高光谱数据降维与水稻氮素含量解析方法

Dimension Reduction of Hyperspectral Data and Analysis of Rice Nitrogen Content

水稻营养监测、病虫害诊断中高光谱技术提供了有效、便捷的技术手段,高光谱数据的降维和特征提取至关重要。为了探索有效的高光谱降维方法,利用2019年沈阳市沈北试验基地水稻分蘖期、拔节期和抽穗期的叶片光谱数据及实测的氮素含量数据,首先分窗口对原始高光谱进行Gram-Schmidt变换找到投影空间并映射出主基底,保留具有显著性概率的主基底,其极大极小值为特征波段,实现高光谱数据降维与特征波段提取;其次利用降维后构建特征建立水稻叶片氮素含量估测模型;最后对比分析本研究方法与主成分分析、植被指数及特征参数等降维方法。结果表明:基于分窗Gram-Schmidt变换可有效实现高光谱数据的降维,高光谱降到4维,水稻叶片氮素含量估测模型5折交叉验证决定系数CV_R2为0.787,均方误差CV_MSE为0.051;基于主成分分析法降维,高光谱降到5维,CV_R2为0.743,CV_MSE为0.056;单个植被指数,效果最好的是GNDI(911,487),CV_R2为0.667,CV_MSE为0.076;多个特征光谱指数最优化选择7个特征,分别为SI(487),RI(601,487),DI(911,487),DDI(990,685,487),NDI(990,685),NDI(990,601)和GNDI(911,487),CV_R2为0.731,CV_MSE为0.072。说明采用全波段降维的方法优于植被指数及特征降维,并且本研究提出的分窗Gram-Schmidt变换在降维的同时也能获取包含主要信息的波段,优于主成分分析法。综上所述,基于分窗Gram-Schmidt变换的高光谱可为高光谱降维与水稻叶片氮素含量估测提供一定的理论基础和技术支撑。

Hyperspectral technology provides an effective and convenient technical means in rice nutrition monitoring and disease and insect diagnosis.The dimension reduction and feature extraction of hyperspectral data are very important.In order to explore an effective method for hyperspectral dimension reduction,the leaf spectral data and the measured nitrogen content data were carried out in three growing stages of rice,i.e.tillering stage,jointing stage and heading stage in Shenyang Shenbei experimental base in 2019.Firstly,the original hyperspectral data were divided into windows,and the projection space was found by the gram Schmidt transform,and the main substrate was mapped out,and the main substrate with significant probability is retained.The maximum and minimum values of the main substrate were the feature bands,which realizes the dimension reduction and feature band extraction of hyperspectral data.Secondly,the estimation model of nitrogen content in rice leaves was established by using the dimension reduction features.Finally,the research method and the main substrate were compared and analyzed Component analysis,vegetation index and characteristic parameters.The results showed that the transformation could effectively reduce the dimension of hyperspectral data based on window gram Schmidt.The dimensionality of spectral data was reduced to 4 dimensions,the cross validation coefficient CV_R2 was 0.787,and the mean square error CV_MSE was 0.051.Based on the principal component analysis method,the hyperspectral dimension was reduced to 5 dimensions,CV_R2 was 0.743,CV_MSE was 0.056;the best effect of single vegetation index was GNDI(911,487),CV_R2 was 0.667,CV_MSE was 0.076;multiple characteristic spectral indices selected seven features,namely SI(487),RI(601,487),DI(911,487),DDI(990,685,487),NDI(990,685),NDI(990,601)and GNDI(911,487),CV_R2 was 0.731,CV_MSE was 0.072.It showed that the method of full band dimensionality reduction was better than that of vegetation index and feature dimensionality reduction,and the windowed gram Schmidt transform proposed in this study could obtain the band containing the main information while dimensionality reduction,which is better than the principal component analysis method.In conclusion,the hyperspectral data based on the window gram Schmidt transform could provide theoretical basis and technical support for hyperspectral dimensionality reduction and nitrogen content estimation of rice leaves.