基于偏最小二乘与随机森林的土壤盐含量反演研究

Research on Inversion of Soil Salt Content Based on Partial Least Squares Combined with Random Forest

针对土默川平原地区的土壤盐分含量提出了偏最小二乘与随机森林相结合(RF-PLSR、PLSR-RF)对土壤盐分含量进行预测的回归反演模型。该研究共采集45份土壤样本,随机选取35份为建模集,10份为验证集。试验首先对采集到的高光谱土壤图像进行分割处理提取出土壤在400~1000 nm的原始反射光谱,其次对原始反射光谱进行4种光谱变换(一阶微分、多元散射校正的一阶微分、SG平滑去噪的一阶微分、对数的一阶微分),并与土壤的实测盐分量进行相关性分析(CA),利用相关系数选取敏感波段,最后建立偏最小二乘与随机森林结合的回归反演模型。结果表明,与偏最小二乘回归、随机森林回归单独建模相比,2种模型结合后的预测精度有明显的改善。光谱经过对数的一阶微分变换建立的PLSR-RF反演模型更为明显,其建模集决定系数R_(c)^(2)为0.852,均方根误差RMSE_(c)为0.102 g/kg,相对分析误差RPD_(c)为2.600,验证集决定系数R_(v)^(2)为0.941,均方根误差RMSE_(v)为0.049 g/kg,相对分析误差RPD_(v)为4.117。

Aiming at the soil salt content in the Tumochuan Plain,a regression inversion model combining partial least squares and random forest(RF-PLSR,PLSR-RF)to predict soil salt content was proposed.A total of 45 soil samples were collected in the study,35 of which were randomly selected as the modeling set and 10 of which were randomly selected as the verification set.The experiment first performed segmentation processing on the collected hyperspectral image of the soil to extract the original reflection spectrum of the soil at 400-1000 nm,and then performed 4 kinds of spectral transformations on the original reflection spectrum(first-order differential,first-order differential of multiple scattering correction,SG smoothing Denoising first-order differential and logarithmic first-order differential).And it performed correlation analysis(CA)with the measured salt content of the soil,utilized the correlation coefficient to select the sensitive band,and finally established a regression model combining partial least squares and random forest.Compared with partial least square regression and random forest regression,the prediction accuracy of the combination of the two models was significantly improved.The PLSR-RF inversion model that established by the first-order differential transformation of the spectrum was more obvious.Its modeling set determination coefficient R_(c)^(2) was 0.852,the root mean square error RMSE_(c) was 0.102 g/kg,and the relative analysis error RPD_(c) was 2.600.The set determination coefficient R_(v)^(2) was 0.941,the root mean square error RMSE_(v) was 0.049 g/kg,and the relative analysis error RPD_(v) was 4.117.