Comparative Analysis of Pre-trained Models and Custom CNN for Weed Seedlings Classification

In the agricultural field, there is a need for effective weed control as it causes high yield loss, especially in large-scale industries. Therefore, we need to identify each species of weed and treat them with weedicides or use other weed control methods to destroy weeds for precision farming. This paper presents a comparative analysis of different approaches for classifying weed seedlings from images. The study focuses on evaluating the performance of a pre-trained Xception model, developed under freezing and fine-tuned scenarios, and a custom Convolutional Neural Network (CNN) model. The dataset comprises 3104 images belonging to six weed classes. Experimental results demonstrate the effectiveness of transfer learning in improving classification accuracy compared to both the pre-trained and custom models. Average accuracy, precision, recall, AUC, and F1-score together with the confusion matrix are used as performance metrics. After training and validating with respective data sets, the networks, pre-trained model architectures both frozen and fine-tuned and the custom CNN architecture were able to achieve classification accuracies of 74.23%, 98.37%, and 90.23% respectively.