Uso de vants no monitoramento da sigatoka-amarela da bananeira

Abstract

The Machine Learning techniques in precision agriculture offer new prospects for the monitoring and identification of phenotypic characteristics in crops,such as physiological, biotic and abiotic features, as a manifestation of diseases and pests, hydric and nutritional stress. The algorithms used in the processes seek the recognition of standards from Remote Sensing data, such as aerial imagesobtained through Unmanned Aerial Vehicles -UAVs. This scenario includes banana farming, an activity of great economic importance, being one of the most consumed fruits in the world, with big nutritional value. The banana cropis affected by severaldiseases and among them, yellow sigatoka, which is one of the main limiting factors to its cultivation, causing considerable losses in fruit production. In this context, searching for the basic assumptions for identification, classification, quantificationand prediction (ICQP) of phenotypic factors, the general objective of this work was to use remote sensing techniques, machine learning and high spatial resolution aerial images to monitor the severity of the yellow sigatoka attack in banana culture. Monthly flights were carried out in banana plantations in the city of Russas, Ceará, Brazil, belonging to the company Frutacor, using UAVInspire 1, shipped with X5 (panchromatic RGB) camera of 16 megapixels and 8 bytes. The algorithms Maximum Likelihood, Mahalanobis Distance and Minimum Distance, were considered as easy interface and fast processing, using the PhotoScan software. The algorithms were evaluated by the Kappa statistic and the Global Accuracy Index and the data obtained by the tests, compared to the field surveys. As a result, the Minimum Distance algorithm achieved better performance (99.28% accuracy) for the month of September 2017, and 2.44% of the degree of severity of the yellow sigatoka, compared to the field survey, which resulted in a degreeof infection from 1% to 5%. For the months of October and November, the Maximum Likelihood algorithm obtained 89.77% and 78.76% of accuracy, approaching the values collected in the field, demonstrating that the tools for monitoringleaf spots can be performed by means of techniques remote sensing, computational learning, and high spatial resolution panchromatic images.