Redes neurais artificiais aplicadas na modelagem de relações dendrométricas e morfométricas de espécies arbóreas do Cerrado

Abstract

The Cerrado has a high biodiversity of plants. However, few studies detail the Cerrado’s dendrometric and morphometric characteristics, which allow to understand the competition and vitality of trees and can assist the monitoring and execution of conservation practices. An alternative technique to describe these relationships is the use of Artificial Neural Networks (ANNs). Thus, the present study aimed to: i) identify and characterize the sampled trees; ii) train ANNs to predict the dendrometric and morfometric relationship; iii) evaluate the performance of ANNs using qualitative and quantitative variables iii) create an algorithm in R language to calculate the dendrometric and morphometric variables. A total of 200 open grown trees were sampled in an area of Cerrado, where trunk diameter measurements were made at different levels above the ground and the following qualitative variables were defined: branch arrangement, geometric shape of the crown, stoniness, tree position on the relief and vitality. In cases of occurrence of more than one stem, the equivalent diameter (deq) was used. In addition, the total height (h), crown base height (cbh), crown diameter (dc) were measured and the morphometric variables were calculated: crown proportion (CP%), slenderness rate (SR), range index (RI), salience index (SI) and crown formal (CF). ANNs were trained with different configurations in the input layer: using only the deq and then the deq and the geometric shape of the crown. The resilient propagation - (RP-) learning algorithm with Multilayer Perceptron (MLP) architecture was used. The sampled trees belonged to different botanical families. ANNs using only deq presented a low performance in the description of the variables: CF, SI and RI. The insertion of the categorical variable crown shape provided the ANN with greater efficiency in the predictions of all variables in training. Synaptic weights and the algorithm created in R can be used to predict the variables of the present study.