Modelagem espacial de indicadores ecológicos em áreas atingidas por rejeitos de minério de ferro

Abstract

Environmental degradation resulting from human activities represents one of the most urgent and complex challenges faced by contemporary society. An example of this scenario is the disaster that occurred in Mariana, Minas Gerais, in November 2015, when the collapse of an iron ore tailings dam resulted in catastrophic consequences for the local environment. Since then, efforts have been directed towards the restoration of these affected areas, which must be monitored to verify the trajectory of ecological processes and evolution in the new ecosystem, as they provide data on errors and successes, allowing readjustment of the methods used. Given this scenario, the main objective of this research is to model and map ecological indicators of diversity (species richness, Model I) and structure (canopy cover, Model II) on the banks of the Gualaxo do Norte River, in Mariana, MG. Two scenarios were evaluated (Scenario A and Scenario B) to understand the influence of field observation variables on the models (Scenario A, generating Models IA and IIA), as well as the exclusive use of remote sensing variables (Scenario B, generating Models IB and IIB). The indicators were collected in 24 transects in six different areas. In addition to the indicators, field observation data (categorical predictor variables) were collected. Remote sensing (RS) data was generated from Planet imagery and used in the modeling. To carry out the modeling and mapping of indicators, the Random Forest (RF) machine learning algorithm was used, generating preliminary models and, after selecting the variables, generating the final models. The models' performance was assessed using the Coefficient of Determination (R²) and error metrics (Root Mean Square Error – RMSE and Mean Absolute Error - MAE). The results show that to estimate species richness, the use of categorical field data is essential for mitigating the effect of data saturation, in addition to improving the global fit and precision of the model. In contrast, the canopy cover model was better fitted using only the remote sensing variables. The variables that contributed most to the spatial estimate of species richness (Model IB) and canopy cover (Model IIB) were, respectively, near-infrared textural mean and NDVI vegetation index.