Identification of weeds in sugarcane through image processing

Abstract

The increase in production without causing damage to the environment is one of the biggest challenges of modern agriculture. In the production of sugarcane, this becomes clearer when it comes to the weed management, since the use of herbicides to configure most widely adopted technique. Weeds cause interference in agricultural production, causing reduction in product quality and crop yields. Therefore, the identification of weed species and the level of infestation becomes very important so that appropriate management strategies can be defined. This study sought to develop and evaluate the performance of an image processing system to identify weeds in sugarcane and estimate their level of infestation, since the existence of a computer tool to recognize plants species should give a great support to decision-making about the management of weed communities. The approach taken to identify weeds and crop plants was based on the methodology of bag-of-words. On this methodology, invariant feature points and multiple images are used to create a dictionary of features, the dictionary is then used to ascertain what his words are present the images to be processed, the quantization of the number of words in the dictionary is present in the image made by a probability density function and the mathematical model of rank was made by support vector machine. Considering the performance measures: overall accuracy and Kappa coefficient, the developed system has processed 435 RGB images, what were obtained from three experimental cultives having plants of sugarcane, corn and six weed species (Urochloa plantaginea, Urochloa decumbens, Panicum maximum, Euphorbia heterophylla, Ipomoea hederifolia and Ipomoea quamoclit). The results show that the method has high power to identify and discriminate weed and crop, reaching overall accuracy and Kappa coefficient of up to 94% and 0.94, respectively. These results give support to premise that an image processing system is capable to identify weeds in sugarcane, estimate the infestation level and to be yet a tool for support the decision-making about the management from the weed species.