基于RF和连续小波变换的露天煤矿土壤锌含量高光谱遥感反演

露天煤矿周边存在潜在重金属污染隐患，快速获取土壤重金属空间分布是土壤污染评价、土地复垦与修复的前提。传统调查方法费时费力且易造成对环境的二次污染，高光谱遥感为土壤重金属反演提供了新的视角。该研究以某露天煤矿土壤锌（Zn）含量为研究对象，采集了111个原位表层（0～20 cm）土壤样品及反射光谱；对样品反射光谱进行Savitsky-Golay（SG）平滑、连续统去除（Continuum Removal，CR）和连续小波变换（Continuous Wavelet Transform，CWT）以降噪和增强；利用Boruta算法确定特征波段；采用偏最小二乘回归（Partial Least Squares Regression，PLSR）和随机森林（Random Forest，RF）构建土壤Zn含量反演模型，使用留一交叉验证评估反演模型精度以确定最优反演模型；基于最优反演模型，利用空间插值方法绘制土壤Zn含量空间分布图。结果表明：1）CWT可有效降低光谱噪声，增强光谱响应。2）Boruta算法能消除光谱信息冗余，并能有效提取特征波段；特征波段的数目随CWT分解尺度和光谱测量条件变化。3）RF估算土壤Zn含量性能优于PLSR，且RF结合CWT具有较好的土壤Zn反演能力；最优野外原位光谱反演模型精度（建模集R2=0.92，验证集R2=0.54）低于实验室光谱反演模型（建模集R2=0.95，验证集R2=0.72）。4） 土壤Zn空间分布表现出显著的异质性，呈现高值集中于研究区西南部和东北部的特征。研究结果可为利用高光谱遥感开展露天矿区土壤重金属反演提供借鉴，为其他类似区域土壤污染评价、土地复垦与整治、土壤修复提供前提与依据。

Inversion of soil zinc contents using hyperspectral remote sensing based on random forest and continuous wavelet transform in an opencast coal mine

Abstract: Zinc (Zn) is one of the common elements of heavy metal pollutants in open-pit coal mines. Heavy metals in tailings have been accumulated around the mining areas, leading to a huge threat to the farmlands and urban ecosystems. The serious environmental issue has been highly urgent for the largest producer and consumer of coal in the world, China. Particularly, coal is one of the major fossil fuels, accounting for more than 70% of the total energy. What is worse, heavy metals may accumulate toxicity via the food chain, which may cause harmful impacts on public health. Specifically, long-term exposure to Zn pollution can lead to gastrointestinal discomfort, nausea, vomiting, abdominal pain, and respiratory symptoms. In addition, exposure to Zn with a high dose may also affect the cholesterol balance and fertility in agricultural production. A promising and practical way can be to determine the spatial distribution of soil heavy metals for soil pollution evaluation and control, land reclamation, and soil remediation. Most efforts were made on the potential risks of heavy metal pollution around the open-pit coal mines. However, the rapid detection is still lacking in the soil element distribution for the proper treatment with the heavy metals pollution. The traditional examination can also be time-consuming, laborious, and easy to cause secondary pollution. Fortunately, hyperspectral remote sensing technology can provide a new perspective to identify the concentrations of soil heavy metals in recent years. This study aims to implement the hyperspectral remote sensing inversion of soli Zn contents in an opencast coal mine using Random Forest (RF) and Continuous Wavelet Transform (CWT). 111 soil samples were collected to measure the Zn concentrations and spectral reflectance. The Savitsky Golay (SG) smoothing, Continuum Removal (CR), and CWT were used to reduce the noise for the spectral response characteristics of soil samples. Moreover, the characteristic bands were selected via the Boruta Algorithm. The Partial Least squares regression (PLSR) and Random Forest (RF) were introduced to construct the inversion model of soil Zn contents. Furthermore, the leave one out cross-validation was carried out to examine the robustness of the optimal model. Last but not least, the spatial distribution of Zn concentrations was mapped by the geographical interpolation using the optimum models. The results indicated that: 1) The CWT significantly enhanced the spectral responses to reduce the noise of spectral data. 2) The Boruta algorithm accurately retrieved the characteristic bands to remove the redundancy of spectral information. The number of characteristic bands varied with the different CWT decomposition scales and spectral measurement conditions. 3) The RF performed better to estimate the soil Zn contents, compared with the PLSR. The CWT combined with RF presented a higher accuracy than the rest. Particularly, the errors of the optimal field in-situ spectral inversion model (the determination coefficients of calibration dataset and validation dataset were 0.92, 0.54 respectively) were significantly lower than those of the laboratory spectral inversion model (the determination coefficients of calibration dataset and validation dataset were 0.95, 0.72 respectively). 4) There was significant heterogeneity in the spatial distribution of soil Zn. Specifically, the high values were concentrated in the southwest and northeast of the study area. This finding can also provide a strong reference to detect the soil heavy metal concentrations using the hyperspectral remote sensing technology, particularly for the soil pollution evaluation, land reclamation, and remediation, as well as soil remediation in the mining areas.