为了探讨基于高光谱的苎麻品种识别和分类的方法， 在大田栽培条件下， 采集了4个不同基因型苎麻品种共927个叶片高光谱数据。 根据苎麻叶片高光谱反射曲线， 选择了2组特征参数： 基于高光谱波形峰谷反射率和位置参数（V1组）、 基于偏度和峰度参数（V2组）。 运用逐步判别的方法， 通过设置不同F值筛选不同个数的变量， 分别建立基于2组特征参数的多个Fisher线性判别函数， 并从计算量、 正确率和稳定性三方面对所建立的判别函数进行分析比较。 结论： (1)所有组合的判别函数总体平均正确率为911%， 标准差总体均值为12%； (2)综合权衡， 在所有组合中， V2组且14≥变量个数n≥8判别效果最好——计算量中等， 正确率和稳定性均高于平均值， 其中， 13个变量的Fisher判定函数平均正确率最高有942%， 标准差最低为0%； (3)若优先考虑正确率， V1组且22≥变量个数≥15正确率最高， 平均正确率最大有955%， 但计算量比较大， 稳定性中等， 标准差最低为09%。 研究表明， 利用高光谱参数结合逐步判别方法识别苎麻品种是可行的。

The hyperspectral data on total 927 leaves of different genotypes, which come from 4 ramie varieties, were collected under the field cultivation conditions to explore the identification and classification of ramie varieties with the hyperspectral as the basis. According to the hyperspectral reflection curve of ramie leaves, two groups of feature parameters were extracted, namely, the hyperspectral wave-valley reflectance and position parameters(the group V1) as well as the skewness and kurtosis parameters(the group V2) . Then, by adopting the stepwise discriminant approach to screen different number of variables under different F-value settings, multiple Fisher linear discriminant functions based on these two groups of feature parameters were created respectively, and further, the created discriminant functions were comparatively analyzed from the computational complexity, the accuracy and the stability. So, we can come to the following conclusions: (1) For discriminant functions under all the combinations, the overall average accuracy was 911% and the overall standard deviation mean was 12%; (2) From the comprehensive trade-off perspective, when the number of variables was between 8 and 14, the discriminant effect of the group V2 was the best among all the combinations, namely, the computational complexity was in the middle level, and both the accuracy and the stability were over their corresponding average values; among them, the discriminant functions with 13 variables had the highest average accuracy and the lowest standard deviation, which were 942% and 0% respectively; (3) When the accuracy was considered preferentially and the number of variables changed between 15 and 22, the group V1 had the highest accuracy which was 955%, however, the computational complexity under this case was higher, the stability was in the middle level and the lowest standard deviation was 09%. Above results showed that it was feasible to utilize the hyperspectral parameters together with the stepwise discriminant approach.