Predição da diversidade arbórea no sudeste do Brasil utilizando sensoriamento remoto e aprendizagem de máquina

Abstract

Knowing the biodiversity of a region is an important management tool with regard to sustainable development. For that, indexes are used that synthesize information on the richness and uniformity of species in an area, based on field surveys, which in turn are time-consuming and expensive. Environmental modelling using easily obtainable variables (such as from remote sensing) has helped a lot in this process, reducing the need for such surveys. On the other hand, in view of the large amount of data available for use in modelling, another challenge arises, which is the selection of variables that most contribute to the prediction of the metric in question and the machine learning algorithms are quite efficient in this matter. Therefore, the objective was to model and map tree diversity for the state of Minas Gerais using the Random Forest (RF) algorithm and climatic, terrain and remote sensing variables. We also tried to answer the following questions: (1) Which variables have the greatest predictive potential for each diversity metric? (2) What is the pattern of spatial distribution of tree diversity for the state of Minas Gerais? (3) Is there a consensus among the diversity metrics? Information from the forest inventory of 2755 sample plots was used to obtain diversity metrics and 115 predictor variables. The modelling was performed using the Rstudio software. The models generated by the RF for all metrics had an average R² of 0.46 and RMSE (%) of 20,73. In general, the most representative selected variables for the metrics were precipitation (climatic), gross primary productivity (remote sensing) and valley depth (terrain). The maps generated for the state show that the tree diversity increases from North to South and Northwest to Southeast, being strongly associated with the morphoclimatic domains. The greatest tree diversity was concentrated in the Atlantic domain, decreasing slightly in the Cerrado domain and the least diversity was in the Caatinga domain. Fisher's alpha index had a greater consensus with richness, and the Shannon and Simpson indexes agreed more with each other. Pielou, for representing uniformity, presented a different pattern. It is concluded that the variables selected by the RF managed to capture the existing differences between the domains and represent well the spatial patterns of tree diversity in Minas Gerais.