基于成像高光谱技术的露天煤矿区复垦土壤剖面有机碳和全氮预测及制图  

Prediction and mapping of organic carbon and total nitrogen in reclaimed soil profiles in surface coal mining areas based on imaging hyperspectral technology

作　　者：彭思涵 包妮沙[1] 雷海梅 刘善军[1] 杨天鸿[1] PENG Sihan;BAO Nisha;LEI Haimei;LIU Shanjun;YANG Tianhong(School of Resources and Civil Engineering,Northeastern University,Shenyang 110000,China)

机构地区：[1]东北大学资源与土木工程学院,辽宁沈阳110000

出　　处：《煤炭学报》2023年第7期2949-2960,共12页Journal of China Coal Society

基　　金：国家重点研发计划资助项目(2022YFC2903900);国家自然科学基金资助项目(52074063,U1903216)。

摘　　要：k露天煤矿区复垦土壤在经过开采—剥离—覆盖过程后,其有机碳(SOC)和全氮(TN)的垂直分布发生巨大变化,并且直接决定植被恢复物种的选择和生态恢复效果。因此,在双碳目标下,复垦SOC和TN在0~100 cm剖面的监测及制图,是评估矿区生态恢复及复垦工程固碳的重要基础。相比于传统的化学方法,高光谱技术是一种快速而且无损,已经广泛应用于土壤属性估测的技术,成像高光谱技术相比于点位光谱测量具有图谱合一的优势。因此以我国北方干旱半干旱草原大型露天煤矿复垦土壤为研究对象,采集不同复垦工程、复垦年限的土壤剖面样本,揭示不同复垦年限和复垦方式下土壤剖面SOC和TN垂直分布规律和高光谱特征,通过深度挖掘和筛选SOC和TN特征波段和指数,构建可解释性机器学习模型,从而实现土壤剖面SOC和TN的制图,结果表明:(1)土壤光谱反射率随SOC和TN质量分数的增加而降低,通过集成皮尔森相关性分析和选择连续投影算法,对点状高光谱数据进行降维和去冗余,最终明确表征SOC质量分数的30个特征波段和表征TN质量分数的18个特征,并基于特征波段建立了SOC和TN三维光谱特征指数;(2)通过对比偏最小二乘回归(PLSR)、随机森林模型(RF)和径向基函数(RBF)神经网络预测模型,发现对光谱数据进行特征波段筛选和变换,可以提高预测精度,其中随机森林模型预测SOC和TN的精度最高,对SOC的预测结果为R^(2)=0.97、RMSE=7.5 g/kg、LCCC=0.84、bias=3.70,TN预测结果为R^(2)=0.78、RMSE=0.33 g/kg、LCCC=0.74、bias=0.19;(3)利用SHAP可解释机器学习模型方法对输入特征进行重要度排序,发现提出的三维光谱指数对模型预测的贡献大于大部分特征波段;(4)使用最佳模型可以快速绘制出成像高光谱图像中每个像素的SOC和TN质量分数,实现矿山土壤剖面中SOC和TN质量分数的制图和可视化。基于成像高光谱技术绘制露�After the process of surface mining,stripping and covering,the vertical distribution of organic carbon(SOC)and total nitrogen(TN) in the reclaimed mine soil undergoes great changes,which directly determines the selection of vegetation species and the effectiveness of ecological restoration.Therefore,under the dual carbon target,monitoring and mapping SOC and TN in reclaimed soil profile(0-100 cm) are the important basis for evaluating ecological restoration and carbon sequestration in mining areas.Compared with traditional chemical methods,hyperspectral imaging is a rapid and non-destructive technique that can be used for the effective estimation of the important indicators of soil properties.Imaging spectroscopy technology has the advantage of spectral integration compared to point spectral measurement.Therefore,the reclaimed soil from arid and semi-arid grassland mining areas in northern China was selected as the research object,and soil profiles(0-100 cm) from different mining areas,modes and years were collected.The vertical distribution patterns and hyperspectral characteristics of SOC and TN in profiles were revealed under different cultivation years and patterns.And then,with the feature bands and spectral indices of SOC and TN,the interpretable machine learning models were constructed to achieve mapping SOC and TN contents in profiles.The results indicated that:(1) The spectral reflectance of the soil samples decreased with the increase of SOC and TN contents.By combining stacked feature selection methods(Pearson correlation coefficient–successive projections algorithm,PCC-SPA),hyperspectral data were reduced in dimension.30 bands of SOC and 18 bands of TN were clearly identified,and three-dimensional(3D) spectral indices were established;(2) By comparing different machine learning algorithms(partial least squares regression;PLSR,random forest;RF,and radial basis function model;RBF),the prediction accuracy could be improved by filtering and transforming the selective bands of hyperspectral data.The RF w

关 键 词：成像高光谱技术 干旱半干旱露天煤矿 复垦土壤 土壤剖面 有机碳 全氮 

分 类 号：TD88[矿业工程—矿山开采]