| Maize is one of the world’s three major crops and is one of the most important food and feed crop in China. The maize’s yield and quality is directly related to the utilization rate of maize, the grade of goods and economic benefits. Therefore, it has a scientific and practical value to find a new method for automatic extracting the maize yield-related traits and determination of main component content in maize simultaneously.This article focused on the digital extraction methods for maize yield-related traits and design of a prototype system for extracting ear length, ear width, ear weight, spike rows, line grain number, cob weight, grain length, grain width, grain thickness, grain colors and100grains weight. The methods of measuring maize amino acid, protein and amylose based on near-infrared spectroscopy(NIR) were studied and the models for measuring maize amino acid, protein and amylase were built. The main results are as follows:1) The automatic extraction system of maize yield-related traits were designed based on machine vision, including conveying module, PLC control module, image acquisition module, serial communication system between PLC and PC.2) The performance of system for automatic extracting maize yield-related traits were studied and the accuracy of the system was verified.200maize ears and200grains were measured with the system and the results showed that the ear length relative errors of dent maize were all within5%and the maize ear width relative errors of91.2%dent maizes were within5%. For durum maize ear,86.1%of the maize ear length relative errors were within5%and97.4%of maize ear width relative errors were within5%.80.61%of maize spike rows relative errors were within10%and82%of maize line grain number were within10%.3) The different methods of eliminating abnormal samples for measuring maize grain components content were studied based on near infrared spectroscopy(NIR). The leverage value method, resampling by half-mean(RHM) and monte carlo sampling (MCS) were used to eliminate abnormal samples from amino acid, protein and amylose in maize grains and their PLS models were built, respectively. Experimental results showed that the RHM was the optimum method for eliminating outlier for measuring amino acid in maize grains, leverage value method for protein content in maize grains and MCS for amylose content in maize grain.4) The different sample set partitioning methods and modeling results were studied based on NIR and sample set partitioning based on joint X-Y distance algorithm (SPXY) was chosen as the optimal method. Random method (RS), Kennard-Stone (KS) and SPXY were used to divide maize grain samples into calibration set and validation set, respectively. The partial least squares (PLS) models were established for measuring amino acid, protein and amylose in maize grains. The results showed that SPXY was the optimal method for sample set partitioning for measuring maize component contents.5) Different parameter optimization methods for partial least squares support vector machine(LS-SVM) modeling were studied. Niche ant colony algorithm(NACA) and grid search were used to optimize the parameters y and a2of LS-SVM, respectively. The results showed that using NACA optimization method could improve the accuracy of LS-SVM model and optimization speed.6) The influence of different preprocessing methods and modeling method was compared for3component models of maize grains. Based on SPXY, models by PLS and LS-SVM were built. The influence of different spectral pretreatments on model performance was analyzed. The results showed that the performance of LS-SVM for measuring maize grain amino acid, protein and amylose content was much better than PLS. The performance of LS-SVM with multiplicative scatter correction (MSC) plus orthogonal signal correction (OSC) and plus autoscale was optimal for maize grain amino acid which the correlation coefficient R of validation set is0.997and root mean square error of prediction (RMSEP) is0.019. The performance of LS-SVM with orthogonal signal correction (OSC) plus autoscale was optimal for maize grain protein which the correlation coefficient R of validation set was0.999and root mean square error of prediction (RMSEP) was0.019. The performance of LS-SVM with detrend plus mean center(MC) was optimal for maize grain amylose which the correlation coefficient R of validation set was0.999and the RMSEP was0.068.7) Different variable selection methods for3component models were studed based on NIR. The models were built in the full spectrum (4000～10000cm-1), the combinaiton region (4000～5500cm-1), the first overtone region (5500～7000cm-1), the second overtone region(7000～10000cm-1), and the wavelengths selected by correlation coefficient method or genetic algorithm(GA). The results showed that the performance of models built with the wavelengths selected by GA were not good. The performance of LS-SVM models built in different regions were better than PLS models. |
PDF Full Text Request