Metaheuristic-based automated design of Convolutional Neural Network architecture for plant seedlings classification

Recently, Convolutional Neural Network (CNN) has been successfully applied in various fields and it is also being applied to the field of agriculture. The use of CNN for automatic weed and plant seedling classification has emerged as an important and promising area of study for enhancing agricultural outcomes. However, convolutional neural networks' performance heavily depends on their architecture. Traditionally, CNN structures have been created through manual trial-and-error procedures, which can be tedious and time-consuming and often require expert knowledge. To address this tedious problem, our paper presents an automatic CNN architecture design approach by using a genetic algorithm (GA), based on pre-trained fine-tuning model as a Backbone. We also present chromosome representation (i.e., an array of parameter values) to encode CNN architectures, where a pre-trained model is considered as a chromosome’s parameter. The suggested method is evaluated using the V2 plant seedling classification dataset, which includes 5539 images of around 960 distinct plants from 12 different species at various growth stages. By leveraging the power of transfer learning technique and the optimization capabilities of GA, our method achieves validation accuracy of up to 97.74%, outperforming state-of-the-art methods, without performing any trial-and-error based architecture design approaches. This method offers a revolutionary tool that is specifically designed for the dataset being used to increase the accuracy of CNN models.