Adaptive feature fusion of ResNet50 and vision transformer for robust agricultural pest classification

The classification of agricultural pests is a complicated issue since the inter-class, intra-class, and environmental factors like changes in illumination and background clutter are hard to classify. These challenges limit the effectiveness of existing deep learning models, which do not always jointly detect fine-grained local texture and global contextual relationships.In order to overcome this problem, we present an Adaptive Feature Fusion Network (AFFN) that combines ResNet50 and Vision Transformer (ViT) based on a temperature-regulated gating system. The suggested framework dynamically balances local and global feature contribution, avoids dominance of representations and enhances stability of training. Moreover, preprocessing using CLAHE is added to improve the discrimination of features in diverse light intensities.Experiments on the Agricultural Pest Dataset show that the proposed approach has a validation accuracy of about 83%, which is far better than standalone CNN, Vision Transformer, and static hybrid models. The model, also, has better convergence, robustness, and class-wise discrimination. These findings suggest adaptive feature fusion as an effective and scalable fine-grained agricultural image classification method which is ideal to real-world precision agriculture systems.