Detection of Water Content of Watermelon Seeds Based on Hyperspectral Reflection Combined with Transmission Imaging

Watermelon is a widely cultivated fruit and vegetable that is native to Africa and has become one of the world’s important summer fruits. Watermelon seed vigor has a critical impact on watermelon planting and yield, and seed water content is a key factor in maintaining vigor during seed storage and germination. In this study, reflectance and transmittance spectral data from hyperspectral imaging were fused to improve the detection accuracy of moisture content in watermelon seeds. First, watermelon seed samples with different water content gradients were prepared by dividing all 456 selected watermelon seeds into 10 groups and drying them in a drying oven at 60 °C for 0, 3, 5, 10, 15, 20, 25, 30, 40, and 50 min. Reflectance and transmission spectra of 456 watermelon seeds were collected by a hyperspectral imaging system, and the single spectral data were subsequently used to build PLSR and LSSVR models for quantitative analysis of watermelon seed moisture content. Model performance is enhanced by Competitive Adaptive Reweighted Sampling (CARS), Unrelated Variable Elimination (UVE), and primary and intermediate data fusion methods. Primary data fusion improves model predictions compared to single models based on reflectance and transmission spectra. The intermediate data fusion of the feature spectral data of reflectance and transmittance selected by the CARS algorithm improves the prediction effect of the model more obviously, in which the model with the best prediction accuracy is Raw-CRAS-LSSVR, whose R P 2 and RMSEP are 0.9149 and 0.0144, respectively, which improves the prediction effect of the model built by a single full-spectrum datum by 5.72%. This study demonstrates that hyperspectral reflectance and transmission imaging techniques combined with data fusion can effectively detect watermelon seed moisture content quickly and with high accuracy.