基于光谱变换的高光谱指数土壤盐分反演模型优选

摘要

该文探索基于光谱变换建立光谱指数,进而建立土壤盐分反演模型的可行性.运用倒数、导数、对数等15种光谱变换对土壤含盐量进行反演,并利用原始光谱的波段反射率构造光谱指数对土壤盐分进行建模.在15种高光谱变换中,一阶微分R?和一阶对倒数(log1/R?)变换下土壤盐分估算模型的精度较高.但总体而言,基于单一光谱变换和光谱指数的模型模拟精度均较低.采用光谱变换建立光谱指数,并进一步建立土壤盐分反演模型,结果表明,基于(log1/R?)光谱变换构建归一化植被指数,然后建立的土壤盐分精度最高,经验证,其R2为0.89,均方根误差为3.34 g/kg,高于单一方法构建的模型,可为半干旱地区土壤盐分反演提供参考.%At present, scholars at home and abroad already use the methods like spectral index or spectral transformation and so on to invert soil salinity separately. However, it is rare to study comprehensive modeling of soil salinity based on spectral index derived from different spectral transformations. In this paper, we studied the feasibility of establishing soil salinity model based on spectral index derived from spectral transformations. The study area was Lake Ebinur wetland nature reserve. Soil samples were collected in July and August in 2016 from 32 representative points. The intervals of sampling points were 3-10 km. The hyperspectral band reflectance of the sample was obtained by the ASD spectrometer. The positions of sampling points were recorded by a handy GPS. The reflectance curves were pretreated with mean value treatment, signal denoising and smoothness. Then, we used 15 kinds of spectral transformations, such as reciprocal, derivative, logarithm and so on. Meanwhile, the band reflectance of original spectrum was used to construct the spectral index to model the soil salinity. The spectral index included the difference soil index (DSI), simple ratio soil index(RSI) and normalized difference soil index (NDSI). On the basis of the modeling of soil salinity under a single spectral transformation or spectral index, we tried to establish the hyperspectral matrix coefficient map of soil salinity and the best spectral transformation of reflectance. Then a new hyperspectral estimation model was built in order to improve the estimation accuracy of soil salinity model. The model construction was based on randomly selected 22 samples and the model validation was based on the 10 samples. The mean of soil salinity for the calibration dataset and the validation dataset was 8.33 and 8.44 g/kg, respectively. The coefficient of variation of soil salinity for the calibration dataset and the validation dataset was 60.53％ and 61.15％, respectively. The results showed that among the 15 spectral transformations, the correlations between the reflectance under the 6 transformations of first-order derivative (R'), second-order derivative (R'), first-order derivative of reciprocal (1/R)', first-order derivative of logarithm (logR)', first-order derivative of logarithm of derivative (log1/R)' and that after removal of contour line (Rcr)and soil salinity were the best. The numbers of characteristic bands were large, and the accuracy of soil salinity was better. The precision of soil salinity estimation model based onR' and (log1/R)' transformation was the highest among all the models although the model accuracy was generally low. Among the models of soil salinity based on the spectral index, the linear regression model based on DSI index had the highest fitting accuracy with R2 of 0.68 and root mean square error of 4.54 g/kg. However, the model accuracy based on a single spectral transformation or spectral index was still low. We constructed the spectral index based on the hyperspectral transformation of reflectance. Then, the soil salinity based on the spectral index derived from the transformation was established by the linear or quadratic regression methods and the accuracy was relatively improved. Among the 6 models, the model based on NDSI with (log1/R)' transformation was the best withR2 of 0.76 and root mean square error of 4.28 g/kg. The validation showed thatR2 was 0.89, and root mean square error was 3.34 g/kg. The proposed model was useful in improving soil salinity inversion accuracy. The study provides an efficient method for the quantitative estimation, inversion and detection of soil salinity in arid and semi-arid regions.