基于高光谱Hyperion数据的线性光谱混合模型与神经网络模型的比较

混合像元问题是定量遥感中的热点问题之一,为了改进从遥感数据中提取定量信息,人们建立了各种混合光谱分解技术,其中线性光谱混合模型和神经网络模型就是两种比较成熟的方法。以陕西省横山地区的高光谱Hyperion数据为研究基础,通过最小噪声变换(MNF)、像元纯度指数(PPI)转换和RMS误差分析的迭代方法相结合提取影像中的纯净像元作为终端端元。分别运用神经网络模型和线性光谱混合模型对影像进行光谱分解,得到各个组分的分解图像。以标准植被指数(NDVI)影像为衡量标准,选取训练样本点,分别对两种模型进行回归分析,结果显示NDVI影像与线性光谱混合模型植被分解图像的判定系数(R2=0.91)要大于其与神经网络模型的判定系数(R2=0.81)。进一步分析表明在一般情况下,线性光谱混合模型具有比神经网络模型略高的分离精度,但是神经网络模型对细部信息的提取的效果要好于线性光谱混合模型,最后提出了端元均方根误差(EAR)指数,一种新的混合像元分解的思路。

A comparison of linear spectral model and neutral net model based on Hyperion data

Mixture pixel, well known as the hot problem in remote sense, is the main obstacle of quantitative remote sensing.In order to improve the quality of extracting information from the remote sensing images, several spectral unmixing techniques have been developed over the past years.The linear spectral mixture model and the neutral net model are the two most widely applied techniques in spectral unmixing.Based on Hyperion image in Hengshan region of China, this paper draws a comparison between linear spectral mixture model and neutral net model by using NDVI image.With Minimum Noise Fraction (MNF), Pixel Purity Index (PPI) and an iterative unmixing method based on RMS error analysis, the pure end member from Hyperion data is extracted.In this way, people could not only obtain the "purest" pixel from image but also prevent to omit the end member spectral.After that, linear regressions of NDVI against linear spectral mixture and neutral net have been built up in order to analyze their relationship.The result suggests that the linear spectral mixture model has a higher coefficient of determination (R2=0.91) than the neutral net model (R2=0.81).With an exhaustive study, it is found that under generalized condition, the linear spectral mixture model has a higher unmixing accuracy than the neutral net model, while the neutral net has a better effect when extracting the discrete information.In addition, this paper brings forward a new measure for spectral umixing which is called end member average RMSE (EAR).