Choosing Classifier for Weed Identification in Sugarcane Fields Through Images Taken by UAV

1 University of Campinas, School of Agricultural Engineering, Campinas – SP, 13083-875, SP, Brazil
2 Embrapa Agriculture Informatics, Campinas – SP, 13083-886, SP, Brazil
3 Federal University of the Jequitinhonha and Mucuri Valleys,Unai Cachoeira CEP 38610-000, MG, Brazil
4 Pontifi cal Catholic University, Electrical Engineering, Faculty, Campinas – SP, 13086-900, Brazil


Yano, I. H., W. E. Santiago, J. R. Alves, L. T. M. Mota and B. Teruel, 2017. Choosing classifi er for weed identifi cation in sugarcane fi elds through images taken by UAV. Bulg. J. Agric. Sci., 23 (3): 491–497

Sugarcane is the main raw material in the world production of sugar and ethanol. The weeds can cause 90% loss in sugarcane production. Thus the weed control is very important and usually made by herbicides application. The estimation of herbicide type and its dosage is in general done by sampling because sugarcane occupies extensive areas. This procedure causes problems of misapplication of herbicide, since the weed species and the level of its infestations could not be uniform in whole fi eld. There are some solutions based on remote sense, using satellite image analysis which covers the whole fi eld, that could solve the problems of the applications of herbicides by sampling, but this solution have problems with image resolution, and can only be used on high weed infestation and in the absence of clouds for good results. This work proposed and tested a process for weed surveying, based on pattern recognition in images taken by an UAV (Unmanned Aerial Vehicle). The UAV can take images very close to the plants, so the plants pattern recognition can be done in lower infestation levels than in images taken by satellites and also is not affected by the presence of clouds. In preliminary testes, three classifi ers were tested; the best classifi er was an Artifi cial Neural Network, which achieved an overall accuracy of 91.67% and a kappa coeffi cient of 0.8958.

Key words: images; machine learning; pattern recognition; sugarcane; UAV; weed

See the article as a PDF.