Evaluation Analysis of Insect Classification Model in Pest Detection through Grad-CAM Approach to Improve Interpretability

Abstract

The article explores the intricate use of technology in agriculture, employing the EfficientNet model for pest classification and incorporating the Grad-CAM method as an interpretive tool. The evaluation encompasses a thorough analysis, including a visual examination of the model. The chosen approach involves presenting an in-depth assessment of the reliable EfficientNet model in insect classification. Grad-CAM, an abbreviation for Gradient-weighted Class Activation Mapping, is utilized as a visualization technique, focusing on generating saliency maps to highlight significant areas in images influencing the classification model's predictions. EfficientNetV2, an evolved version of the CNN model, demonstrates faster training speed and superior parameter efficiency compared to its predecessor, EfficientNet. The study conducts a comprehensive analysis of insect classification models for pest detection, utilizing Grad-CAM techniques and the architecture of EfficientNetV2-L. The approach integrates the advantages of EfficientNetV2-L, known for its faster training speed and better parameter efficiency, with the implementation of Grad-CAM, a technique providing visual insight into critical areas in model predictions. In conclusion, the research underscores that combining EfficientNetV2-L with the Grad-CAM approach in insect classification enhances model performance and interpretability. The integration of advanced model architecture with visualization techniques like Grad-CAM offers profound insights into understanding insect classification models, establishing a crucial foundation for more effective and efficient pest detection efforts in the future.