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

High-quality tea is often a story of damage. In the case of Zijin Chan, a distinctive tea from southern China, the very character of the final cup depends on the fresh leaves being nibbled by a tiny insect called the tea green leafhopper. As the insect feeds, it triggers chemical changes within the plant that create a unique honey-like aroma and flavor. The severity of this feeding determines the tea's grade and its market value. For centuries, farmers have judged this damage by eye, looking for specific signs like leaf color, the presence of brown spots, and how much the leaf has curled. However, human judgment is slow and inconsistent; it varies from person to person and changes with the light or the observer's fatigue. To produce a consistent, high-quality product, the industry needs a way to measure this damage quickly and objectively, without relying on the subjective eye of a human expert.

This is where the intersection of two different ways of seeing the world comes into play. One way is machine vision, where a camera captures the external appearance of an object, much like a human eye but with the ability to process thousands of images in seconds. The other is near-infrared spectroscopy, a technique that looks inside the material to read its chemical composition by analyzing how it absorbs light. While a camera can see a brown spot, it cannot see the sugar or amino acids changing inside the leaf that give the tea its flavor. By combining these two methods, researchers at Jiangnan University have developed a new system that acts like a two-step filter, using the speed of the camera to screen most leaves and the chemical insight of the spectroscopy to double-check the ones the camera finds confusing.

The researchers began by training a computer model to recognize and sort fresh tea leaves based on the severity of the insect damage. They used a sophisticated image-recognition system, an improved version of a technology known as YOLO, which is designed to find objects in images and classify them instantly. The team fed the system thousands of images of tea leaves, teaching it to distinguish between three levels of damage: slight, moderate, and severe. They also taught the model to ignore the messy reality of a pile of leaves, where leaves overlap, curl, and hide parts of each other. The improved model learned to focus on the specific visual cues that matter, such as the color of the tea bud, the extent of the brown scorching on the leaf edges, and the degree of curling.

When tested on a large batch of images containing four thousand individual leaf targets, this improved camera system performed impressively. It successfully identified and classified the vast majority of the leaves it could see clearly. However, like any system relying on sight, it hit a wall when the visual information was too ambiguous. In about twelve percent of the cases, the camera could not make a confident decision. These were the "no-result" targets: leaves that were too curled, too obscured by neighbors, or simply too similar in appearance to tell the difference between slight and moderate damage. If the process stopped here, these difficult leaves would remain ungraded, or worse, be graded incorrectly, potentially ruining the quality of the final tea batch.

To solve this problem, the researchers introduced a second stage for these stubborn samples. Instead of trying to force the camera to see better, they turned to the chemical signature hidden inside the leaf. They took the physical samples corresponding to the leaves the camera couldn't classify and subjected them to near-infrared spectroscopy. This process involved freezing, drying, and grinding the leaves to create a uniform powder, which was then scanned with light. The light interacted with the chemical compounds inside the leaf—such as tea polyphenols, sugars, and amino acids—creating a unique spectral fingerprint. The researchers found that leaves with different levels of insect damage had distinct chemical profiles. For instance, leaves with moderate damage tended to have higher levels of certain polyphenols, while severely damaged leaves showed a rise in free amino acids.

Using these chemical fingerprints, the team built a second model to act as a referee. This model was trained on a separate set of two hundred and forty samples to learn the difference between the slight, moderate, and severe classes based purely on their internal chemistry. When this chemical model was applied to the difficult samples that the camera had rejected, it proved to be remarkably accurate. It correctly classified nearly ninety-six percent of the problematic leaves. By combining the speed of the camera with the chemical insight of the spectroscopy, the researchers created a workflow that could handle the entire batch of leaves. The camera did the heavy lifting, sorting the easy cases instantly, while the spectroscopy stepped in only when necessary to resolve the difficult ones.

The result of this two-stage approach was a significant leap in overall accuracy. If the researchers had relied on the camera alone, the system would have correctly classified only about seventy-eight percent of all the leaves. By adding the chemical reassessment for the difficult cases, the total accuracy rose to nearly ninety percent. This improvement demonstrates that while a camera is excellent at seeing the surface, it cannot always see through the complexity of nature. The chemical analysis provides a necessary backup, ensuring that even the most confusing leaves are graded correctly.

However, the researchers are careful to note that this success was achieved in a controlled laboratory setting. The chemical analysis required the leaves to be processed—frozen, dried, and ground—which is not how tea is handled in a real-time sorting line. The current system is a proof of concept, showing that the combination of visual and chemical data works in principle. The next step, which the authors identify as essential, is to adapt this method for fresh leaves that are moving on a conveyor belt, without the need for grinding or freezing. Until that happens, the system remains a powerful tool for laboratory grading and a blueprint for future online inspection systems, offering a path toward more consistent and higher-quality tea production.