Abstract:

Estimation of heavy metal content in soil has important scientific significance for regional soil quality assessment and soil pollution remediation, as well as providing important support for ensuring food security. In order to estimate the level of Copper (Cu), zinc (Zn), nickel (Ni) and lead (Pb) in the sloping farmland of Chongqing, and to explore the feasibility of retrieving heavy metal content using soil spectra, this study firstly analyzed the statistical characteristics and occurrence relationship of soil heavy metals based on the hyperspectral data of soil samples and the laboratory test results. The original spectral variables were transformed into first-order differential (FD), second-order differential (SD), logarithmic reciprocal (LR) and continuum removal (CR) respectively to analyze the correlation between soil heavy metal content and spectral variables, thus the characteristic band of the soil spectrum could be determined. Partial least square regression (PLSR) and support vector machine (SVM) were used to invert the heavy metal content, and the accuracy and stability of the modeling set and verification set were analyzed. Finally, the optimal spectral transformation and model combination of each heavy metal inversion were determined through the comparative analysis of model accuracy. The results showed that there were significant positive correlations (P<0.05) between soil heavy metals, Fe, Mn and soil organic matter. There were significant correlations between wavelength (around 445, 530, 1002, 1414, 1913, 2218 and 2320 nm) and heavy metals. The accuracy of PLSR modeling and LR variable combination simulation results was higher, but in general, the model did not have the ability to estimate elements with relatively low content especially for Cr, Ni and Pb. The overall accuracy of SVM simulation was better and more stable than that of PLSR. The SVM model based on CR had the highest inversion accuracy for the five heavy metal elements (r_M² is between 0.68 and 0.86), and the content of heavy metals could be effectively retrieved by soil spectrum. Soil spectroscopy can provide an important means for estimating soil heavy metal content in this study area, and can also provide a reference for soil heavy metal content monitoring and soil environmental assessment in Southwest China.

Key words: soil heavy metals, hyperspectral, inversion, support vector machine, partial least squares

摘要：

土壤重金属含量估算对于区域土壤质量评估和土壤污染修复具有重要的科学意义，也是保障粮食安全的重要支撑。为估算重庆市耕地中土壤重金属铜（Cu）、锌（Zn）、铬（Cr）、镍（Ni）、铅（Pb）的含量，探讨利用土壤光谱进行重金属含量反演的可行性，基于土壤样品的室内测试高光谱数据和重金属含量，首先进行土壤重金属统计特征和赋存关系分析，然后对原始光谱数据分别进行一阶微分（FD）、二阶微分（SD）、倒数对数（LR）及去包络线（CR）变换，分析土壤重金属含量与光谱变量之间的相关性，从而确定土壤光谱特征波段；利用偏最小二乘回归（Partial Least Square Regression，PLSR）和支持向量机（Support Vector Machine，SVM）进行重金属含量的反演建模，对建模集和验证集进行模型精度和稳定性分析；根据模型精度对比分析，确定不同重金属反演的最佳光谱变换和模型组合。结果表明，土壤重金属元素与Fe、Mn元素、土壤有机质之间存在显著正相关关系（P<0.05）；土壤光谱与重金属元素间存在显著相关的波长主要包括445、530、1002、1414、1913、2218、2320 nm；PLSR建模与LR变量组合模拟结果的精度较高，但是总体上该模型对5种重金属含量的建模精度都较低，尤其对于含量比较低的Cr、Ni和Pb元素，不具备估算能力；而SVM模拟精度整体优于PLSR，且基于CR光谱变换的SVM模型对5种重金属元素的反演精度最高（r_M²介于0.68—0.86之间），SVM模型更加具有稳定性。土壤光谱可为本研究区土壤重金属含量估算提供重要手段，也可为西南地区土壤重金属含量监测及土壤环境评价提供参考。

关键词: 土壤重金属, 高光谱, 反演, 支持向量机, 偏最小二乘