Weed Identification using Convolutional Neural Network and Convolutional Neural Network Architectures

Abstract

In order to overcome this threat imposed by weeds in agriculture, a measure is taken to identify the weeds that grow along with the seedlings with the help of deep learning (DL) technique. Convolutional neural network (CNN), a class of DL render a good way to identify the weeds that harm the plant's growth. Aiming at achieving a greater accuracy, the models such as four convolution layered, six convolution layered, eight convolution layered and thirteen convolution layered architecture were built. Comparatively, eight convolution layered architecture resulted with 97.83% as training accuracy and 96.53% of validation accuracy than the VGG-16 model resulted with. The use of CNN architectures paved way to reach training accuracy of 96.27% and validation accuracy with 91.67% in ZFNet and 97.63% as training accuracy and 92.62% of validation accuracy in ALEXNET. Therefore, by the use of this technology and suggested method there is a lot of possibilities to avoid the manual field work of identifying the weeds. Our results suggest that more of datasets can be used and fine-tuning of parameters can be done.