土壤铬的航空高光谱间接反演模型及可迁移性研究

铬（Cr）是我国东北黑土区土壤重金属污染评价中的主要目标元素之一，随着航空高光谱技术的引入，Cr含量的高光谱反演具备了大范围应用的数据基础，其中影响调查质量的关键是高光谱模型的精度及应用范围。常用的建模方法是利用各类统计学手段进行光谱特征提取并建模，局限性表现为建模结果受样本选择的影响较大，模型泛化能力不强。研究以土壤Cr的赋存规律为切入点，设计了一种新的基于Cr影响因素及光谱特征的间接反演模型，改善了模型在不同地区的适用性。选择黑龙江省建三江和海伦两个研究区，高光谱数据来自CASI/SASI航空高光谱成像系统，波段范围380～2 450 nm，建三江和海伦研究区地面采样数量分别为225个和121个，分析获得土壤Cr元素及SOM，N，P，K₂O，SiO₂，Al₂O₃，Fe₂O₃，CaO，MgO，Na₂O和pH等理化参数，建模方法采用偏最小二乘法。Cr的赋存规律分析结果表明，Cr在两个研究区均表现为与Al₂O₃，Fe₂O₃，MgO，K₂O和pH呈极显著的正相关关系，与SiO₂，Na₂O和SOM呈极显著的负相关关系，该特征为间接反演模型的建立提供了基础。两个地区Cr的光谱特征分析结果共同表明，光谱反射率经变量标准化（SNV）变换后与Cr含量的相关性最明显，特征波段为1 520，2 195，2 210和2 225 nm。将光谱SNV变换后的特征波段作为纯光谱模型自变量，将SNV特征波段和上述与Cr密切相关的土壤组分作为间接反演模型的自变量，建模结果显示，相比于纯光谱模型，建三江研究区的间接反演模型将建模R2由0.643提升到了0.751，验证R2由0.571提升到了0.687，海伦研究区的间接反演模型将建模R2由0.537提升到了0.676，验证R2由0.471提升到了0.643，间接模型相应的均方根误差（RMSE）也得到了降低，可见间接反演模型明显提升了Cr的反演精度。两个研究区之间的模型迁移性实验结果表明，纯光谱模型的可迁移性较差，模型迁移后实测与预测值的回归R2接近于0，而间接模型在两个研究区间的迁移能力得到了显著提升，海伦的间接反演模型应用到建三江时，实测与预测值的回归R2达到0.597 5，而建三江的间接模型应用到海伦时，回归R2为0.577 3。研究结果可为土壤Cr在不同地区的大范围反演制图提供一种新的途径。

Study on Indirect Inversion Model and Migration Ability of Chromium in Soil by Aerial Hyperspectral Method

Chromium(Cr)is one of the main target elements in the evaluation of soil heavy metal pollution in the black soil area of Northeast China.With the introduction of aerial hyperspectral technology,the hyperspectral inversion of Cr content has the data basis for a wide range of applications,among which the accuracy and application range of the hyperspectral model are the key factors affecting the quality of the investigation.The commonly used modeling method is to extract spectral features and build models by various statistical means.The limitations are that the model parameters are difficult to be explained,the modeling results are greatly influenced by sample selection,and the generalization ability is poor.In this study,based on the occurrence law of Cr in soil,a new indirect inversion model based on the influencing factors and spectral characteristics of Cr was designed to improve the applicability of the model in different regions.Two research areas of Jiansanjiang and Hailun in Heilongjiang Province were selected.The hyperspectral data came from the CASI/SASI aerial hyperspectral imaging system.The band range was 380~2450 nm.The number of ground samples in Jiansanjiang and Hailun were 225 and 121 respectively.The physical and chemical parameters of Cr,SOM,N,P,K₂O,SiO₂,Al₂O₃,Fe₂O₃,CaO,MgO,Na₂O and pH were obtained by chemical analysis.The partial least square method was used in modeling.The analysis results of the occurrence rule of Cr show that Cr in both research areas shows a very significant positive correlation with Al₂O₃,Fe₂O₃,MgO,K₂O and pH,and a very significant negative correlation with SiO₂,Na₂O and SOM.This feature provides a foundation for the establishment of indirect inversion models.The analysis results of the spectral characteristics of Cr in the two regions together show that the spectral reflectance has the most obvious correlation with the Cr content after standard normalized variable(SNV)transformation,and the characteristic bands are 1520,2195,2210 and 2225 nm.The characteristic band after SNV transformation was taken as the independent variable of the pure spectral model.The SNV characteristic band and the above-mentioned soil components closely related to Cr were taken as the independent variable of the indirect inversion model.The modeling results show that,compared with the pure spectral model,the indirect inversion model significantly improved the inversion accuracy of Cr.In the Jiansanjiang area,the modeling R2 has been improved from 0.643 to 0.751,and the verification R2 has been improved from 0.571 to 0.687.In the Hailun area,the modeling R2 has been improved from 0.537 to 0.676,and the verification R² has been improved from 0.471 to 0.643.The root mean square error(RMSE)of the indirect model is also reduced.The experimental results of model migration between the two study areas show that the migration ability of the pure spectral model is poor,and the regression R2 of measured and predicted values after model migration is close to 0,while the migration ability of the indirect model in the two study areas is significantly improved.When indirect inversion model of Hailun was applied to Jiansanjiang,the regression R² of measured and predicted values reaches 0.5975,while the indirect inversion model of Jiansanjiang was applied to Hailun,the regression R² is 0.5773.The results can provide a new method for large-scale inversion mapping of Cr in the soil in different areas.