淮北矿区多种类型植被地上生物量反演研究  

Research on Biomass Inversion of Multiple Vegetation Types on the Surface of Mining Areas

作　　者：杨可明[1] 彭里顺 张燕海[2] 谷新茹 陈新阳 江克贵 YANG Keming;PENG Lishun;ZHANG Yanhai;GU Xinru;CHEN Xinyang;JIANG Kegui(College of Geoscience and Surveying Engineering,China University of Mining and Technology(Beijing),Beijing 100083,P.R.China;General Defense Geological Survey Department,Huaibei Mining Co.,Ltd.,Huaibei 235000,P.R.China)

机构地区：[1]中国矿业大学(北京)地球科学与测绘工程学院,北京100083 [2]淮北矿业股份有限公司通防地测部,安徽淮北235000

出　　处：《生态环境学报》2024年第7期1027-1035,共9页Ecology and Environmental Sciences

基　　金：国家科技基础资源调查专项(2022FY101905);企、事业单位委托项目(2023-129);国家自然科学基金项目(41971401)。

摘　　要：生物量是植被碳的重要组成部分,对于评估矿山生态环境修复效果至关重要。在淮北矿业闸河矿区,生态破坏与修复并存,地表植被类型丰富多样,有林木、草本和农作物等。在这种多植被类型交叉覆盖的地区,传统的单一物种生物量模型难以应用。为准确反演该区域综合植被的生物量,以sentinel卫星数据为基础,利用植被指数和波段运算方法构建的光谱指数作为特征数据集,通过不同特征和模型的组合构建该区域生物量反演模型。在特征选择阶段,利用随机森林和相关性分析两种方法对特征进行综合评价,最终筛选出关键的5个光谱指数(增强植被指数、归一化差分水体指数、改进的归一化差分水体指数、Sentinel-2红边位置、陆地叶绿素指数)以及5个基于波段运算的自建光谱特征(1/B1、1/B2、1/B7、B1/B2、B5/B6)。基于这些筛选出的特征,设计了5种不同的特征组合方案,并结合传统方程回归和机器学习模型,在林地和综合植被数据上分别构建了生物量反演模型。研究结果表明,相较于传统回归模型,机器学习模型在生物量反演领域具有更高的精度。基于波段运算的自建光谱特征可以提高机器学习模型生物量反演的精度,其中支持向量回归(SVR)模型结合自建光谱特征与原始波段得到的反演结果精度最高,模型验证集的决定系数R^(2)为0.74,均方根误差为8.14kg·m^(-2)。将SVR模型应用于研究区进行生物量反演,得到的结果与当地植被分布特征高度一致。本研究成果不仅为矿区生态修复评价提供了数据支撑,而且为类似生态系统的后续研究提供了借鉴和经验。Biomass is an important component of vegetation carbon,and is crucial for evaluating the effectiveness of ecological restoration in mining areas.In the Zhahe mining area of Huaibei Mining,ecological destruction and restoration coexist with a rich variety of surface vegetation types including trees,herbs,and crops.Traditional single-species biomass models are difficult to apply in areas where multiple vegetation types overlap.To accurately invert the biomass of the composite vegetation in this region,this study used Sentinel satellite data and constructed a spectral index feature dataset using vegetation indices and band operation methods.Different features and model combinations were used to construct biomass inversion models for each region.In the feature selection stage,two methods,random forest and correlation analysis,were used to evaluate the features comprehensively.Ultimately,five key spectral indices(enhanced vegetation index,normalized difference water index,improved normalized difference water index,Sentinel-2 red edge position,and land chlorophyll index)and five self-constructed spectral features based on band operations(1/B1,1/B2,1/B7,B1/B2,B5/B6)were selected.Based on these selected features,five different feature combination schemes were designed,and biomass inversion models were constructed separately for forest and composite vegetation data by combining the traditional equation regression and machine learning models.The results showed that compared with traditional regression models,machine learning models have higher accuracy in biomass inversion.Self-constructed spectral features based on band operations can improve the accuracy of biomass inversion using machine learning models.Among them,the support vector regression(SVR)model combined with self-constructed spectral features and original bands achieved the highest accuracy in inversion results,with a determination coefficient R^(2) of 0.74 and a root mean square error of 8.14 kg·m^(-2) for the model validation set.Applying the SVR model to bi

关 键 词：矿山 多植被交叉覆盖区 综合生物量 机器学习 反演模型 植被类型 

分 类 号：P23[天文地球—摄影测量与遥感] Q948[天文地球—测绘科学与技术] X173[生物学—植物学]