改进SVR的内陆水体COD高光谱遥感反演

Inland Water Chemical Oxygen Demand Estimation Based on Improved SVR for Hyperspectral Data

高光谱数据可以捕获内陆水体中不同浓度的化学需氧量(COD)引起的光谱变化,因此研究光谱反射率与COD浓度之间的关系对于COD的遥感估算至关重要。支持向量回归模型(SVR)具有适合小样本、泛化能力好的特点,基于SVR模型能够更加准确获得COD浓度和光谱数据之间的关系,但仍然存在参数选取困难和易陷入局部极值的问题。为了解决这个问题,将模拟退火—粒子群算法(SA-PSO)引入到支持向量回归机的参数优化过程中,提出了一种改进SVR(SA-PSO-SVR)的内陆水体COD高光谱遥感反演方法。以潍河流域为研究区域,通过野外测量获得了COD浓度和水表面光谱反射率。首先根据光谱反射率对COD的响应来确定敏感因子,把SA-PSO算法引入SVR的参数优化过程中建立了COD浓度与敏感因子之间的反演模型。最后利用珠海一号高光谱数据验证模型的准确性,进而获得了COD浓度的分布情况。通过光谱分析,可知该区域实测的水面光谱具有典型的二类水体特征,光谱曲线形状呈现明显的双峰特征,当浓度增加时,反射峰具有向短波长方向移动而反射谷向长波长方向移动的趋势。通过计算Pearson相关系数分析COD浓度和光谱之间的相关性,结果表明最佳的反演因子为518 nm/940.4 nm, 623.6 nm/636.8 nm, 729.2 nm/890.9 nm和752.3 nm/857.9 nm的四个波段比值组合;经过SA-PSO-SVR方法建立的COD估计模型的平均相对误差(MRE)和均方根误差(RMSE)分别为1.62%和2.99 mg·L-1(R2=0.86),反演结果优于其他模型(SVR、 BP神经网络和线性回归模型)。将实测水面光谱建立的最优模型应用于高光谱卫星影像上,RMSE和MRE分别为4.47 mg·L-1和11.87%。获得的潍河-峡山水库区域的COD反演结果显示:COD的整体浓度介于17~42 mg·L-1之间,韩信坝、峡山水库的东北部、渠河注入潍河的交汇处等区域的COD浓度高于其他水域。证实了SA-PSO-SVR是一种有效的COD高光谱反演方法,可供潍河流域水资源管理提供参考。

Hyperspectral data can capture the spectral changes caused by different concentrations of chemical oxygen demand(COD) in inland water bodies, and it is important to study the relationship between spectrum and COD concentrations for COD estimation. Support Vector Regression(SVR) model has the advantages of being suitable for small samples and good generalization ability, but it is difficult to select a parameter and prone to fall into the local extremum. In order to solve this problem, this study introduced Simulated Annealing-Particle Swarm Optimization(SA-PSO) into the parameter optimization process of SVR and proposed an improved SVR(SA-PSO-SVR) method to estimate the inland waters COD. This paper takes the Weihe River Basin as the research area, obtained the COD concentrations and spectral curves through field measurement. The sensitive band was determined by analyzing the response of spectral reflectance to COD at first in this paper, and the Simulated Annealing-Particle Swarm Optimization(SA-PSO) was introduced into the parameter optimization process of Support Vector Regression(SVR) to established an inversion model between the cod concentration and the sensitivity factor. The Orbita Hyper Spectral(OHS) hyperspectral data was used to verify the accuracy, and the distribution of COD concentration was obtained at last. Through spectral analysis, it can be seen that the measured above-surface spectra in this area demonstrated typical spectral signatures of second-class water, and the shape of the spectrum curve shows obvious double-peak characteristics. When the concentration increases, the reflection peak tends to move to the short wavelength direction and the reflection valley to the long wavelength direction. The Pearson’s correlation coefficient was used to analyze COD concentration and the spectral, the result showed that the best inversion factors are four band combinations of 518 nm/940.4 nm, 623.6 nm/636.8 nm, 729.2 nm/890.9 nm and 752.3 nm/857.9 nm. The model established by the SA-PSO-SVR method is accurate compared with models established by SVR, Back Propagation neural network, and linear regression method. The Mean-Relative-Error(MRE) and Root-Mean-Square-Error(RMSE) of the COD estimation model established by the SA-PSO-SVR method are 1.62% and 2.99 mg·L-1(R2=0.86), respectively. The optimal model established by the measured water surface spectra was applied to the hyperspectral satellite image. The RMSE and MRE are 4.47 mg·L-1 and 11.87% respectively. The obtained COD inversion results of the Weihe-Xiashan reservoir area show: the overall concentration of COD is between 17 and 42 mg·L-1, and COD concentration in the Hanxinba, the northeast region of XiaShan Reservoir, the confluence of the Qu River into the Wei River are higher than other waters. The experimental results show that SA-PSO-SVR is a feasible approach for the COD inversion of hyperspectral data, providing a reference for water resources management in the Weihe River Basin.