Agricultural Pest Classification Using Transfer Learning: A Process Control and Monitoring Perspective ^†

Context: Insect pests are insects that destroy or harm crop plants by cutting their roots, stems, and leaves, and they also absorb the plant’s cell sap from its diverse parts, impacting the crop’s basic health as well as its productivity.

Objective: The main objective of this article is to develop a robust model that is capable of accurately classifying and identifying different types of insect pests from images. The ultimate goal of this article is to help and aid in pest detection and management. In this article, we not only classify the pests through transfer learning but also manage and control the agricultural processes to minimize pest infestation. The process control includes crop rotation, integrated pest management (IPM), etc. During monitoring, we also estimate the severity level of the pests using different sensors.

Dataset description: In this article, we have used the dataset from Kaggle, which consists of 12 folders of different types of insect pests. They are ant—Formicidae (498), bee—Apis (500), beetle—Coleoptera (421), caterpillar—Lepidoptera (454), earthworm—Lumbricina (323), earwig—Dermaptera (466), grasshopper—Caelifera (484), moth—Lepidoptera (496), slug—Gastropods (389), snail—Gastropoda (499), wasp—Hymenoptera-Apocrita (997), and weevil—Curculionoidea (487).

Methods and materials: The given image classifier model is trained using different convoluted neural network architectures such as inceptionV3, vgg 16, and Resnet50, where the accuracy of the inceptionV3 transfer learning model is greater than that of the other two transfer learning models. In the trained model, the dataset consists of 5494 images belonging to 12 different classes.

Results: Here, we have not used the first and last layer of the inceptionV3 architecture, so we are flattening the downloaded inceptionV3 architecture. Dense is used to customize the number of output nodes. We have multiple categories, so we are keeping the activation as softmax and the trained dataset is rescaled. While compiling the model, the optimizer is adam, and the accuracy is 93% with a value loss of 57% and a value accuracy of 97%.