Estimating chromium concentration in arable soil based on the optimal principal components by hyperspectral data

• Main Authors: Guo, F. (Author), Li, K. (Author), Liu, F. (Author), Liu, X. (Author), Ma, H. (Author), Peng, M. (Author), Tang, S. (Author), Xu, Z. (Author), Yang, Z. (Author), Zhang, L. (Author)
• Format: Article
• Language: English
• Published: Elsevier B.V. 2021
• Subjects: agricultural soil; Agriculture; arable land; Chromium; Chromium concentration; concentration (composition); data set; Decision trees; estimation method; Gradient boosting; Gradient boosting decision tree; Gradient boosting decision tree (GBDT); Health risks; heavy metal; Heavy metals; Hyperspectral models; Hyperspectral remote sensing technology; Least squares approximations; Linear regression; Nonlinear systems; Optimal principal component; Optimal principal components; Partial least square regression; Partial least squares regression (PLSR); Principal component analysis; Principal Components; remote sensing; Remote sensing; Soil chromium; soil pollution; Soil pollution; Soils
• Online Access: View Fulltext in Publisher

 The heavy metal pollution in arable soil poses a significant threat to human health. Thus, it is of great significance to investigate the contamination of heavy metal elements in the soil. As the soil polluted by heavy metal is sensitive to spectral reflectance, thus the hyperspectral remote sensing technology could be a valuable tool for retrieving heavy metal components in the soil. This study, taking chromium (Cr) concentration as an example, proposes an optimal model for estimating heavy metal components in the soil by comprehensively taking account of the spectral pretreatment, dimensionality reduction with optimal parameters, and hyperspectral model. To this end, both the linear model, i.e., partial least squares regression (PLSR), and the nonlinear model, i.e., the gradient boosting decision tree (GBDT), are applied in this study. It is found in the study area, the Savitzky-Golay (SG) method can be regarded as an excellent spectral pretreatment for the hyperspectral data regardless of the applied model. By contrast, the dimensionality reduction in terms of the Principal Component Analysis (PCA) is closely related to hyperspectral model: the optimal principal components (PCs) in the estimation of Cr concentration are the first 9 PCs for the GBDT (nonlinear model), while that for the PLSR (linear model) become the first 8 PCs. Moreover, the examination of hyperspectral model shows the GBDT model has slightly better performance than the PLSR model for the Cr concentration estimation under most conditions. Finally, when the spectral pretreatment, dimensionality reduction, and hyperspectral model are fully considered, the best retrieval model for the Cr concentration in the study area is the SG-PCA-GBDT model. Numeric measures of model accuracy show the proposed model has a determination coefficient of 0.80 and a residual prediction deviation of 2.04, which provides a potentially new method for estimating Cr concentration in the polluted soil. © 2021