ARTIFICIAL NEURAL NETWORKS FOR MODELING HYPSOMETRIC RELATIONSHIPS OF Pinus caribaea Morelet var. caribaea Barr. & Golf.

Abstract

The present study was carried out to compare the performances of regression models and Artificial Neural  Networks (ANNs) in hypsometric relationships modeling and to analyze the influence of ANN type  and sample size on ANN performance. The database was consisted by 65 circular plots of 500 m² in which  Diameter at Breast Height - DBH (cm) and Total Height - Ht (m) of 2538 trees were measured in plantations of Pinus caribaea var. caribaea in Macurije forest company, Cuba. The study was carried out in three  stages: i) Fit of traditional hypsometric models and sigmoidal growth models; ii) ANNs training and comparison of the selected ANN with the regression model selected; iii) Analysis of sample size and ANN type influences on the estimates precision by means of a completely random experimental design with 5x2 factorial arrangement, with the factors sample size (N) and ANN type (R). The results indicated that the best equation to estimate trees heights was that of Gompertz. The ANNs MLP 1-4-1 and MLP 8-4-1 were superior to the selected equation (Gompertz). Multi-Layer Perceptron ANNs generated more accurate estimates and their performances were less influenced by the sample size.