Aprendizado de máquina para a conservação da biodiversidade: adequabilidade de habitat nas unidades de conservação do estado de São Paulo

Abstract

Habitat Suitability Models (HSMs) are statistical models that relate the location of species and environmental variables that restrict their distribution. They can be developed by Machine Learning algorithms, aiming to discover patterns in the data from training and testing in sample points of species occurrence. The general objective of this research is to evaluate Habitat Suitability (HS) in state and federal Conservation Units in São Paulo State based HSMs, as a way of measuring the potential of these protected areas for the maintenance of mammalian species. The hypothesis of this research is that the most restrictive categories of Conservation Units present environmental characteristics more favorable to the occurrence of terrestrial mammals. The models used as predictor variables: Precipitation, Temperature, Artificial Nightlights, Altitude, Terrain Slope and Land Use Land Cover. Data on the occurrence of mammalian species were extracted from GBIF - Global Biodiversity Information Facility. To generate Pseudo-absence points, the OC-SVM classification was performed based on the Presence points; the non-presence area was used as a background to randomly generate the Pseudo-absence points. Then the Presence and Pseudo-Absence points were used for training and testing the Decision Tree, Random Forest, KNN, SVR, XGBoost and LightGBM regression models. Results showed that tree-based Machine Learning methods (Random Forest and Decision Tree) performed better than distance-based methods (KNN) and margin-maximizing methods (SVR). The modeling of ecological processes seems to be better adjusted to the separation of groups with successive divisions than to the attempt to discriminate these groups in an attribute space. The predictor variables that were most important in the modeling were Minimum Temperature, Artificial Nightlights and Precipitation (numerical variables) and Mosaic of Agriculture and Pasture and Forest Formation (categorical variables). In São Paulo State, Environmental Protection Areas (EPAs), as they are a category with less restrictions on land use, have an mean HS similar to areas not protected by Conservation Units. The other Sustainable Use Conservation Units, although they have some degree of anthropism, have a good mean HS for mastofauna; Integral Protection Conservation Units presents the highest mean HS. The models indicate three main areas in São Paulo State with high AdH: 1) Northwest of the state, where the landscapes are strongly marked by agriculture; 2) Northeast of the state, where sugarcane production is dominant, but where there are still remnants of Cerrado not protected by Conservation Units; and 3) geomorphological domain of the Southern Coastal where there is a overlapping of Conservation Units. The occurrence of the species signals but does not define the habitat suitability. Agricultural (and even urban) ecosystems can be used for the transit of individuals or populations of mammals that, eventually, can be observed in these places. However, these ecosystems are not enough to supply all the resources for the maintenance of wild life, in terms of food, safety and thermal comfort, especially in the case of specialist species. The use of agricultural landscapes is due to the high forest fragmentation that suppressed the original habitat of mastofauna in the São Paulo state.