Classification of viable/non-viable seeds of specialty maize genotypes using spectral and image data plus morphological features

Abstract

Seed viability is an important consideration for agricultural production. The number of studies on the measurement of seed viability in specialty maize genotypes via new approaches is limited. This study was carried out to determine the viability of the seeds (n = 950) of two specialty maize (high oil and high protein) populations using spectral measurements and imaging techniques. Spectral data from the seed embryos were collected between 1200 and 2400 nm. Image data were taken with 300 dpi resolution. From the collected images, red (R), green (G) and blue (B) [RGB] data were extracted, and morphological features (M) were also determined. Then, the seed samples were separated into two sets and the viability of the samples was determined using two different methods [standard germination (SG) test and triphenyl tetrazolium chloride (TTC) test]. Support Vector Machine (SVM), Random Forest (RF), and Classification and Regression Tree (CART) methods were used to develop the classification models (n = 36). Classification accuracy of the models was comparable for the SG test (0.56–0.91) and TTC test (0.53–0.85). However, the classification models based on TTC test results had higher sensitivity (0.86–0.99) than specificity values (0.07–0.74), which indicated that the viable seeds were more accurately identified than the non-viable seeds. The RF model, created using the NIR+M dataset, based on the SG test (sensitivity = 0.89, specificity = 0.94, accuracy = 0.91), was most effective for determination of the seed viability of specialty maize genotypes used in this study.