土壤高光谱噪声过滤评价研究

Quantitative Evaluation of Soil Hyperspectra Denoising with Different Filters

以ASD FieldSpecProFR测试的土壤高光谱数据为研究对象，探讨光谱噪声分布以及不同滤波器去噪效果定量评价。土壤高光谱曲线目视及其包络线去除、一阶微分和高通滤波分析表明，除起始波段350nrn最前端40nm范围内有部分噪声外，其他UV／vNIR（350-1050nm）范围内基本不存在噪声，而整个SWIR（1000-2500nm）范围内存在一定的噪声，SWIR2（1800-2500nm）后半段噪声较大，并且组成光谱仪的3台分光计在相互结合处，噪声比邻近谱段更大。采用六种不同滤波器进行噪声去除，通过构建光谱平滑指数（SI）、横向特征保持指数（HFRI）和纵向特征保持指数（VFRI）进行定量评价各滤波器去噪能力，评价结果对比发现WD和MA既能达到曲线平滑又能较好地保持波段特征。此外在PLSR模型下，以66个土壤样本的光谱为例，将六种滤波器去噪后的一阶微分光谱作为模型输入，比较分析不同滤波器对砂粒含量预测精度影响，精度对比表明，相比滤波器的特征保持能力，平滑能力更能影响砂粒含量精度。本研究为开展光谱预处理和光谱分析技术提供有益探索，并能为光谱学相关应用提供了科学依据。

The noise distribution of soil hyperspectra measured by ASD FieldSpec Pro FR was described, and then the quantita- tive evaluation of spectral denoising with six filters was compared. From the interpretation of soil hyperspectra, the continuum removed, first-order differential and high frequency curves, the UV/VNIR （350-1050 nm） exhibit hardly noise except the coverage of 40 nm in the beginning 350 nm. However, the SWIR （1 000-2 500 nm） shows different noise distribution. Especially, the latter half of SWIR 2（1 800-2 500 nm） showed more noise, and the intersection spectrum of three spectrometers has more noise than the neighbor spectrum. Six filters were chosen for spectral denoising. The smoothing indexes （SI）, horizontal feature reservation index （HFRI） and vertical feature reservation index （VFRI） were designed for evaluating the denoising performance of these filters. The comparison of their indexes shows that WD and MA filters are the optimal choice to filter the noise, in terms of balancing the contradiction between the smoothing and feature reservation ability. Furthermore the first-order differential data of 66 denoising soil spectra by 6 filters were respectively used as the input of the same PLSR model to predict the sand content. The different prediction accuracies caused by the different filters show that compared to the feature reservation ability, the filter＇s smoothing ability is the principal factor to influence the accuracy. The study can benefit the spectral preprocessing and analyzing, and also provide the scientific foundation for the related spectroscopy applications.