Analysis of various crown variables on stem form for Cunninghamia lanceolata based on ANN and taper function

Abstract

Crown, with many dimensions greatly influences the stem taper of a tree. However, few taper models have accounted for its impact on diameter estimation. In order to investigate and quantify the effects of various crown factors on stem taper and develop new taper models incorporating crown information for Cunninghamia lanceolata in Southeast China, a sample data of 1100 taper measurements from 108 trees and two different modeling methods were utilized. A set of traditional non-linear regression (NLR) models with linear and non-linear functions composed of crown factors introduced respectively, were developed for stem diameter prediction, as well as artificial neural network (ANN) models based on different input variables. ANN technology was applied to variable screening prior to developing models, and the evaluation statistics and graphics were used to assess the models. The results showed that crown length (CL) and height to live crown base (HCB) had larger influence on the model accuracy than other crown variables, demonstrating that variables screening based on ANN is feasible and efficient when numerous potential variables are available. The accuracy of taper model was improved when incorporating different crown variables or their combinations in terms of higher R2 and lower RMSE, however, the degree of improvement varied, depending on the variables added and the modelling approach. The inclusion of CR and HCB presented the highest improvement, whether using ANN or NLR modelling methods. The ANN model decreased 13.96 % in RMSE and 17.00 % in MAE. Similarly, the NLR model reduced 2.1 % in RMSE and 1.78 % in MAE. This study indicated that the refined models with crown variables included were more in line with the biological logic of nature and the accuracy has been improved, although the improvement of nonlinear regression models was not as significant as expected. In addition, it also suggested that in forest resource inventory, ANN was a recommended technique for variable screening and an alternative method for model development.