Performance of modeling for classification of forest sites in databases with outliers

Abstract

The information used to estimate the productive capacity of forest sites comes from forest inventory databases that may contain discrepant observations (outliers). Thus, consistency analysis is required to exclude these. However, the outliers may represent a certain growth pattern existing in the forest, so their exclusion may be a mistaken action. The objective was to compare the performance of different modeling techniques for forest site classification, considering a database with the presence of outliers. We used pairs of data of age and dominant height (HD) of permanent parcels of Eucalyptus urophila x Eucalyptus grandis located in the north of Minas Gerais. A HD outlier was simulated. The database was modeled, with and without the presence of outliers, by linear regression (RL) and artificial neural networks Multilayer Perceptron (MLP) and Radial Basis Function (RBF). The methods were analyzed by means of precision statistical criteria: bias, square root of mean error, Pearson correlation, mean percentage error and residual scatter plot. The MLP was superior for site index estimation. Therefore, the MLP is indicated for forest site classification when there are outliers in the database.