Evaluation and selection of predictive variables in the estimation of wood density of Eucalyptus
Name: ISÁIRA LEITE E LOPES
Publication date: 28/02/2018
Examining board:
Name |
Role |
|---|---|
| GILSON FERNANDES DA SILVA | Internal Examiner * |
| GRAZIELA BAPTISTA VIDAURRE | Co advisor * |
Summary: The objective of this work was to evaluate and select the most relevant predictor variables for estimating the basic density of eucalyptus wood. The qualitative variables obtained from cadastral data (clone, sub-region and relief), the quantitative variables obtained from the Continuous Forest Inventory - IFC (total volume with bark, diameter at breast height and total height) and quantitative data from the climatic information of the study area (wind speed, mean temperature, mean total precipitation, vapor pressure deficit, water deficit and altitude) were used to estimate the wood density of 386 trees. The methods of evaluation and selection of variables used were: brute force with application of Artificial Neural Networks (RNA) testing all possible combinations between variables; algorithm of Garson and Random Forest, that quantify the individual importance of the predictor variables. The classification of the predictor variables varied among the methods, which can be attributed to their different mathematical approaches. The clone variable stood out from the others, in all methods. For the brute force method, the simplification of the artificial neural network with the use of 5 variables resulted in a higher degree of accuracy of the basic density estimates, WHERE the optimal combination consisted of clone, age, total volume with bark, mean temperature and water deficit. As for the Garson algorithm, the 5 variables with the highest importance were: clone, subregion, relief, age and water deficit. Random Forest presented, among the 5 most important variables, clone, age, total height, mean total precipitation and mean temperature. However, in the face of computational effort to apply the brute force method, an alternative is the use of Random forest or Garson algorithm, since the variables selected in these methods also provided good estimates of the basic density of wood.
Keywords: Random forest, Garson algorithm, Artificial Neural Networks, wood, forest measurement.
