【摘要】 镉(Cd)作为土壤重金属污染物，由于其难降解特性及较强的迁移能力，使其极易在农作物内富集，进而通过食物链威胁人体健康。高光谱遥感技术利用高光谱分辨率和图谱合一的优势，为土壤重金属的定量监测提供了一种高效且成本较低的解决方案。以宁夏回族自治区为研究区域，采集了104个土壤样本，利用ASD FieldSpec4地物光谱仪测定土壤光谱数据。首先，针对高光谱数据易受噪声干扰和多重共线性影响，需通过对原始光谱数据进行Savizky-Golay平滑(SG平滑)及一阶微分(FD)、二阶微分(SD)、倒数对数一阶微分(ATFD)、倒数对数二阶微分(ATSD)和连续统去除(CR)等光谱预处理来减弱土壤散射和噪声的影响，增强信息并提高模型精度。其次，结合皮尔逊相关系数和竞争自适应重加权采样算法筛选特征波段，基于偏最小二乘回归(PLSR)、随机森林(RF)和支持向量机(SVM)三种反演模型构建土壤重金属镉含量的高光谱估算模型。最后，通过建模集和验证集对模型的精度和稳定性进行分析，确定最佳的光谱变换和模型组合。结果表明，随机森林(RF)模型在建模精度方面表现最佳，其中RF-CR组合模型的决定系数(R~2)为0.947,均方根误差(RMSE)为0.010 3。然而，在预测精度方面，支持向量机(SVM)模型表现更优，SVM-CR组合模型的RMSE为0.115 5,R~2为0.414。虽然镉的建模和预测系数较低，但其均方根误差(RMSE)极小，可能与土壤中镉含量较低有关，说明拟合效果较好。通过比较不同光谱预处理方法对模型精度的影响，能够有效提高土壤重金属镉含量反演的精度，为宁夏农田的重金属治理与监管提供更便捷的方法和优质农产品安全生产提供新的技术手段。后续研究可融合多种特征波段选择方法、预处理手段和建模方式，增加样本多样性以验证模型泛化能力。此外，开展控制实验筛选对镉含量敏感的特征波段，有望进一步提升反演精度。更多还原

【Abstract】 Cadmium(Cd), as a heavy metal pollutant in soil, due to its difficult degradability and strong migration ability, is prone to accumulate in crops, thereby threatening human health through the food chain. Hyperspectral remote sensing technology, leveraging its high spectral resolution and integrating spectra and images, provides an efficient and cost-effective solution for quantitative monitoring of soil heavy metals. This study selected the Ningxia Hui Autonomous Region as the research area and collected 104 soil samples. The soil spectral data were measured using the ASD FieldSpec4 spectrometer. Firstly, to address the issues of noise interference and multicollinearity in hyperspectral data, spectral preprocessing methods such as Savizky-Golay smoothing, first derivative(FD), second derivative(SD), inverse logarithmic first derivative(ATFD), inverse logarithmic second derivative(ATSD), and continuum removal(CR) were applied to the original spectral data to reduce the influence of soil scattering and noise, enhance information, and improve model accuracy. Secondly, the feature bands were selected using a combination of the Pearson correlation coefficient and the competitive adaptive reweighted sampling algorithm. Based on three inversion models, namely partial least squares regression(PLSR), random forest(RF), and support vector machine(SVM), hyperspectral estimation models for soil cadmium content were constructed. Finally, the accuracy and stability of the models were analyzed using the modeling and validation sets to determine the best spectral transformation and model combination. The results showed that the random forest(RF) model performed best in terms of modeling accuracy, with the RF-CR combination model having a coefficient of determination(R~2) of 0.947 and a root mean square error(RMSE) of 0.010 3. However, in terms of prediction accuracy, the support vector machine(SVM) model performed better, with the SVM-CR combination model having an RMSE of 0.115 5 and an R~2 of 0.414. Although the modeling and prediction coefficients for cadmium were relatively low, the RMSE was extremely small, which might be related to the low cadmium content in the soil, suggesting a good fit. By comparing the effects of different spectral preprocessing methods on model accuracy, the precision of soil cadmium content inversion can be effectively improved, providing a more convenient method for heavy metal governance and supervision in Ningxia farmland and a new technical means for the production of safe, high-quality agricultural products. Future research can integrate multiple feature band selection methods, preprocessing techniques, and modeling approaches, and increase sample diversity to verify the model’s generalization ability. Additionally, conducting controlled experiments to screen for feature bands sensitive to cadmium content is expected to enhance inversion accuracy further.更多还原