Weight-based ensemble method for crop pest identification

Crop pests cause significant losses to agricultural production. Pests can be detected and controlled over time using accurate and effective methods, thereby reducing potential losses. However, there are challenges in realistic agricultural scenarios, such as diverse pest species and complicated environments, which render manual recognition and conventional machine learning methods insufficient. To address this issue, deep learning methods that can automatically extract features have recently been widely used for pest identification. However, accurately recognizing images that resemble complex real-world scenarios remains a challenging task for a single deep learning model. The ensemble method, which combines multiple basic models, provides a solution for improving recognition performance. In this study, we proposed two weight-based ensemble methods, VecEnsemble and MatEnsemble, constructed from vector- and matrix-based weights, respectively. The weights that combine basic models significantly influence the performance of the ensemble methods. Therefore, to effectively combine the basic models, we formulated the weight design problem as a quadratic convex optimization problem whose solution has a closed-form expression and can be computed efficiently. Our method achieved the highest accuracy of 77.39% on the large-scale complex-scene IP102 dataset, which was competitive with those of other state-of-the-art methods. Furthermore, we conducted comprehensive ablation experiments to compare our proposed methods with voting-based approaches and illustrate the scenarios in which they are applicable. These results highlight the practical significance of our method for agricultural production and provide a foundation for further research on crop pest identification. The source code is available at https://github.com/shiguangqianmo/WBEnsemble.