基于PCA的土壤Cd含量高光谱反演模型对比研究

A Comparative Study of the Hyperspectral Inversion Models Based on the PCA for Retrieving the Cd Content in the Soil

土壤重金属污染对人类健康造成了极大的威胁,如何快速摸清土壤污染情况尤为重要。高光谱遥感具备光谱分辨率高,快速无损等优势,使其在土壤组分反演方面具有巨大的潜力。针对高光谱信息冗余及光谱变换对土壤镉(Cd)含量估算的影响进行分析,并利用变换前后的高光谱数据对比研究了不同高光谱模型对土壤Cd含量反演的性能。首先利用等离子体质谱法和FieldSpec4地物光谱仪收集了56组土壤样品的Cd含量和对应的高光谱曲线(350~2500 nm);为了弱化光谱测定中光亮变化和土壤表面凹凸对实验结果的影响,研究对高光谱数据进行倒数对数预处理;考虑到高光谱数据中存在大量的信息冗余,研究采用了主成分分析(PCA)对高光谱数据进行降维处理并最终保留了前12个主成分量作为特征变量。针对高光谱反演模型,研究选择了偏最小二乘(PLSR)、支持向量机(SVM)、人工神经网络(ANN)和随机森林(RF)四种回归模型建立PCA主成分与Cd含量之间的关系;最后,研究选取了决定系数(R^(2))、均方根误差(RMS E)和RPD三种精度评估指标评估回归模型的拟合精度,结果表明针对光谱采用PCA波段降维的方法处理后,选取的12个主成分对变化前后的光谱累计贡献率均达到99.99%,作为模型的输入变量,四种模型均具有一定的预测能力。无论光谱变换与否,PCA-RF反演模型的预测能力均为最好(R^(2)分别为0.856和0.855,RPD均高达3.39)。利用PCA对高光谱数据降维处理可以有效降低高光谱数据冗余,有力的保证模型的预测能力。以PCA筛选出的主成分量可以作为模型极好的输入变量,以RF为基础的高光谱反演模型在反演土壤Cd含量时具有最佳效果,可为该区域及类似地区的土壤重金属污染物反演提供新的方法支撑。

The soil heavy metal pollution poses a great threat to the human health,thus,it is quite important make out the contamination in the soil.There are a series of advantages in the hyperspectral remote sensing technology,such as the high spectral resolution,rapid response,non-destructive,etc.,making it a well-suited in retrieving the soil’s components.In this study,the impacts of the information redundancy in the spectral and spectral transformation on the inversion of Cd content in the soil are investigated.Further,based on the hyperspectral data before and after spectral transformation,the performance comparations of hyperspectral models are carried out in this paper,as well.By so doing,the Cd contents and the corresponding lab spectrum(350~2500 nm)of 56 soil samples are measured by the ICP-MS and ASD Fieldspec4.Then,the reciprocal and logarithm changes are performed to weaken the impacts of the light variation and soil surface roughness on the experimental results.Due to the fact that there is much redundant information in the obtained data,the Principal Component Analysis(PCA)is carried out to reduce the dimensionality of the spectral bands in the data.After this processing,only 12 principal components are selected as the input variables of the model.Regarding the hyperspectral models,the Partial Least-Squares Regression(PLSR),Support Vector Machine(SVM),Artificial Neural Network(ANN)and Random Forest(RF)are chosen to establish the relationship between the Cd content and PCA components.Finally,for evaluating the prediction capabilities of the regression models,three precision evaluation indexes are preferred to assess the accuracy of regression models in this study,they are the correlation coefficient(R^(2)),Root Mean Squared Error(RMSE)and Residual Predictive Deviation(RPD).Analysis results show that the cumulative contribution rate of 12 principal components of the original data after processed by the PCA can be up to 99.99%.Using principal components as the inputs,all four hyperspectral models show excellent performances in predicting the Cd content in the soil.The PCA-RF,in particular,has the most accurate prediction capability regardless of whether the spectral transformation is performed or not(whose R^(2) before and after spectral transformation are 0.856 and 0.855,respectively,while the RPD under both conditions are 3.39).In conclusion,the PCA is used to reduce hyperspectral data’s dimensionality,this processing can effectively reduce the redundancy of hyperspectral data and guarantee the predictive capability of hyperspectral models.Also,the principal component selected by the PCA method could be excellent input variables of the hyperspectral models.Further,the hyperspectral model based on the PCA-RF shows the most excellent performance for rapid detecting the Cd element in the soil within the study area and similar regions,which could be a new supplement for the inversion of heavy metals in the soil.