Modeling the leaf area of Ormosia paraensis Ducke by statistical models and artificial neural networks

Abstract

The leaf is a very important plant structure because it allows gas exchange and transformation of light energy into chemical energy. The objective of this research study was to test artificial neural networks (ANNs) to estimate the leaf area (LA) of Ormosia paraensis Ducke and compare their performance with adjusted statistical models. One hundred forty leaves were selected from seedlings of the species at the leaf age of 100 days after sowing (DAT). The LA was calculated by indirect estimation with ImageJ software, and the linear length (L) and width (W) dimensions were measured with a ruler graduated in centimeters. Afterward, 90 leaves were randomly separated to generate mathematical equations of LA (Y) as a function of linear dimensions (L and W), and 50 leaves to validate the equations. Similarly, 100 networks of the multilayer perceptron (MLP) type were trained with the backpropagation algorithm and the best networks were selected for later validation. The choice of the best equations and ANNs was based on precision and dispersion statistics. The quadratic equation obtained from model (4) demonstrated consistent statistics with the coefficient of determination (R²aj = 87.39) and a low result for the standard error of estimate (SEE = 12.07%). For the ANNs, a correlation coefficient (Ryŷ), varying between 0.9316 and 0.9521, was obtained in the training phase, while it ranged from 0.8522 to 0.8825 in the validation phase, generating lower residues. It is concluded that the ANN performance was higher compared with the conventional regression technique.