Rapid Vitality Estimation and Prediction of Corn Seeds Based on Spectra and Images Using Deep Learning and Hyperspectral Imaging Techniques

Abstract

Highly viable seeds are of great significance for agricultural development, and the traditional corn seed vigor detection method is time-consuming and laborious. In this paper, the spectral and image information of hyperspectral imaging was used, and a distinction between seed vigor detection and prediction was proposed. The potential of hyperspectral imaging technology and convolutional neural networks (CNNs) to identify and predict maize seed vitality was evaluated. The hyperspectral information in 10 hours before the germination of four vigor level seeds (144 samples each) was collected. A support vector machine, extreme learning machine, and one-dimensional convolutional neural network (1DCNN) were used to model the spectral data set, comparing the effects of multidimensional scattering correction and principal component analysis. 1DCNN performed best on the original spectral data, reaching an accurate recognition of 90.11%. According to the spectral changes of the seed germination, the first three hours of data were selected for prediction, which had higher recognition accuracy than the test set. The image-based 2DCNN model achieved 99.96% accurate recognition at a fast convergence speed. By differentiating the spectra and image information, the various CNN models can achieve accurate detection and prediction, providing a framework to advance research on seed germination.

Prediction of Sweet Corn Seed Germination Based on Hyperspectral Image Technology and Multivariate Data Regression

“…From Table 1 , we can notice that the best correlation coefficient was 0.7805, which indicated the relevance between the hyperspectral data and root length of seeds. The result further verified the feasibility of the hyperspectral method in seed germination prediction, as in many other studies [ 33 , 34 , 35 , 36 , 37 ].…”

“…The accuracy of the partial least squares discriminant a nalysis (PLS-DA) to distinguish aged (heat treated) and normal (untreated) corn seeds was up to 95.6% [ 35 ]. To combine the spectral and image information of HSI for seed vigor prediction, a multi-channel data acquisition system was used for image and spectral measurement in [ 36 ]. Hyperspectral information for a 10 h period before the germination of four vigor level seeds was collected and convolutional neural networks were applied for vitality evaluation.…”

“…Hyperspectral information for a 10 h period before the germination of four vigor level seeds was collected and convolutional neural networks were applied for vitality evaluation. By comparing different preprocessing and pattern recognition models, the convolutional neural network model had the best recognition by integrating the spectra and image information, and relatively high accuracy for the prediction of the four vigor grades was obtained [ 36 ]. Moisture content, which directly affects the storage time and seed germination rate, is also predicted by HSI in [ 37 ].…”

Application of Hyperspectral Imaging and Deep Learning for Robust Prediction of Sugar and pH Levels in Wine Grape Berries

“…Despite being mainly applied to classify two-dimensional (2D) images [ 27 ], CNNs have also been used to analyze one-dimensional (1D) data, such as speech recognition, text classification, and, more recently, spectral analysis. However, there are only a few recent studies showing that one-dimensional convolutional neural networks (1D CNNs) can be successfully applied to spectroscopic measurements (1D) in both classification [ 28 , 29 , 30 , 31 , 32 ] and regression [ 33 , 34 , 35 , 36 , 37 , 38 , 39 ] problems.…”

“…Regarding task (i), although deep learning can be trained on datasets without the use of preprocessing [ 40 ], there have been some recent works published in the scientific literature where the use of preprocessing combined with deep learning was applied to spectroscopic data, leading to results improvement. For instance, in [ 37 ], the raw spectrum was standardized using the standard normal variate (SNV) method before being fed into the CNN model; [ 39 ] evaluated the effect of using the original spectra or of the spectra preprocessed by the multiplicative scatter correction (MSC) method in a 1D CNN model for the prediction of corn seed, showing an improvement in accuracy for MSC + 1D CNN; [ 28 ] used a preprocessing strategy, combining different spectral preprocessing techniques, to develop CNN models in different spectroscopic datasets; [ 38 ] implemented extended multiplicative scatter correction (EMSC); [ 31 ] applied a Savitzky–Golay (SG) filter and logarithm methods to the reflectance spectra before 1D CNN.…”