基于sCARS-PSO-SVM的土壤硒含量高光谱定量反演

Hyperspectral Quantitative Inversion of Soil Selenium Content Based on sCARS-PSO-SVM

硒(Se)是人体必需的微量元素之一。人主要通过食用农产品来获取硒,而农产品中的硒主要来自土壤。因此,研究土壤中硒的含量和分布,对人体健康和农作物生产具有重要的意义。高光谱遥感技术的发展,使得高效、低成本、大范围估测土壤中硒的含量和分布成为可能。但是,土壤中硒的含量对光谱的敏感性较弱,严重影响了高光谱硒含量定量反演精度。该研究以广东连州地区富硒土壤为研究对象,系统采集研究区土壤样品50份,分析土壤样本硒含量,同步采集土壤反射光谱数据;利用Savitzky-Golay卷积平滑算法、多元散射校正(MSC)、对数一阶微分(lg(R)-FD)、标准正态变量校正(SNV)、多元散射校正一阶微分(MSC-FD)对原始光谱进行增强处理;应用稳定竞争自适应重加权采样(sCARS)算法结合皮尔逊相关性分析(PCC)进行特征波段选择;对比分析偏最小二乘(PLS)、支持向量机(SVM)和粒子群优化支持向量(PSO-SVM)模型土壤硒含量高光谱定量反演效果。结果表明:将sCARS算法应用于光谱增强后的回归模型,并结合皮尔逊相关性(PCC)选择与土壤硒含量敏感性较大的特征波段,不仅可以降低土壤硒含量高光谱预测模型复杂度并有效避免大量有用信息的损失,还能提高高光谱回归模型的反演效率;对比不同回归模型训练集和预测集的决定系数R^(2)和均方根误差RMSE,发现支持向量(SVM)模型比偏最小二乘(PLSR)模型预测效果更好,模型稳定性更高,且非线性模型更适用于土壤硒含量的预测;通过粒子群(PSO)算法优化SVM的核函数和正则化参数,SVM模型的反演精度和稳定性都有所提升;MSC-PSO-SVM模型(R^(2)=0.53、RMSE=0.34)和MSC-FD模型(R^(2)=0.50、RMSE=0.04)预测效果较为突出。综上所述:利用sCARS结合PSO-SVM算法建立土壤硒含量的高光谱定量反演模型,能够为土壤硒含量的高光谱大面积估测提供新的途径。

Selenium(Se)is part of the essential trace elements in the human body.People obtain selenium mainly through the consumption of agricultural products,and selenium in agricultural products mainly comes from the soil.Therefore,studying the content and distribution of selenium in the soil is very important to human health and crop production.However,the development of hypersensitive remote sensing technology has made it possible to estimate the content and distribution of selenium in soil in an efficient,low-cost and large-scale manner.However,the sensitivity of soil selenium content of spectra is weak,which seriously affects the accuracy of quantitative inversion of hypersecretion selenium content.In this study,50 soil samples were systematically collected from the study area to analyse the selenium content of the soil samples,and the soil reflection spectral data were collected simultaneously;the Savitzky-Golay convolutional smoothing algorithm,multiple scattering corrections(MSC),first-order logarithmic differentiation(lg(R)-FD),standard normal variance correction(SNV),multiple scattering corrected first-order differentiation(MSC-FD)for raw spectra enhancement;application of the stable competitive adaptive benighted sampling(sCARS)algorithm combined with Pearson correlation analysis(PCC)for feature band selection;comparative analysis of partial least squares(PLS),support vector machine(SVM)and particle swarm optimisation support vector(PSO-SVM)models for the quantification of soil selenium content in hypersecretion The results showed that the sCARS algorithm was applied to the inversion.The results show that applying the sCARS algorithm to the spectrally enhanced regression model and combining Pearson correlation(PCC)to select the feature bands with greater sensitivity to soil selenium content can not only reduce the complexity of the hypersecretion prediction model for soil selenium content and effectively avoid the loss of a large amount of useful information,but also improve the inversion efficiency of the hypersecretion regression model;comparing the training and prediction sets of different regression models The comparison of the coefficient of determination R2 and root mean square error RMSE between the training and prediction sets of different regression models showed that the support vector(SVM)model had better prediction results and higher model stability than the partial least squares(PLSR)model,and the non-linear model was more suitable for the prediction of soil selenium content;the inversion accuracy and stability of the SVM model were improved by optimizing the kernel function and regularization parameters of the SVM through the particle swarm(PSO)algorithm;the MSC-PSO-SVM model(R^(2)=0.53,RMSE=0.34)and MSC-FD model(R^(2)=0.50,RMSE=0.04)had more outstanding prediction results.In summary:the hypersensitive quantitative inversion model of soil selenium content using sCARS combined with the PSO-SVM algorithm can provide a new way to estimate the soil selenium content in a large hypersensitive area.