Deep Transfer Learning based Light-weight Agricultural Pest Classification Model using Convolutional Neural Networks

A variety of pests and insects can hinder the sprouting of major field crops including wheat, rice, maize, sugarcane, and soybeans, as well as lower agricultural productivity. Since all agricultural insects have similar appearances during the initial phases of crop growth, farmers find it challenging to appropriately classify and identify these insects. Deep architecture Convolutional Neural Networks (CNNs) are employed to tackle this issue, as they can automatically extract and comprehend complex features suitable for applications in image classification. This paper seeks to highlight the competence of transfer learning and the implementation of a pre-trained model customized for the Pest Dataset in the classification of pests into 12 categories. The proposed framework was developed by incorporating the convolutional and pooling layers from pre-trained CNN frameworks, namely ResNet50, Xception, DenseNet121, and EfficientNetB5 for feature extraction. Furthermore, new fully-connected layers and an output layer were developed particularly for the categorization process. The Agricultural Pests Image Dataset is used to examine the proposed framework. The experimental findings demonstrate that Xception, when employed with the proposed architecture, obtained the highest accuracy, with a training accuracy of 92.09% and a testing accuracy of 89.35%, surpassing the performance of the other four pre-trained networks.