Abstract
The identification of pests hidden in stored wheats, essential to grain storage safety, is a key difficulty in the research of target detection. This paper introduces the support vector machine (SVM) classifier to identify the pests hidden in wheat kernels, and selects the proper kernel function and parameters to classify various samples. It is verified that the proposed method could accurately detect the pests in wheat kernels. This research provides new insights into the application of pattern recognition in bio-photon detection of pests in stored grains.
Introduction
Detecting pests hidden inside grain kernels is a major hurdle for food safety, as traditional methods often only spot infestations once adults emerge. The "hidden enemy" in grain storage is the early-stage pest—larvae or eggs tucked away inside individual wheat kernels. While current granary technology can effectively monitor temperature, humidity, and the presence of adult insects, there is a significant gap in detecting these internal infestations before they escalate.
By the time pests are visible to the naked eye or traditional sensors, the damage is often already severe. This research proposes a non-invasive detection method using ultra-weak bio-photon emission analysis—measuring the light naturally emitted by biological tissues—paired with a Support Vector Machine (SVM) classifier to distinguish between healthy and infested wheat without damaging the grain.
Methodology
The researchers combined quantum optics with machine learning to create a diagnostic tool. The process involved three primary stages:
- Bio-photon Acquisition: Using a specialized detection system to measure the ultra-weak photon radiation from both healthy wheat and wheat infected with maize weevils. The study utilized 200g samples of each, maintained at 14% moisture and 28°C.
- Feature Extraction: Ten specific "characteristic parameters" were extracted from the time-domain signals. These included standard statistics like mean, variance, and kurtosis, alongside specialized metrics like photon statistical entropy, which characterizes the order or disorder of the photon distribution.
- SVM Classification: An SVM was employed to classify the samples. Because the relationship between bio-photon signals and infestation is non-linear, the team tested four different kernel functions—Linear, Polynomial, Gaussian (RBF), and Sigmoid—to determine the optimal mapping for high-dimensional separation.
Results and Discussion
The experimental results validated that the SVM-based approach is highly effective for this application. A key reason for choosing SVM over deep learning models was its proven performance in "small-sample" problems, providing a robust mathematical foundation for high-dimensional pattern recognition in niche agricultural sensing.
Kernel Performance Comparison
The Gaussian (Radial Basis Function) kernel significantly outperformed the others. While linear kernels struggled with the complexity of biological signals, the Gaussian kernel’s ability to handle non-linear relationships allowed it to achieve a best-case recognition accuracy of 93.33%.
As shown in the data, the Gaussian model achieved a missing rate (false negatives) of only 6.67% and a false alarm rate of 13.33%. In contrast, the linear kernel had a missing rate of 26.89% and a false alarm rate of 33.56%, proving that simple linear separation is insufficient for bio-photon data. Furthermore, the inclusion of photon statistical entropy suggests that the structure of light emission changes when a pest is present, as biological interactions alter the deterministic value of the photon distribution.
Conclusion
This study demonstrates that SVM-based classification of bio-photon emissions is a viable method for the early detection of pests hidden within wheat kernels. While the laboratory results are promising, transitioning this technology to real-world granaries involves overcoming environmental noise and scaling the hardware for massive storage volumes. Future work will focus on verifying species specificity and integrating these sensors into portable granary monitoring systems.
Summary updated