Avaliação da eficácia de controle de plantas daninhas por glyphosate com uso de análise computacional de imagens por Random Forest | Evaluation of the weed control efficacy with glyphosate using computational image analysis with Random Forest

One of the biggest challenges in agricultural production is the management of weeds, which can harm crop productivity and viability. Different control methods can be used in weed management, with chemical control using herbicides being the most widespread of them. After applying the herbicides, an analysis of the efficacy of their action is carried out. This analysis is carried out based on an on-site evaluation by a trained person, which can be a time-consuming, inefficient, and imprecise process. To confirm the control the method of collecting the remaining plant mass after applying the herbicide is also used, separating it into living and dead material, which makes it possible to calculate the control. This study aimed to develop models generated by Random Forest capable of evaluating the efficacy of weed control by the herbicide glyphosate based on images obtained by smartphone and drones (at four different heights). The images were obtained in an area with uniform infestation of U. brizantha desiccated by glyphosate doses of 0, 144, 288, 576, 864, 1152 and 1440 g a.e. ha-1. Twenty-seven vegetation indices were used as input variables for the model, in order to estimate control based on data from human visual evaluation and control determined by plant mass killed by the herbicide. The proposed models were evaluated using the Pearson’s coefficient of determination (r2). The first article considered the images obtained by smartphone, with the developed models presenting r2 greater than 90% for the training and validation samples, being able to predict with high accuracy the level of weed control by glyphosate. The models were robust against overfitting, showing their potential for use in new data sets. The second article used the drone images, and the developed models also presented high accuracy in predictions (r2 > 90%), for all training and validation samples, regardless of flight height. The model developed to evaluate control independently of height also achieved high accuracy and high performance in the validation samples (r2 greater than 90%), allowing the use of images that vary in flight height and/or image quality (within the range of heights tested). The use of common cameras was sufficient to achieve the study objectives, which means greater economy and accessibility of the process compared to the use of special cameras. The study presents a new alternative for evaluating weed control, with the possibility of implementing these models in mobile and desktop applications.

CAPES - Coordenação de Aperfeiçoamento de Pessoal de Nível Superior