Algoritmos híbridos aplicados a biometria florestal

Abstract

Understanding and quantifying the growth of trees are very important for providing available information, in order to enable an adequate management of forest resources. Thus, modeling has becoming an important tool, because it has the ability to synthesize knowledge, identify gaps, design and predict behavior according to different conditions. In this context, many techniques are used, highlighting computational intelligence with great growth in the forest sector and good results. As an example, we have the Random Forest, Simulated Annealing and Genetic Algorithm. In this dissertation, these algorithms were applied in different situations for native and planted forests. The first article had as main objective to evaluate the efficiency of the genetic algorithm and Simulated Annealing in the prediction of parameters of the Weibull function of 2 parameters. From the results it was concluded that the tested methods are consistent and stable, surpassing the classic methods regardless of silvicultural regimes involving thinning evaluated. The second article approach the selection of variables and modeling of the above-ground carbon stock, at tree level, of native vegetation, through the construction of a hybrid method involving the Simulated Annealing and Random forest. The results obtained indicated that the hybrid method was efficient, since it managed to reduce the number of variables (less than 16 variable) and identify those that contribute most to explain the carbon stock, in addition to reducing the error of the estimates. In general, the ability to apply the machine learning technique was observed in the cases evaluated.