| Agriculture is the nation’s foundation, while seed is the most important element of agriculture. Seed quality has close relationship to crop yields and farmer’s benefits. Improving the quality of seeds and regulating the management of seed market have important implications for increasing crop yield and farmer’s incomings. Maize is one of the world’s three major grain crops. Maize seeds have high commercialization degree and many cultivars. The development of precision planting technology requires higher seed quality. The research of seed sorting and nondestructive test technology is benefit to seed breeding and maize farming.Hyperspectral imaging technology was employed to investigate the methods of how to sort immature maize seed, recognize the authenticity of seed cultivar, classifying the cold damage seed and the degree of cold damage. Results prove that hyperspectral imaging technology combined with proper data processing method has the potential to detect the quality of single maize seed.The main results and conclusions are as follows:(1) A method for removing background of seed hyperspectral image by using binaried variance image of 500-580 nm was introduced. The binaried variance image was used as mask to remove the background of each band image in the hyperspectral image cube. This method can overcome the effect of shadow at single seed edge and segment seed from background.(2) A method for identifying immature maize using image technology was introduced. ROI spectral of immature and mature regions on maize seed were extracted. KW examination was employed to select the best band-ratio wavelength. The segmentation results of the band-ratio image prove that this method is appropriate for the segmentation of immature area. The correct recognition rate of mature and immature seed is 93.9% using 864 seeds samples.(3) A method for classifying maize seed cultivar was introduced. After selecting characteristic wavelength by SPA, the characteristic spectrum, mean texture features of characteristic band images and morphological features were extracted from hyperspectral images of seed embryo and endosperm side. The fusion of these features were used as the input variables of SVM. Using ten cultivars maize seed as test samples (1855 kernels), the accuracy classification rate of embryo and endosperm test dataset is 96.8% and 96.6%.(4) Considering the large number of maize seed cultivars, all of the cultivars can’t be collected for the multi-class classifier in many circumstance. For this reason, a method based on SVDD-KNNDD was introduced for justify the authenticity of seed cultivar. Users only need the cultivar samples which need to build the classify model when training this classifier, and the performance of the classifier is only relate to this cultivar samples. We have no need to collect all of the cultivars to train our model. The best test results is occurred at the embryo side spectral preprocessed by SGFD and build model using SVDD-KNNDD. For ten cultivars of test samples, the correct recognition rate of most of the abnormal cultivars is 100%, a few samples have a low recognition rate. The largest mistake classification rate of abnormal cultivars is 68.5% and the lowest correct classification of correct cultivars is 84%, most of the correct classification rate of correct cultivars is over 90%. This method can be used as a rapid test technology for seed authenticity testing, not only for maize seed, but also can use other seed samples.(5) Seed vigor decrease as the maize seed cold damaged. The cold damaged maize seed almost indistinguishable by eyes. We try to classify cold damaged seed samples from normal ones using hyperspectral images. Raw spectral preprocessed by MSC, SNV and SGFD can promote the correct recognition rate. The research results proves that mean spectral extracting from both embryo and endosperm side images combined with SGFD and SVM can classify the damaged samples and normal samples. The lowest correct recognition rate is 88% of samples cold at -5℃ using embryo side spectral and 73.6% using endosperm side spectral. The lowest correct recognition rate is 96.9% of samples cold at -18℃ using embryo side spectral and 95.2% using endosperm side spectral.(6) Considering lack of cold damage sample in reality circumstances, SVDD-KNNDD was employed to build abnormal sample detection model only use the spectral extracted from hyperspectral images of normal seed samples. When the samples is cold processed at -5℃ by 16 h, the correct recognition rate of cold damaged seeds is 96% using endosperm side data and 90% using embryo side data. The correct recognition rate of other samples processed at -5℃ by 4 h,8 h and 12 h are not very good. When the samples is cold processed at -18℃ by 4 h,8 h,12 h and 16 h, the lowest correct recognition rate is 89.7% and the highest is 100% using endosperm side data, the lowest correct recognition rate is 92% and the highest is 100% using embryo side data.(7) Hyperspectral imaging allows the evaluation of small samples with interactive selection of regions of interest, something that is not possible with conventional spectroscopy. To analyze the cold damage extent of embryo, embryo region hyperspectral image was segmented by band-ratio images, the best band-ratio was selected by KW test method. Extracted the spectral of pixels in embryo area and employ SVM to build classify models.Calculating the ratio of pixels which recognized as normal or various cold extent samples to the whole embryo area, the ratio used to classify the cold damage extent. When the samples is cold processed at -5℃, the correct recognition rate of control samples and cold damaged by 16h is the same 99.2%,4.2% of cold damaged by 4h samples recognized as control samples,6.7% of cold damaged by 12 h samples recognized as cold damaged by 16 h samples. The correct recognition rate of cold damaged by 4,8 and 12 h is not very good. The spectral of pixels in segmented embryo region can be used to separate the normal and cold damage seed. This method can be used as reference method for classifying serious cold damaged seed and normal seed, but is not fit for classify the cold damage extent.The above work provided an important foundation for developing on-line and fast detection equipment of seed maturation, cultivar authenticity and cold damage. |
PDF Full Text Request