Seed vigor is a notable characteristic that allows efficient pre- and postharvest management by seed scientists. As a field emergence test, the direct seed vigor test is laborious and resource-consuming [1, 8]. In addition, the results are subjective, as they are heavily dependent on the expertise of the tester. Therefore, seed scientists have been attempting to find a convenient, economical and standardized method for seed vigor testing . Unfortunately, to date, there has been no report of a universally acceptable single indirect test for assessing seed vigor . For this reason, multiple tests should be used together to make the best assessment. The indirect methods commonly used produce one parameter per test. Therefore, an indirect method of seed vigor testing requires more than one method to confirm the results. In the past, seed scientists have used two or three laboratory-based indirect methods together to accurately determine seed vigor, including the stress test, electrical conductivity test and seed growth rate test [2, 31, 33]. Other methods, such as fast ethanol assays, do not consider seedling performance but rely on expensive sensors and have not been reported to be successful with rice seeds [34, 35].
Recently, the radicle emergence test has been recognized as being effective and fast for seed vigor testing, as supported by the metabolic repair hypothesis . This hypothesis provides the overall physiological basis to explain the principles behind the standard germination and vigor tests, but the radicle emergence test is sensitive and can vary according to species and test conditions [9, 28, 36]. Interestingly, Joosen, et al.  offered GERMINATOR software that can calculate more than one parameter in a single germination test with a cumulative germination curve through the four-parameter Hill function. Furthermore, it is possible to classify seed vigor automatically using machine vision combining the concepts of the GERMINATOR software with image processing and cluster analysis as revealed in the SVRice package. After imaging (image or video), the information on radicle emergence for each seed at a particular time is processed and plotted as a cumulative radicle emergence curve based on the Hill function. This allows the radicle emergence indices to be calculated for multiple parameters (MaxRE, MRET, t50, U7525 and AUC). The MRET, U7525 and t50 values are of particular importance because they had a strong negative correlation with the SSAA test (Fig. 5), and they could accurately determine rice seed vigor after cluster analysis. The SSAA test classified seed vigor into 2 groups using ANOVA and a post hoc test (Fig. 2). However, the analytical results were questionable due to entanglements in the classification. In contrast, the results from SVRice could be clearly separated using the five radicle emergence indices (MaxRE, MRET, t50, U7525 and AUC), as shown in Fig. 6e. SVRice successfully differentiated samples from the same variety but in different production years. The number of clusters was reliable because relative clustering validation was used to determine the optimal number of clusters and evaluated the clustering structure using varying parameters for the same algorithm (varying the number of k clusters) . In comparison with high-end seed phenotyping devices such as the SeedGerm system , our image-processing algorithm and hardware design follow a specific-use and easy-to-build strategy. Unfortunately, we cannot compare the results of SVRice with other germination phenotyping platforms using the same images because their commercial versions are not made for rice seeds.
A germination time of 90 hours at 25°C was sufficient for effective classification based on SVRice, whereas the SSAA test took approximately 400 hours to complete. These results were consistent with Onwimol, et al. , who proposed a single count of radicle emergence at 110 hours after setting to germination at 25°C for the identification of the vigor of rice seeds. In the current experiment, especially under ambient storage, the radicle emergence behavior of rice seeds was very different for rice seed samples covering the popular rice varieties in Thailand (Fig. 4). The SVRice software algorithm was set up to be especially suitable for assessment after 6 months under controlled atmosphere storage because this storage condition and period are common for rice production in Thailand (Fig. 3b).
Our image-processing algorithm can be easily adapted to desired varieties or radicle emergence behavior through the adjustment of the number of groups in Otsu’s multiple thresholding in the seed segmentation step and the structuring element for morphological dilation in the counting step. For calibration, the results after adjustment of the algorithm can be compared with the results from visual analysis—manual scoring—that can be used as a reference. Uncontrolled illumination in the image acquisition step might result in failure to obtain a result in the seed segmentation step. Otsu’s multiple thresholding might group some parts of the seed into the background. Long, messy overlapping or touching roots, as shown in Fig. 4c, might result in an incorrect number of seeds in the counting step. The counted seed number might be greater than the number of seeds because of substantial variations caused by too many overlapping radicles at the late germination stage. Due to the potential of the K-means, the limitation of SVRice is the classification ability of the package that can be classified up to the grouping level but cannot be classified to a single sample level . However, it should be apparent that the classification of seed vigor up to a single sample level does not provide much information to seed scientists to make decisions during seed inventory management, which, if necessary, can be solved by using an appropriately trained protocol with reference to an authentic dataset of the in-house protocol.
The developed software is suitable for high-throughput seed vigor classification. The SVRice package utilizes dynamic imaging to compare the image to that from previous time points. Hence, it can detect changes in radicle emergence better than using single end-point imaging analysis. Consequently, the results are more accurate, and less time is required. SVRice software can be applied to detect seed health and seed damage from insects during storage in a warehouse if multispectral imaging is utilized . The SVRice pipeline offers a well-defined and robust experimental setup but is flexible in terms of numbers and treatments. Improved efficiency and the absence of subjectivity are great advantages in computer-aided assessment. Automatic seed vigor classification was optimized for indica rice and would most likely work for many other species as well.