Answers are grounded in this paper and cite their sources.
Reliable in-field rice pest monitoring is essential for intelligent crop protection, but it remains challenging because pest targets are usually small, visually similar to background foliage, and affected by variable illumination, weather, occlusion, and long-tailed category distributions. Existing studies are further limited by the lack of field-representative benchmarks that simultaneously cover pest categories, beneficial insects, and hard negative samples. To address these issues, this study first constructs RicePest22, a 22-category field benchmark designed for realistic rice-pest surveillance under diverse field conditions. We then propose RSOP-DETR, an end-to-end Transformer-based detector that integrates adaptive hierarchical augmentation, a Dual Cross-layer Connection Module (DCCM), RepNCSPELAN-based feature aggregation, and Wise-IoU v3 localization loss. These components jointly improve minority-class learning, shallow spatial-detail preservation, multi-scale feature aggregation, and hard-sample localization. Experimental results on RicePest22 show that RSOP-DETR achieves an F1-score of 87.0%, an mAP 0.5: 0.95 of 60.8%, and an mAP s of 14.4%, outperforming RT-DETR and recent YOLO-style detectors in overall detection accuracy. Although very small pest detection is still difficult under severe occlusion and low-contrast backgrounds, the results indicate that the combination of a realistic benchmark and task-oriented detector design provides a practical basis for AI-assisted rice pest monitoring and precision crop protection.
No takes yet. Share an insight, caveat, or question.
Lin et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: