闲鹤野云
第1楼2009/06/09
英文摘要
From five paddy rice cultivars Tainung Sen 20, Taichung Sen 10, Tainung 67, Taikeng 8, and Taikeng 9 grown in central, eastern and south Taiwan, and harvested in the summers of 1997, 1998, and 1999, calibrated models were established by discriminant analysis and backpropagation neural network program through near infrared absorbance and external morphological and color features selection. The calibrated models were used to classify the above mentioned five paddy rice cultivars harvested in the same area in the summer of 2000.
From 1100 nm to 2500 nm in 3-nm steps, the reflectance spectrum absorbance was collected as a variable. Totally three hundred fifty-one variables were used to develop the discriminant analysis and backpropagation neural network models, and the average classification rates were 98.1% and 92.5%, respectively. Sixty-nine variables were selected by using stepwise discrimination to develop the discriminant analysis and backpropagation neural network models, and the average classification rates were 98.5% and 85.5%, respectively. Sixty-nine variables were selected by using the correlation matrix to develop the discriminant analysis and backpropagation neural network models, and the average classification rates were 72.0% and 72.3%, respectively. Sixty-nine variables were selected by loading on the first and second principal components to develop the discriminant analysis and backpropagation neural network models, and the average classification rates were 69.1% and 60.6%, respectively. In selecting wavelength of near infrared spectroscopy for establishing models, the classification rate by stepwise discrimination method was superior to the results by correlation matrix and loading value method. Reducing 351 variables to 69, the same classification rate was still kept. Using the same variables the classification by discrimation method was better than that by backpropagation neural network.
Using morphological and color features of five paddy rice cultivars, there were 60 variables including single kernel area, perimeter, shape factor (4π×area/perimeter2), area/perimeter, maximum width, maximum length, maximum length/ maximum width, average intensities of red, green, and blue, and 50 widths on the maximum length. The following models were trained by backpropagation neural network program to establish classification models. With 60 features, the average classification rate of Model 1 was 92%. With the most effective 50 features, by loading in the first principal component, the average classification rate of Model 2 was 90.0%. With 35 features selected from the correlation coefficient matrix, the average classification rate of Model 3 was 91.0%. With the most effective 20 features of area, area/perimeter, 48th width, shape factor, maximum length/maximum width, average intensity of blue, maximum length, average intensity of green, 47th width, 50th width, average intensity of red ,1st width, 19th width, 5th width, 6th width, 29th width, perimeter, 46th width, 42nd width, and 4th width based on the contribution of the training model, the average classification rate of Model 4 was 91.8% and would be recommended for classifying five paddy rice cultivars of set trading prices because it required fewer features and held a stable classification rate.
17 variables from 60 morphological and color features and 54 variables from 351 near infrared reflectance absorbance were selected by stepwise discriminant method. Totally 71 variables were input into backpropagation neural network to establish classification model. Selecting 71 variables according to loading value on the first principal component from high to low, the number of variables were used from 10 to 71 with 5 steps were input backpropagation artificial neural network program to establish Model 1. The average validation rates of model were 76.2%, 82.4%, 92.9%, 92.7%, 95.5%, 95.8%, 96.5%, 94.1%, 95.5%, 95.5%,95.5%,95.5%,and 96.5%, respectively, and the average classification rates were 78.0%, 78.5%, 83.5%, 51.9%, 66.8%, 51.2%, 77.1%, 73.7%, 59.0%, 82.0%, 83.2%, 78.6%, and 79.3%, respectively. With 71 variables arranged according to correlation coefficient from low to high, and with the number of variables from 10 to 71 with 5 steps to establish Model 2, the average validation rates were 83.3%, 83.5%, 83.5%, 93.9%, 93.4%, 96.2%, 96.0%, 96.7%, 97.2%, 95.8%, 97.2%, 96.7%, and 96.2%, respectively, and the average classification rates were 87.0%, 85.4%, 76.4%, 70.4%, 54.4%, 45.0%, 49.8%, 55.3%, 62.0%, 62.0%, 61.9%, 76.8%, and 79.0%, respectively. According to contribute to backpropagation neural network model from high to low, and the number of variables from 10 to 71 with 5 steps were used to establish Model 3. The average validation rates were 85.2%, 91.5%, 91.1%, 92.5%, 94.1%, 95.3%, 96.2%, 95.5%, 96.0%, 95.5%, 96.0%, 96.0%, and 96.2%, respectively, and the average classification rates were 65.1%, 66.3%, 64.9%, 62.0%, 74.7%, 70.8%, 66.6%, 76.4%, 78.8%, 78.5%, 72.5%, 71.8%, and 85.1%, respectively. Collecting image features from single kernel and average 10 kernel data, and scanning the near-infrared absorbance in a vessel (8 g, about 90-100 kernels), two different data treatment method were combined to establish models, it may exist some interference and reduce the classification rates only by image features or near infrared absorbance. Further studying scanning the near-infrared absorbance in single kernel see whether can improve the classification rates.