The level of moisture content and size of winter melon seeds affect their storage, germination and growth processes. Moderate moisture and size are essential for seed germination and growth. Therefore, detecting the moisture and size of winter melon seeds is beneficial to improve their germination rate. Traditional seed moisture testing methods are complex to operate and require destructive sample preparation or chemical treatment. Common dimensional measurement methods are also time-consuming and laborious. Hyperspectral imaging technology can acquire information about the surface and internal structure of the target separately, it can be used to quickly and non-destructively detect the moisture and size of winter melon seeds. In this study, partial least squares regression (PLSR) and partial least squares support vector machine (LSSVM) models were established to predict the moisture content of winter melon seeds by using reflection and transmittance spectral data. The models were optimized using five variable selection methods. The optimal performance of the LSSVM model based on the CARS algorithm was achieved in both single reflection and transmission spectra. The RP2 and RMSEP of the model based on the reflection spectra were 0.9667 % and 0.0215 %, respectively. The RP2 and RMSEP of the model based on the transmission spectra were 0.9644 % and 0.0223 %, respectively. In low-level data fusion, the LSSVM model based on the CARS algorithm also achieved optimal performance, but with only a little improvement compared to a single model (reflection spectra or transmission spectra), with RP2 and RMSEP of 0.9679 % and 0.0212 %, respectively. In the mid-level data fusion, the LSSVM model also based on the CARS algorithm achieved the optimal performance, and the performance of the model was further improved. The RP2 and RMSEP of the model were 0.9738 % and 0.0191 %, respectively. Finally, the image processing algorithm and the convex polygon size measurement method was proposed to measure the size of winter melon seeds. The absolute error between the calibrated winter melon seed length and width and the true length and width was less than 0.22 mm, and the relative error was less than 2 %. The results show that hyperspectral imaging technology can accurately detect the water content of winter melon seeds. Data fusion method and LSSVM model based on CARS algorithm can detect the water content of winter melon seeds more accurately. The image processing algorithm combined with the convex polygon size measurement method can be effectively used to detect the size of winter melon seeds with high precision.
Read full abstract