Data mining applied to feature selection methods for aboveground carbon stock modelling

Abstract

The objective of this work was to apply the random forest (RF) algorithm to the modelling of the aboveground carbon (AGC) stock of a tropical forest by testing three feature selection procedures – recursive removal and the uniobjective and multiobjective genetic algorithms (GAs). The used database covered 1,007 plots sampled in the Rio Grande watershed, in the state of Minas Gerais state, Brazil, and 114 environmental variables (climatic, edaphic, geographic, terrain, and spectral). The best feature selection strategy – RF with multiobjective GA – reaches the minor root-square error of 17.75 Mg ha-1 with only four spectral variables – normalized difference moisture index, normalized burn ratio 2 correlation texture, treecover, and latent heat flux –, which represents a reduction of 96.5% in the size of the database. Feature selection strategies assist in obtaining a better RF performance, by improving the accuracy and reducing the volume of the data. Although the recursive removal and multiobjective GA showed a similar performance as feature selection strategies, the latter presents the smallest subset of variables, with the highest accuracy. The findings of this study highlight the importance of using near infrared, short wavelengths, and derived vegetation indices for the remote-sense-based estimation of AGC. The MODIS products show a significant relationship with the AGC stock and should be further explored by the scientific community for the modelling of this stock.