Robust Fine-Grained Pest Classification via Boundary-Aware Attention and Growth-Stage Supervision.

Accurate pest identification plays a critical role in effective pest management and sustainable agricultural production. However, fine-grained pest classification is inherently challenging due to low inter-class separability, significant intra-class variability, and complex environmental interference. To address these challenges, we develop a boundary-aware channel-spatial attention network to strengthen discriminative feature learning while suppressing background noise. The proposed attention module enhances fine-grained structural and boundary cues to improve inter-class separability under cluttered field conditions. Furthermore, auxiliary label supervision based on pest growth stages is incorporated to model developmental variations and enhance intra-class consistency. Experiments on the IP102 dataset demonstrate that the proposed method consistently outperforms state-of-the-art baselines in classification accuracy, validating its effectiveness for fine-grained agricultural pest classification. These results highlight the potential of integrating boundary-aware attention mechanisms with growth-stage supervision for robust real-world pest classification.