Two-stage machine vision and near-infrared spectroscopy for grading leafhopper damage in fresh tea leaves

Abstract

Abstract Leafhopper feeding-damage severity is an important factor in the raw-material grading and final quality of Zijin Chan tea, but conventional assessment relies heavily on subjective visual judgment. This study developed a two-stage workflow combining machine-vision screening with offline near-infrared (NIR) spectroscopic reassessment. First, an improved small You Only Look Once version 8 (YOLOv8s) model localized tea-leaf targets and classified slight, moderate, and severe feeding damage. Physical samples corresponding to machine-vision no-result targets were then reassessed using an NIR model integrating standard normal variate preprocessing, Pearson correlation-based spectral-region selection, the successive projections algorithm, and a support vector machine. On 200 test images containing 4,000 annotated targets, the improved YOLOv8s model achieved precision of 88.6%, recall of 87.7%, and mean average precision at an intersection-over-union threshold of 0.5 of 89.3%, representing improvements of 12.2, 9.3, and 8.7 percentage points over the baseline, respectively. Developed from 240 independent physical samples, the NIR model achieved macro-averaged recall of 94.03% on the cross-validated training set and 95.12% on the spectral model-selection set. When the fixed model was applied to physical samples corresponding to 492 machine-vision no-result targets, 473 targets were classified correctly, yielding a reassessment accuracy of 96.14%. Across all targets, standalone machine vision achieved an end-to-end accuracy of 77.7%, whereas the two-stage workflow achieved 89.5%. These results support the feasibility of combining visual and spectral information for laboratory-based grading. Validation using in situ spectra and samples from multiple batches and seasons remains necessary before online deployment.

Code and data availability

The paper describes a 2,000-image tea-leaf dataset (40,960 bounding boxes) and NIR spectral data, but no public deposit or URL is provided. Both datasets and custom code are stated to be available only from the corresponding author on reasonable request, so no public, actionable paper-specific asset exists in the text.

No evidence-backed public reproduction asset is currently recorded.