Desenvolvimento de algoritmo baseado em redes neurais convolucionais para detecção de ervas daninhas em áreas de cana-de-açúcar

Abstract

Weed infestation is an important factor in productivity loss for sugarcane cultivation. The use of remote sensing data, whether coming from satellites or from Remotely Pi- loted Aircraft (RPA), in conjunction with Artificial Intelligence (AI) techniques, can lead the cultivation of sugarcane to a new level in terms of refers to weed control. For this purpose, an algorithm based on Convolutional Neural Networks (CNN) was developed for detection, quantification and mapping of weeds in sugarcane areas in the state of Alagoas. Data from images of the PlanetScope satellite were subdivided, separated, trained in different scenarios, Classified and georeferenced, being finally turned into a map. Scenario one of the CNN training and test presented the best ac-curacy result (98.3%), being selected to generate the geometric shape from the clas-sify algorithm, which was then georeferenced and transformed into a map of the sugarcane, weed and forest classes. The comparison of the accuracy results showed that the values of the results obtained by the CNN algorithm were higher than the Kappa coefficient val- ues. The quantitative analysis of the area (ha) infested by weed indicated a high prob-ability of a negative impact on sugarcane productivity. The orthomosaic of RPA images and the collaboration of the Plant technicians were essential to prove the occurrence of weed infestation, as well as the recognition of the species (Momordica charantia L.) predominant on the farm. It is recommended that, in the future, the adequacy of CNN's algorithm for RPA images be carried out, aiming at the differentiation between weed species, as well as its application in the detection of weeds in areas with other cultivars of great relevance to the sectors agricultural and economic state of Alagoas.

Key-words: CNN. PlanetScope image. Python. Precision Agriculture.