植被参数高光谱遥感反演最佳波段提取算法的改进

摘要

Hyperspectral data involve a huge amount of information, and how to select the best combination band to build high-accuracy spectrum model, is the key work for the developing of remote sensing inversion model of vegetation parameter. Maximum correlation coefficient (MCC) is a method of extracting combination bands by using the correlation coefficient between vegetation parameter and hyperspectral band as the selection index of feature combination band. Due to simple calculation and easier operation, MCC iswidely used in dimensionality reduction and extracting the useful bands of hyperspectral data. However, the method only considers to extract the maximum correlation bands of vegetation parameter variable, which may lead to overlook the indirect effect of other factors. Optimum index factor (OIF) is the extraction method of feature combination bands based on the basic idea that the amount of information is proportional to the sum of mean square deviation, and inversely proportional to the sum of correlation coefficient for each band. Although OIF can obtain feature combination bands of abundant information and small redundancy, the highest correlation between vegetation parameter and the selected band can’t be ensured, which may cause the decreasing of estimation ability of the model which is built by the extraction bands. Obviously, OIF method and MCC method in the extraction of feature combination bands have complementary advantages. In this study, by using the evaluation method with entropy coefficients, OIF and MCC are endowed with objective weight respectively, and based on this the optimal combination band is obtained, which is named as optimum index factor and correlation coefficient (OIFC). The feature combination bands extracted by OIFC, have the highest correlation with the corresponding vegetation parameters, as well as with ensured rich information. In order to demonstrate OIFC’s practicality, feature combination bands of winter wheat leaf chlorophyll which was extracted by OIFC were given as an example. The 760, 1860 and 1970 nm were considered as feature combination bands and selected by OIFC. Then, hyperspectral model of wheat chlorophyll content was built by partial least squares. Compared with the models from vegetation indices such as normalized difference vegetation index (NDVI) and soil-adjusted vegetation index (SAVI) and OIF method, the precision of the model built by OIFC was the highest and the coefficients of determination was 0.7390. The determination coefficient of linear fitting between predicted values by OIFC and measured chlorophyll contents was 0.827, and the root mean square error was 5.440. The results show that the extracted bands of winter wheat chlorophyll based on OIFC has higher modeling precision. It also proves that OIFC can reliably and effectively extract feature combination bands of vegetation parameters from hyperspectral data. The method of OIFC can also provide theoretical basis and technical support for further improving the accuracy of hyperspectral estimation model of physical and chemical parameters of vegetation.%高光谱信息量巨大，如何选取最佳组合波段构建高精度光谱模型，是植被参数遥感反演模型研究的重要工作基础。该研究将最佳指数与相关系数通过熵权评价值进行融合，提出最佳指数－相关系数法（optimum index factor and correlation coefficient，OIFC）。基于OIFC法选取了小麦叶片叶绿素含量的最佳组合波段，并利用最佳组合波段的高光谱数据建立小麦叶片叶绿素含量预测模型。结果表明：利用OIFC法所提取的小麦叶绿素最佳组合波段是760、1860、1970 nm；对比最佳指数法（optimum index factor，OIF）、最大相关系数法（maximum correlation coefficient，MCC）提取波段以及归一化植被指数（normalized difference vegetation index，NDVI）、土壤调和植被指数（soil-adjusted vegetation index，SAVI）所建立的叶片叶绿素含量高光谱模型，基于OIFC法构建的模型预测值与实测值具有显著的线性关系，决定系数达0.827，且均方根误差最小（RMSE=5.44）。可见，基于 OIFC 法构建的小麦叶绿素含量模型具有更高的精度，该结果验证了利用OIFC法提取高光谱特征波段的可行性，并且能够获得更高建模精度的特征波段。