中文摘要：

高光谱数据波段多、波段之间相关性强,导致信息冗余严重,增加了数据处理的工作量,有效准确地在众多波段中选择具有代表性的波段尤为重要。首先用Wilks＇Lambda（WL）,随机森林（random forest,RF）与自适应波段选择（adaptive band selection,ABS）这3种方法对高光谱数据进行降维处理。然后提出了基于曲线误差指数的评价方法,用此指数的趋势来确定每种降维方法所要选择的合适波段数量,同时用指数的大小评价不同降维方法的优劣,并用分类方法对评价结果加以验证。结果显示：Wilks＇Lambda最终选择的波段数为10个,α6-α平稳值（选择6个波段时的曲线误差值与曲线误差平稳值之间的差值）为0.05;随机森林最终选择的波段数为13个,α6-α平稳值为0.06;自适应波段选择方法最终选择的波段数为20个,α6-α平稳为0.14。Wilks＇Lambda的总体分类精度为80.56%,Kappa系数为0.77;随机森林的总体分类精度为79.11%,Kappa系数为0.76;自适应波段选择方法的总体分类精度为49.94%,Kappa系数为0.41。得出以下结论：（1）基于曲线误差指数的方法得出Wilks＇Lambda有最小的α6-α平稳值,是最佳的波段降维方法 ;分类结果显示：Wilks＇Lambda有最大的总体分类精度与Kappa系数,是最佳的波段降维方法。（2）基于曲线误差指数的评价方法与基于分类结果的误差一致,说明此方法具有可行性。

英文摘要：

Hyperspectral datasets have been widely used in monitoring analyses of vegetation due to their abundant spectral information;however,their spectral resolution greatly increases information redundancy causing more data processing work.To find an efficient method for selecting the most representative bands to reduce redundancy before using them,three traditional dimension reduction methods,namely Wilks＇ Lambda（WL）,random forest（RF）,and adaptive band selection（ABS）,were used to select optimal bands among the64 bands.Then,a new evaluation method based on a curve error index was proposed to determine the appropriate number of bands through analysis trend of the index values and to select the best dimension reduction method.Lastly,classification results were used to demonstrate the validity of this method.Results indicated that WL selected 10 bands and the α6-αsmoothvalue（the curve error difference between selecting six bands and when the curve becomes stable） was 0.05.The RF method selected 13 bands and its corresponding α6-αsmoothvalue was 0.06.The ABS method selected 20 bands and its α6-αsmoothvalue was 0.14.The overall accuracy of WL based band selection was 80.56% with a Kappa coefficient of 0.77;RF was 79.11% with a Kappa coefficient of0.76;and ABS had an overall accuracy of 49.94% and a Kappa coefficient of 0.41.Thus,（1） with the curve error index method WL presented the smallest α6-αsmoothvalue and had the highest overall classification accuracy,suggesting WL was the optimal method for dimensional reduction,and（2） the two evaluation methods had the same results,illustrating the curve was feasible.