基于CARS-PLSR算法的土壤有效磷高光谱反演研究

Prediction of soil available phosphorous using hyperspectral data based on CARS-PLSR

关键词:  土壤有效磷  高光谱  竞争性自适应重加权算法  连续投影算法  偏最小二乘回归法

Key words:  soil available phosphorous（SAP）  hyperspectral  competitive adaptive reweighted sampling algorithm（CARS）  successive projections algorithm（SPA）  partial least square regression（PLSR）

摘要:

【目的】 剔除土壤高光谱中包含的大量冗余和无效信息，探明土壤有效磷（SAP）的敏感波段，简化SAP的高光谱估算模型并提高模型的预测精度。【方法】 文章以四川省崇州市西河流域110个土壤样本为研究对象，利用ASD Fieldspec3地物光谱仪在室内条件下测定350~2 500 nm波段范围的土壤高光谱数据。对光谱数据进行预处理后，采用连续投影算法（SPA）和竞争性自适应重加权算法（CARS）优选的波长变量作为建模参数，运用偏最小二乘回归（PLSR）方法建立模型并比较其精度。【结果】 结果表明，标准正态变换预处理方法是SAP的最佳土壤光谱数据预处理方法。基于标准正态变换后的光谱数据，CARS、SPA算法可将预测SAP的关键波段变量分别压缩至54和13个，CARS-PLSR模型与SPA-PLSR模型相比，相关系数由0.894提高到0.945，均方根误差由5.73降低到3.56。【结论】 土壤高光谱数据经标准正态变换后，采用CARS-PLSR算法可有效提高有效磷含量预测的鲁棒性。该结果可为高光谱数据快速反演土壤有效磷含量提供理论依据。

Abstract:

【Purpose】 To remove a large amount of redundancy and invalid information in soil hyperspectrum，select the sensitive wavelengths of soil available phosphorus （SAP），simplify the SAP hyperspectral estimation model and improve the prediction accuracy of the models.【Method】 A total of 110 soil samples were collected from Xihe basin in Sichuan province. The raw hyperspectral reflectance of soil samples in the range of 350~2 500 nm wavelengths were measured by the standard procedure with a spectrometer of ASD Field Spec3 equipped with a high intensity contact probe under the laboratory conditions. Based on the pre-processing spectra，the optimized wavelength variables were selected by using the successive projections algorithm（SPA）and competitive adaptive reweighted sampling algorithm（CARS）. Final model with partial least square regression（PLSR）were established.【Result】 The results showed that SNV was the best soil spectral data pre-processing method of SAP. Based on the pre-processing spectral data by SNV，the CARS and SPA selected 54 and 13 key wavelength variables from full-spectrum to predict SAP concentrations，respectively. Compared with the SPA-PLSR model，the correlation coefficient of the CARS-PLSR model increased from 0. 894 to 0. 945，and the root mean square error reduced from 5. 73 to 3. 56.【Conclusion】 The utilization of CARS-PLSR algorithm based on soil hyperspectral data that is pre-processed by SNV could enhance SAP forecasting capability and reduce the model complexity，which could provide theoretical basis for the rapid inversion of SAP concentration by using hyperspectral data.