Imagens multiespectrais e inteligência artificial para predição da densidade de plantas espontâneas em plantio de Eucalyptus saligna

Abstract

The management of Eucalyptus production has its technical operational processes well defined and consolidated throughout the country. However, the management of weeds, which compete with Eucalyptus plants, decrease the final productivity of the plantation, this monitoring of weed control is still dependent on a technical inspection in loco and its quantification is not accurate. Therefore, the present study aims to map the density of weeds in commercial plantations of Eucalyptus saligna through artificial intelligence techniques applied to multispectral images of very high spatial resolution. Thus, a study was developed based on a bibliometric review on the state of the art of the research developed with RPAS (Remotely Piloted Aircraft System) for the mapping weeds in forest and agricultural areas. In four Eucalyptus saligna production areas in the state of Rio Grande do Sul, Brazil, with an average age of 54 days after planting, eight sample plots were evaluated to identify and obtain hyperspectral reflectances readings of weeds and Eucalyptus saligna with the FieldSpec® 3 spectroradiometer. Using the artificial intelligence RF (Random Forest) algorithm with an accuracy of 95.44%, it was determined that the most important wavelength ranges are from 510 to 589 nm, 400 to 423 nm, 674 to 731 nm and 886 at 900 nm were able to distinguish weeds from Eucalyptus saligna individuals in commercial plantations. In these same areas, multispectral images were also obtained with the Parrot Sequoia sensor embedded in the RPAS Phantom 4 Pro, using a flight height of 30 m. From these images, the four sensor bands and five more vegetation indices were used as predictors. The K-Means algorithm was applied for image segmentation and vegetation discrimination in the classes Eucalyptus saligna, weeds and regrowth of Eucalyptus saligna. These data were partitioned into 70% training and 30% testing, to be modeled by the RF algorithm, whose model obtained an accuracy of 95.49% in the classification of weeds, which enabled the elaboration of the weed density map for the study areas, composing the final product of the study