Non-destructive assessment of hemp seed vigor using machine learning and deep learning models with hyperspectral imaging

Abstract

Hyperspectral imaging was employed to capture spectral information from entire trays of hemp seeds. Individual seed spectral data was extracted using a region-of-interest analysis, isolating each seed for detailed examination. To simplify the analysis and reduce computational complexity, a subset of key spectral wavelengths was selected using a successive projection algorithm. Deep learning models were trained on these selected wavelengths to directly learn patterns from the raw spectral data. The performance of these deep learning models was compared to traditional machine learning approaches. Particularly, an EfficientNetB0 convolutional neural networks achieved the most impressive results, demonstrating a high sensitivity of 98.85, a specificity of 99.22, and a Matthews correlation coefficient of 0.98. It indicated its ability to accurately distinguish between high-vigor and low-vigor hemp seeds. Our findings demonstrated the potential of data-driven models trained on hyperspectral imaging data for non-destructive assessment of hemp seed vigor. This approach offers an advantage over traditional methods, which often involve destructive testing or time-consuming manual evaluation. By enabling rapid and objective seed selection, this technology can improve the efficiency of hemp seed production and ultimately lead to higher crop yields. This innovative approach has the potential to revolutionize the agricultural industry by providing a powerful tool for assessing seed quality and optimizing crop production.