Automatic Classification of Bagworm, Metisa plana (Walker) Instar Stages Using a Transfer Learning-Based Framework

Abstract

Bagworms, particularly Metisa plana Walker (Lepidoptera: Psychidae), are one of the most destructive leaf-eating pests, especially in oil palm plantations, causing severe defoliation which reduces yield. Due to the delayed control of the bagworm population, it was discovered to be the most widespread oil palm pest in Peninsular Malaysia. Identification and classification of bagworm instar stages are critical for determining the current outbreak and taking appropriate control measures in the infested area. Therefore, this work proposes an automatic classification of bagworm larval instar stage starting from the second (S2) to the fifth (S5) instar stage using a transfer learning-based framework. Five different deep CNN architectures were used i.e., VGG16, ResNet50, ResNet152, DenseNet121 and DenseNet201 to categorize the larval instar stages. All the models were fine-tuned using two different optimizers, i.e., stochastic gradient descent (SGD) with momentum and adaptive moment estimation (Adam). Among the five models used, the DenseNet121 model, which used SGD with momentum (0.9) had the best classification accuracy of 96.18% with a testing time of 0.048 s per sample. Besides, all the instar stages from S2 to S5 can be identified with high value accuracy (94.52–97.57%), precision (89.71–95.87%), sensitivity (87.67–96.65%), specificity (96.51–98.61%) and the F1-score (88.89–96.18%). The presented transfer learning approach yields promising results, demonstrating its ability to classify bagworm instar stages.