Índices espectrais para gestão do controle químico de plantas daninhas em povoamento de eucalipto

Abstract

One of the main components of eucalyptus production cost is weed management. Due to the growth of the areas and operational schedules of the companies, the management of weed competition control has increased costs and compromised results. Insertion of precision farming tools can be an alternative to managing chemical control. The aim of the project was to determine and relate the spectral indices of vegetation after application of glyphosate and simulations of drift and rainfall in eucalyptus and signalgrass plants. Two field experiments were set up at the Federal University of Uberlândia, on the campus of Monte Carmelo, MG: in the first, eucalyptus seedlings were transplanted in a spacing of 2.5 x 0.8 m and developed up to 225 days. The treatments, delineated in four blocks, were: glyphosate drift simulations in eucalyptus plants corresponding to 15, 10, 5 and 2.5% of the commercial dose (1,240 g ha-1). In the second experiment, in an area occupied by signalgrass plants, glyphosate (2,170 g ha-1) was applied and the treatments were related to the simulation of rainfall (10 mm) after application: 30, 60, 120 and 240 minutes. The experiments were carried out at the same time and images were captured weekly from the plots using a UAV with a 20 megapixel RGB camera coupled to determine the following spectral indices VARI, MPRI, MGVRI, ExG, GLI, RGVBI. BGI, HI, HUE, SI, SCI and BI. Weekly in loco evaluations of the herbicide intoxication symptoms in the plants were also carried out. The data generated after calculating the spectral indices, with the generation of orthomosaic of the captured images, were compared with the data observed in loco by means of Pearson's correlation at 5% significance, and subsequently, the best index was selected for the use of it with the aid of neural networks, selecting the best learning method, namely: GBM (Gradient Boosting Algorithm), Random Forest and Deep Learning (DL). Regarding the detection of intoxication symptoms in eucalyptus plants, no treatment effect was observed. However, with regard to the evolution of symptoms of poisoning in signalgrass plants, a strong correlation was observed between the studied indices and the grades of poisoning caused by the herbicide. It is concluded that both the spectral indices and the use of neurais networks, with the use of Deep Learning, can be used to evaluate the evolution of signalgrass poisoning by glyphosate and for the eucalyptus experiment, it is concluded that the indices spectral determined in images collected by drones in a eucalyptus stand do not correlate with the intoxication data collected in loco.