Investigation on Weeds Detection in Crops Using AlexNet Deep Learning Algorithms

Abstract

newline The presence of weed in the agriculture farms reduces 30% to 80% of entire newlineproduction in India. Detection and removal of weeds is essential to improve the production of newlineagriculture yields. In the weed management agrarians are using herbicides or physical man newlinepower presently. In order to bypass the adverse effects of herbicides and huge wastage of man newlinepower, precision or digital agriculture is preferred by the researchers. In the precision newlineagriculture capturing images and processing images to identify the weed from the main crop is newlinethe prime and fundamental process. So, for the weed detection with image processing datasets newlineare needed. The data sets collected from the Mahatma Phule Krishi Vidyapeeth, Rahuri, Shri newlineShivaji agriculture college, Amravati and different field of Vidarbha region, State of newlineMaharashtra INDIA. To process the images in the identification and classification of the newlineweeds AlexNet network is preferred because of automatic feature extraction. So this research newlinefocus on the training AlexNet with the collected datasets. For the training optimizing learning newlinerate important. So that, It is important to find the optimal value of learning rate for proposed newlinealgorithm on the collected data sets using transfer learning approach. The Training of Deep newlineNeural Networks is handled with the parameters of batch size, learning rate and training time. newlineIn this research work, the experimentation of training the AlexNet network is done using newlineMATLAB R2020a tool. The investigation of trained network s training time requires the newlinevariable batch size and learning rate consecutively. In addition to that the behaviour of batch newlinesize on the performance of AlexNet is also analyzed. The higher learning rate needs more newlinetraining time to train the network. keeping learning rate low and small batch size will allow newlinetraining better but it requires more training time additionally. Since the optimal learning rate newlinecannot be find out by analytical calculation trial and error method is preferred. From the trial newlineand error method