Large scale multi-class pest image classification using structurally adapted DenseNet architecture.

The United Nations' Sustainable Development Goals, SDG 12: Responsible Consumption and Production, and SDG 13: Climate Action highlight the importance of environmental conservation and reducing pesticide use. Early and accurate pest identification is essential for implementing targeted pest control measures, which helps reduce unnecessary and incorrect pesticide use. While effective pest recognition and classification are crucial for ecological research and biodiversity conservation, traditional methods remain labor-intensive, time-consuming, and dependent on experts. Several deep learning techniques have been introduced in recent years, leading to more efficient and accurate identification and classification of crop pests. This research presents a structurally adapted DenseNet model for multi-class pest image classification based on dense connections. The model is fine-tuned through hyperparameters involving dense blocks and transition layers to perform consistently across three different datasets, including the IP102 dataset, which contains over 75,000 images of 102 pest species. The study also addresses dataset imbalance to prevent biased outcomes by deep learning models. The proposed structurally adapted model for fine-grained classification achieves 82.69% accuracy and 81.45% F1 score on the IP102 dataset, complementing existing advanced methods.