Genetic Algorithm-Based Hyperparameter Optimization for Convolutional Neural Networks in the Classification of Crop Pests

Crop pest classification is essential for a strong and sustainable agricultural economy and food safety. However, the classification of pests is a time-consuming process that requires domain knowledge and relies on expertise. Therefore, automation of the classification process can reduce cost, improve accuracy, and facilitate analysis. In recent years convolutional neural networks (CNNs) and transfer learning fine-tuning methods have gained popularity in solving many computer vision problems in agriculture. The main advantage of pre-trained CNN models is instead of designing and training a model from scratch to solve various classification problems utilizing pre-trained models via transfer learning fine-tuning methods. However, it is important to determine transfer learning and fine-tuning hyperparameters of a pre-trained CNN model to achieve a successful classification performance. But this is a challenging task that requires experience, knowledge, and a lot of effort. This study proposed a new genetic algorithm-based hyperparameter optimization strategy for pre-trained CNN models in insect pest type classification. The proposed method was tested with three CNN models at different scales (MobileNetV2, DenseNet121, and InceptionResNetV2) on three insect datasets; Deng's dataset with 10 classes, Xie2's dataset named D0 with 40 classes, and Wu's dataset named IP102 with 102 classes. The optimized CNN models have achieved state-of-the-art accuracies on D0 (99.89%) and Deng (97.58%) datasets and showed the closest performance to the literature on the IP102 (71.84%) dataset. According to the test results, the proposed method effectively classifies various crop pests and can be used in farming to save crop fields.