The Convolutional Neural Network approach in a weed detection: a case of the rice crop

Abstract

This paper investigates the efficiency of a Convolutional neural network in detecting weed in case of rice field. Generally speaking, rice is considered as the most crop presents in farming system of Malagasy household. However, yield may be altered as these weeds represents one of the most threats encountered on rice growing. They enter in competition with crop plant for light, nutrients water, and space. Yet, the objective of this paper is to obtain the best Convolutional Neural Network Model by using rice and weed dataset in terms of accuracy and loss in order to make a distinction between weeds and rice crop. The method is based on training each of pretrained model and the baseline one, use the weed and rice dataset, making feature extraction and, at the end, take an unseen dataset and test it on the selected CNN Model. 10 different model has been lined up and among of them, the INCEPTION_V3 obtain the best general score (0.652 for loss and 0.804 for accuracy). In addition, out of 8 samples, 7 has been «correctly classified »and one classified as "not classified". Regarding the performance of the INCEPTION_V3 model on the weed and rice dataset, it did a pretty good Job in the prediction exercise. However, some disadvantages have been spotted like the lack of dataset even when we use a data augmentation technique.