PestNet50: Pest Classification via Swin-Hybrid Networks with Auxiliary Supervision

Abstract

Pest classification is an important task in precision agriculture because early identification of harmful species can support timely treatment and reduce unnecessary pesticide use. However, many public insect datasets combine multiple developmental stages, geographic regions, and acquisition conditions, which increases intra-class variation and complicates fine-grained adult pest recognition. We introduce PestNet50, a dataset containing more than 46,000 RGB images from 50 adult pest species relevant to Asian agriculture. Images were collected from public iNaturalist observations, manually verified, cleaned, and standardized to 75×75 pixels. We also propose Swin-Hybrid with Auxiliary Supervision, a lightweight architecture that inserts depthwise convolution into the Swin Transformer feed-forward network to strengthen local texture modeling. Lightweight auxiliary classifiers provide intermediate supervision during training and are removed during inference. Experiments against five representative CNN and transformer baselines show that the proposed model achieves the best performance, reaching 94.3% test accuracy and 0.941 macro F1-score with 28.5M parameters. These results demonstrate that adult-stage standardization and local enhancement of hierarchical transformers are effective for practical fine-grained pest recognition.