Floresta de precisão na identificação de espécies florestais exóticas invasoras

Abstract

Invasive Exotic Forest Species (IEFE) have the potential to transform the structure and formation of ecosystems, due to their ability to exclude native species and destroy characteristics that local biodiversity provides. Given the concern that IEFE bring, their identification and monitoring are necessary, however, such tasks in protected areas are complex, due to the large territorial dimension, difficulty of access and high costs to inventory these species in the field. In this way, remote sensing through aerial platforms, such as RPAS (Remotely Piloted Aircraft Systems), stands out as an important approach to identify and monitor IEFE, since they allow fast and frequent mapping together with the machine learning (ML) techniques for incorporating specialized knowledge into processing. Therefore, the present research aimed to evaluate the combination of the use of images obtained by RPAS and machine learning algorithms to identify invasive exotic forest species in the Quarta Colônia State Park, Rio Grande do Sul, Brazil. Field data were obtained in two different sampling areas, where the IEFE Hovenia dulcis and Psidium guajava census was carried out, measuring the variables CAP ≥ 5 cm, height, and geographic coordinates collected with GPS. To obtain the remote data, an RPAS coupled to a Parrot Sequoia multispectral camera was used. Subsequently, the images were processed using the Pix4D® application to generate reflectance maps, and the ArcMap® 10.8 Geographic Information System (GIS) to generate vegetation indices and also the spatial distribution of the inventoried species, which was made from manual photointerpretation. In sequence, for the training process of the models, in the GIS ArcGis Pro® 2.8, four classes of interest were defined, being for sample area I (H. dulcis) and sample area II (P. guajava) and the classes similar species, shade and other species applied to both areas. The supervised classification process involved two approaches (pixel-by-pixel and object-based analysis – OBIA) and two ML algorithms in comparison (Random Forest – RF and Support Vector Machine – SVM). The samples were separated into 90% for training/testing and 10% for model validation. For performance analysis, overall accuracy and Kappa index metrics were calculated. The results demonstrate that the RF algorithm in the pixel-by-pixel approach had the best performance in classifying the IEFE H. dulcis, obtaining a kappa of 0.87 and an overall accuracy of 91.5%. For IEFE P. guajava, with the composition of RGB images, the best result was obtained using the OBIA method and the RF algorithm (Kappa of 0.89 and overall accuracy of 92.5%). The RGCV multispectral composition proved to be excellent in differentiating the IEFE P. guajava pattern in the OBIA approach with the RF algorithm (Kappa 0.90 and overall accuracy 93%). In view of the results obtained, the present study has a new and important methodology for identifying the IEFE Hovenia dulcis and Psidium guajava in areas of Seasonal Deciduous Forest, since it can be used in management strategies suitable for the control and eradication of these species.