聂小军, 洪雯雯, Ammara Gill,等.基于CWT-CARS-CNN综合方法的矿区土壤煤源碳质量分数高光谱估测[J].河南理工大学学报（自然科学版）,doi:10.16186/j.cnki.1673-9787.2023030004

NIE X J, HONG W W, GILL A, et al.Hyperspectral estimation of coal-derived carbon content in mine soils based onthe CWT-CARS-CNN integrated method[J].Journal of Henan Polytechnic University(Natural Science) ,doi:10.16186/j.cnki.1673-9787.2023030004

基于CWT-CARS-CNN综合方法的矿区土壤煤源碳质量分数高光谱估测(网络首发)

聂小军, 洪雯雯, Ammara Gill, 于海洋, 陈晓东

河南理工大学 测绘与国土信息工程学院，河南 焦作 454000

摘要:针对目前尚无可靠方法定量识别土壤中煤源碳的问题，采集焦作矿区耕地土壤，配制不 同质量分数的煤源碳土壤样品249个，利用ASD FieldSpec4获取样品光谱数据，采用连续小波 变换（continuous wavelet transform，CWT）-竞争性自适应重加权采样（competitive adaptive reweighted sampling，CARS）-卷积神经网络（convolutional neural networks，CNN）方法估测 土壤中 煤源碳质量分数，并对比CWT-CARS-CNN方法与传统光谱指数法构建的煤源碳估测模型估测 效果，检验CWT-CARS-CNN估测模型的适用性。结果表明：350~2 500 nm波段内，煤与土壤光 谱特征截然不同，含煤土壤样品光谱反射率随煤源碳质量分数增加而降低；CWT方法提高了 光谱对土壤煤源碳的敏感性，CARS方法提取的煤源碳敏感波段数量明显增加，基于CWT-CARS-CNN 方法的煤源碳含量估测模型精度明显高于传统光谱指数法的；以L8分解尺度构建 的 CWT-CARS-CNN 模型精度最高，验证集 R²=0.999 3，RPD=40.308 1。基于 CWT-CARS-CNN 的高光谱估测方法能较准确地估测矿区不同土地利用类型下土壤中的煤源碳质量分数，可为 双碳背景下矿区土壤碳固存与肥力评估提供参考。

关键词:煤源碳;碳固存;高光谱估测;深度学习;矿区土壤

中图分类号：S153.6;TP183;O657.3

doi:10.16186/j.cnki.1673-9787.2023030004

基金项目:国家自然科学基金资助项目（41977284）；河南省科技攻关项目（222102320032）

收稿日期:2023-03-02

修回日期:2023-05-05

网络首发日期: 2023-10-12

Hyperspectral estimation of coal-derived carbon content in mine soils based onthe CWT-CARS-CNN integrated method（Online）

NIE Xiaojun, HONG Wenwen, GILL Ammara, YU Haiyang, CHEN Xiaodong

School of Surveying and Mapping Land Information Engineering，Henan Polytechnic University，Jiaozuo 454000，Henan，China

Abstract:As a high-carbon （C） organic matter，coal particles diffused in the soils can lead to apparent overestimation of soil organic C stocks.However，there is a shortage of reliable methods in quantitating coal- derived C in soils. In this study，soil samples from cultivated lands of Jiaozuo mining area were collected， 249 artificial mixtures of soil and coal samples were prepared to form different contents of coal-derived C， and spectral data of the samples were obtained by ASD FieldSpec4. The Continuous Wavelet Transform （CWT）-Competitive Adaptive Reweighted Sampling（CARS）-Convolutional Neural Network（CNN） method was used to estimate coal-derived C content in soil.The estimation effect of coal-derived C content between the CWT-CARS-CNN and traditional spectral index models，was compared，and the applicability of the CWT-CARS-CNN model was also tested.It was found that in the range of 350-2500nm，the hyperspectralcharacteristics between coal and soil are largely different. The spectral reflectance of coal-contained soilsamples decreases with increasing coal-derived C content.The CWT method improved the sensitivity of thespectrum to the coal-derived C content in soil.The number of feature waveband of coal-derived C extractedby the CARS was obviously increased.In general，accuracies of coal-derived C estimation models based onthe CWT-CARS-CNN integrated method were significantly higher than those based on traditional spectral in⁃dex method.Especially，the CWT-CARS-CNN model constructed with L8 decomposition scale exhibited thehighest accuracy，showing R²=0.999 1 and RPD=40.308 1 for its validation set. In addition，the estimationmodel showed good applicability to coal-contained soils under different land use types.Our study suggests that hyperspectral estimation based on the CWT-CARS-CNN integrated method can provide method support for accurate assessment of soil C sequestration and fertility in coal mining area under the “Double C” background.

Key words:coal-derived carbon;carbon sequestration;hyperspectral estimation;deep learning;mine soils

CLC：S153.6;TP183;O657.3