Inteligência computacional na modelagem florestal: teor de carbono e distribuição geográfica de espécies

Abstract

Forest modeling, whether for its dendrometric variables or geographical distribution, is a consolidated practice in Forest Engineering in which traditional statistical models are typically employed. However, the progress made by computer science in recent decades has made new challenges and solutions possible, which may limit the use of traditional statistical models. Machine learning algorithms gain their use in the forestry sector within the context of little-known non-linear relations derived from large databases. Among these algorithms, the random forest (RF) is prominent due to its robustness, ease of parameterization, and internal metrics. Despite its great potential, this algorithm demands further studies to consolidate its use. In this dissertation, we applied the algorithm in three different situations for native forests, addressing classification and regression issues, as well as heterogeneous data from different sources. The first article (1) aimed to evaluate three machine learning methods (decision tree - J48, RF, and artificial neural networks) for the potential distribution modeling of the ten most abundant tree species in a subbasin of the Sao Francisco River, in Minas Gerais, Brazil. In conclusion, the RF algorithm presented the most robust tree species potential distribution model. With the results obtained by the algorithm, we wrote the second article (2) seeking to model the potential distribution of Eremanthus erythropappus, considering climatic change scenarios. The hypothesis tested is associated with future effects (2050 and 2070) on the ecological niche of the species. The results indicate a good accuracy of the method used, highlighting the Espinhaço, Canastra, and Mantiqueira mountain ranges as the three primary ecological refuges of the species. We verified a drastic reduction in the potential area of species development regarding its interaction with the local climate if the climate changes scenarios be come real. The last article (3) focused on the application of RF on regression problems involving the prediction of the quantitative variable of carbon content above the soil. The objective was to test the combination of methods and strategies in the selection of variables using the random forest. The results obtained indicate that the RF algorithm is a robust method, little affected by the inclusion of many correlated variables. Even with the slight improvement in algorithm errors, the use of variable selection techniques is justified since it considerably reduces the number of variables used. The multi-objective genetic algorithm obtained a smaller set of selected data and lower error. Given the results found, the RF has great potential to explore relationships still little known in Forest Engineering, whether they are classification or regression issues.