Cutting-edge ensemble framework of deep convolutional neural networks for high-precision insect pest classification

Abstract

In response to the pressing agricultural concern posed by insect pests, leading to substantial crop losses and compounded by the challenges of distinguishing between similar species, this study presents an innovative solution using convolutional neural networks (CNNs) for rapid and accurate insect species recognition, addressing the agricultural challenge of insect pests and species differentiation. Initially, six pre-trained CNN base models (VGG16, VGG19, ResNet50, Inception-V3, Xception, and MobileNet) are fine-tuned and perform classification on our unique dataset from Kerala, India, called KSDAgriPest dataset with 33 insect classes. Later, four best-performing base models, VGG16, Inception-V3, Xception, and MobileNet, were modified and retrained using appropriate transfer learning and fine-tuning strategies and are ensembled via all possible combinations of three base models using genetic algorithm (GA) optimized weighted voting, is called GAEnsemble and the generated models are called Ensemble Variants (EV). In the final stage, two top-performing EVs are ensembled. The proposed “Genetic Algorithm-based Ensemble of Ensemble” (GA2Ensemble) achieves an impressive 99.34% accuracy on the KSDAgriPest dataset and competitive results on other datasets (DO: 98.99%, SMALL: 96.21%, IP102: 69.56%). GA2Ensemble proves effective for insect pest species identification, particularly on challenging datasets.