the Branch Growth Model developed by scientists in New Zealand Forest Research Institute. Initially the ANN’s were trained using individual branch information (i.e., diameter of branches in each whorl) in the real data set. To do this, multilayer feed-forward networks with back propagation algorithm were used. If information for branches in quadrants 1 and 2 (top half whorl) was used as input, then the largest branch diameter in quadrants 3 and 4 (bottom half whorl) was set as the desired output, and vice versa. Networks trained using individual branch information in the real data set had very low coefficients of determination (ranging from 0.37 to 0.39). For Networks that were trained using pre-processed data (the range of branch diameter and the average branch diameter) as input, coefficients of determination for all the networks trained, except one, were improved. The best R² obtained was 0.52. This R² value was not satisfactory either. The usefulness of the ANN approach for modelling branch diameter distribution patterns in the context of this study proved to be poor. Sensitivity analysis revealed that the most important inputs, in the order of importance, were the largest branch diameter, the average branch diameter, Logbot (the height at the bottom of a log), Top (the height at the top of a whorl), Bot (the height at the bottom of a whorl), and Sedtop (the diameter of the top of the log). These inputs accounted for more than 90% of the total effect of the model. Changing the PE (processing element) number in the hidden layer, and the learning rates did little to improve the performance of the trained networks. Networks trained using the pre-processed artificial data had greater accuracy in predicting the largest branch diameter in the bottom half whorl (R² ranged from 0.72 to 0.74). This was to be expected, for the artificial data set was generated using an existing model. Sensitivity analysis showed that the largest, average, and smallest branch diameters explained about 95% of the model's total effect. The multiple regression approach to predict the largest branch diameter in the bottom half whorl produced much worse results than the ANN’s based on either the pre-processed real or artificial data set. However, consistent with the ANN work, the regression model developed using the artificial data set was able to account for a greater percentage of the variation in the dependent variable. Furthermore, the largest branch diameter in the top half whorl was the most important in predicting the largest branch diameter in the bottom half whorl. Suggestions were made to improve the utility of ANN for modelling branch diameter distribution by changing the way data is collected and by using multiple artificial intelligence tools