Abstract: Rice crops are frequently threatened by pests such as rice planthoppers (Nilaparvata lugens, Sogatella furcifera, and Laodelphax striatellus) and leafhoppers (Cicadellidae), leading to significant agricultural yield losses. Accurate identification of both pest developmental stages and natural predators is crucial for effective pest control and ecological balance. Conventional field surveys often suffer from subjectivity, inefficiency, and lack of traceability. To overcome these challenges, this study proposes RiceInsectID, a two-stage cascaded detection method for identifying and counting tiny rice pests and their natural predators in white flat plate images captured by head-worn augmented reality (AR) glasses. The method can recognize 25 insect classes, including 17 instars of rice planthoppers, two instars of leafhoppers, four spider species (Araneae), Miridae, and rove beetles (Staphylinidae Latreille). At the first coarse-grained detection stage, 16 visually similar classes are consolidated into 6 broader categories, detected using an enhanced YOLOv6. To improve the detection of small objects and address class imbalance, the full-region overlapping sliding slices and target pasting (FOSTP) algorithm increases mean average precision at 50% IoU threshold (mAP50) by 35.46% over the baseline YOLOv6. Feature extraction and fusion are improved by efficient channel attention path aggregation feature pyramid network (ECA-PAFPN) and adaptive structure feature fusion (ASFF) modules, while the balanced classification mosaic (BCM) enhances the detection of minority classes. Test-time augmentation (TTA) further improves mAP50 by 2.06%, achieves 84.71%. In the second fine-grained classification stage, each of the six broad classes from the first-stage output is further classified using individual ResNet50 models, with online data augmentation and transfer learning to significantly enhance generalization. The two-stage cascaded detection method improves recall by 4.06%, precision by 3.79%, and the F1-score by 3.92% over the baseline YOLOv6. RiceInsectID achieves 82.85% recall, 80.62% precision, and an F1-score of 81.72%, offering an efficient solution for monitoring tiny rice pests and their natural predators in paddy fields. This study provides valuable insights for ecosystem monitoring and sustainable pest management in rice agriculture.

Key words: tiny rice pest, natural predator, AR glasses, intelligent survey, object detection, fine-grained recognition