Performance Comparison of Different Convolutional Neural Network Architectures for Plant Seedling Classification
Nawmee Razia Rahman, Md. Al Mehedi Hasan, Jungpil Shin
Abstract
Agriculture is the backbone of many nations. Many developing and under-developed countries' economies still depend on agriculture. The main target is to increase production while reducing cost, supply food to the increased population worldwide. In the agricultural system, weeds are a major concern for farmers. They compete vigorously with the crops for nutrition and water and cause huge loss. The use of chemical herbicides eliminate weeds but do a lot of damage to the environment also increases cost. Regarding the environment and cost an automated machine vision system is required that can identify crops and weeds in a safe and cost-efficient manner. In this work, we have implemented several Convolutional Neural Network architectures to find the best performing architecture that classifies plants and weeds accurately. For this reason, we have used a database that contains a total of 5539 images of 12 classes (3 crops and 9 weeds) at their several early stages. We have calculated the accuracy, precision, recall, and f1-score of each architecture. From this study, we have found that ResNet-50 is the best performing architecture with an overall test set accuracy of 96.21% that is higher than other architectures used in this task. This model can be a useful tool for farmers in identifying weeds at their early stages.