Deep learning based identification of Crocidolomia pavonana larvae on mustard plants using Grad-CAM

Abstract

Mustard greens are an important vegetable commodity, but their production is often affected by pest attacks, especially the cabbage worm Crocidolomia pavonana (C. pavonana). The larvae damage leaf tissues and cause significant yield losses, while chemical control is often ineffective due to differences in insecticide sensitivity across larval instars. This study proposes a deep learning based classification approach combined with gradient weighted class activation mapping (Grad-CAM) to identify larval instars of C. pavonana on mustard plants. A dataset of 684 images covering instars 1 to 4 was collected from laboratory rearing and field observations, then processed using resizing and augmentation techniques and divided into training, validation, and testing sets with an 8 to 1 to 1 ratio. Two convolutional neural network (CNN) models, visual geometry group 19 (VGG19), and Xception, were implemented with additional fully connected layers. The VGG19 model achieved 94.20% accuracy and outperformed Xception. Grad-CAM successfully highlighted larval regions and supported visual interpretation. The results show that the proposed method can improve pest identification accuracy and support more effective pest management.