Vis-NIR hyperspectral imaging for sorghum variety discrimination: a systematic evaluation of preprocessing, feature selection, and classification models

In the vast world of agriculture, a single grain of sorghum holds a story of its origin, its quality, and its potential use as food, feed, or fuel. Sorting these grains by hand is slow and prone to human error, while chemical tests destroy the very seeds farmers need to sell. To solve this, scientists have turned to a technology that sees more than the human eye can: hyperspectral imaging. This method captures a detailed "fingerprint" of light reflecting off a surface across hundreds of narrow color bands, stretching from the visible spectrum into the near-infrared. Because different chemical bonds in starch, protein, and water absorb light in unique ways, this fingerprint reveals the hidden biochemical makeup of the grain. The challenge, however, lies in interpreting this massive amount of data. Researchers must decide how to clean up the noisy signals, which specific colors of light matter most, and which mathematical tools are best at sorting the varieties. Without a clear guide, different studies often use different methods, making it difficult to know which approach truly works best.

A team of researchers at Jilin Business and Technology College set out to bring order to this complexity by testing every major strategy in a single, controlled experiment. They gathered 1,800 intact seeds representing nine distinct sorghum varieties and scanned them with a hyperspectral camera that recorded 300 different wavelengths of light. The goal was not just to find a working method, but to determine the most reliable one by systematically comparing five ways to clean the data, five ways to select the most important information, and five different computer models to do the sorting. The researchers treated the data with extreme care, splitting it into separate groups for training and testing to ensure the results were genuine and not just lucky guesses. They also applied rigorous statistical checks to measure exactly how confident they could be in their conclusions.

The investigation began by asking how best to prepare the raw light signals. The researchers tried smoothing out the noise, correcting for how light scatters off the grain surface, and even calculating how quickly the light intensity changes across the spectrum. They found that a specific combination worked best: a technique that smoothed the data and then calculated the first derivative, which essentially highlights the steepness of the light curves. This process sharpened the subtle differences between the varieties, making them easier to distinguish. When they fed this sharpened data into a classic statistical model known as linear discriminant analysis, the system correctly identified the variety of the sorghum in more than 90 percent of cases. This was the highest success rate achieved in the study, establishing a new benchmark for how to sort these grains.

The team also explored whether they could simplify the task by using fewer colors of light. They tested methods that tried to pick out only the most useful wavelengths, hoping to make the process faster and cheaper. One method, which used a smart sampling technique to select the most informative bands, came very close to the performance of using all 300 colors, maintaining high accuracy while discarding nearly a quarter of the data. However, other methods that tried to compress the data into a few broad categories performed poorly, suggesting that for this specific task, keeping the full spectrum of information is crucial. The study also tested a modern deep learning model, a type of artificial intelligence that often excels at complex pattern recognition. Surprisingly, this advanced system struggled, achieving an accuracy of less than half that of the simpler statistical methods. The researchers concluded that for datasets of this moderate size, the complex deep learning approach was not the right tool, and that well-tuned, traditional models remain the most practical choice for real-world grain screening.

Ultimately, the study provides a clear roadmap for the future of grain quality control. It demonstrates that rapid, non-destructive identification of sorghum varieties is not only possible but can be achieved with high reliability using established, robust methods. The findings suggest that the key to success lies not in the most complex technology, but in the careful preparation of the data and the selection of the right, proven mathematical tools. By confirming that a specific combination of light processing and statistical analysis works best, the researchers have offered a dependable standard for breeders and quality control experts to use in protecting the integrity of the grain supply chain.