GEPC: Generative-Explanatory Pest Classification with Multi-Modal Priors and Hierarchical Concept Constraints

Abstract Accurate and interpretable pest classification underpins precision agriculture and food security, yet three obstacles remain. Field-collected datasets are severely long-tailed, so rare and often high-risk pests are under-represented; closely related species are morphologically near-identical, limiting single-modality discrimination; and standard end-to-end classifiers emit labels without auditable reasoning. We present GEPC (Generative-Explanatory Pest Classifier), a unified framework organized around one principle: every prediction should be backed by a verifiable taxonomic pathway. To counter the long tail, a semantics-guided diffusion pipeline synthesizes ecologically plausible samples for rare classes through structured prompts and CLIP-based quality filtering, grounding augmentation in real taxonomic and habitat priors. For interpretability and fine-grained discrimination, GEPC casts feature extraction as non-negative, path-constrained sparse pursuit via Hierarchical Beam Orthogonal Matching Pursuit (HB-OMP) over an offline differential dictionary built on the Order--Family--Species tree, so the recovered root-to-leaf path denoises a frozen DINOv2 visual representation while doubling as a transparent diagnostic cue. This hierarchy-aligned signal is fused with species-anonymized morphological descriptions encoded by a frozen CLIP text encoder, and a lightweight classifier issues the final prediction. Across multiple random seeds on AP162, GEPC attains $90.4 \pm 0.3\%$ overall accuracy, $82.4\%$ tail-class recall, and $94.6\%$ agreement between the visual pursuit path and the final decision, consistently outperforming end-to-end supervised and zero-shot baselines while exposing the taxonomic evidence behind each prediction.