Simulated annealing in feature selection approach for modeling aboveground carbon stock at the transition between Brazilian Savanna and Atlantic Forest biomes

Forest ecosystems are important in the carbon storage process. Thus, the objective was to investigate the effectiveness of the Simulated Annealing meta-heuristic analysis for selecting variables to maximize the accuracy of the aboveground carbon prediction at the tree level. We used data from uneven-aged forests located in the Rio Grande Basin - Minas Gerais, Brazil, where 227 trees had their carbon stock measured. The classic Spurr linear model, stepwise linear regression and pan-tropical coverage, Random Forest (RF), and the hybrid SARF method (Simulated Annealing and Random Forest) were used to estimate the carbon stock from the selection of variables for the different compartments of the tree (total, stem, branch, and leaf). The SARF consisted of the metaheuristic to select the variables to be used in the RF. These methods were evaluated by the root mean square error (RMSE), coefficient of determination (R²), and residual graph. As a result, the pan-tropical equation demonstrated superior performance than the Spurr model due to its greater homogeneity of residues. The stepwise technique reduced the number of variables and the error of the estimates, mainly for the validation set. SARF showed better adjustments than RF, as it reduced in on average 99.2% of the number of variables and 9% of the error of estimates considering all compartments. In general, variables such as volume, basic wood density, canopy projection area, diameter at 0%, diameter at breast height, height, and latitude contributed strongly to the carbon independent of the tree compartment. Among the methods, SARF is an alternative to the traditional method, as it can extract accurate information from a large data set.