Deep learning-based recognition of insect pests in plant images

Early and effective insect pest detection and control help to protect plants, increase crop yields, and reduce losses in the agricultural economy. In this dissertation, deep learning approaches for identifying insect pests, which pose a significant threat to agricultural production, are proposed. The first proposed method involves classification through transfer learning using different pre-trained deep neural networks (DNN). In the second method, features extracted from deep layers of these networks with the Support Vector Machine (SVM) classifier are utilized, Third method involves the use of a neural network model that uses transformer architecture known as the vision transformer. Finally, ensemble learning, an approach that combines predictions from multiple models to achieve stronger and more accurate predictions, is used to merge the best-performing models from the three different methods mentioned above. Additionally, a comprehensive performance analysis was conducted using various image preprocessing techniques both individually and in combination, including green color channel extraction, data augmentation, histogram equalization, and deep learning-based segmentation with background elimination. Experiments were carried out on the Li and D0 datasets, which contain 10 and 40 plant pest species, respectively. The highest performance was achieved with 98.35% accuracy using the majority voting method in the Li dataset, and 99.78% accuracy in the D0 dataset. The results demonstrate that the proposed models can be effectively used in insect pest control.