引用本文:曹肖奕,丁建丽,葛翔宇,王敬哲.基于光谱指数与机器学习算法的土壤电导率估算研究[J].土壤学报,2020,57(4):867-877. DOI:10.11766/trxb201902190024
CAO Xiaoyi,DING Jianli,GE Xiangyu,WANG Jingzhe.Estimation of Soil Electrical Conductivity Based on Spectral Index and Machine Learning Algorithm[J].Acta Pedologica Sinica,2020,57(4):867-877. DOI:10.11766/trxb201902190024
【打印本页】   【HTML】   【下载PDF全文】   查看/发表评论  【EndNote】   【RefMan】   【BibTex】
←前一篇|后一篇→ 过刊浏览    高级检索
本文已被:浏览 611次   下载 389 本文二维码信息
码上扫一扫!
分享到: 微信 更多
基于光谱指数与机器学习算法的土壤电导率估算研究
曹肖奕 ,2,3, 丁建丽1,2,3, 葛翔宇1,2,3, 王敬哲1,2,3
(1.新疆大学资源与环境科学学院,乌鲁木齐 830046;2. 新疆大学绿洲生态教育部重点实验室,乌鲁木齐 830046;3. 新疆大学智慧城市与环境建模自治区普通高校重点实验室,乌鲁木齐 830046)
摘要:
土壤盐分是干旱区土壤盐渍化评价的重要指标。以新疆维吾尔自治区渭干河-库车河三角洲绿洲为例,基于土壤电导率 (Electrical conductivity,EC) 及可见光-近红外 (Visible and near infrared, VIS-NIR) 光谱数据,通过蒙特卡洛交叉验证 (Monte Carlo cross validation, MCCV) 确定364个有效样本。采用原始光谱 (Raw reflectance, R) 及其经过微分、吸光度 (Absorbance, Abs)、连续统去除 (Continuum removal, CR) 等6种预处理后的数据构建光谱指数。基于遴选出的21个最优指数,采用BP神经网络 (Back propagation neural network, BPNN)、支持向量机 (Support vector machine, SVM)、极限学习机 (Extreme learning machine, ELM) 三种算法对EC进行估算,并引入偏最小二乘回归 (Partial least squares regression, PLSR) 进行比较。结果表明:在基于R与6种光谱预处理数据构建的21个最优光谱指数之中,R_FD_RSI (R1913,R2142) 表现最佳 (r = 0.649) ;与PLSR相比,机器学习算法能够显著提高模型的估算精度,R2提高了34.55%。三种机器学习算法模型中,ELM表现最优 (R2 = 0.884, RMSE = 3.071 mS?cm-1, RPIQ = 2.535) 。本研究中所构建的光谱指数在兼顾遥感机理的同时能深度挖掘更多的隐含信息,并且基于机器学习算法的土壤EC估算模型精度显著提高,为干旱区土壤盐分定量估算提供了科学参考。
关键词:  光谱  土壤电导率  光谱预处理  光谱指数  机器学习
基金项目:国家自然科学基金项目(41771470)
Estimation of Soil Electrical Conductivity Based on Spectral Index and Machine Learning Algorithm
CAO Xiaoyi 1, 2,3, DING Jianli 1, 2, 3, GE Xiangyu 1, 2,3, WANG Jingzhe 1, 2,3
1. College of Resources and Environmental Science, Xinjiang University, Urumqi 830046, China;2. Key Laboratory of Oasis Ecology under Ministry of Education, Xinjiang University, Urumqi 830046, China;3. Key Laboratory of Smart City and Environment Modelling of Higher Education Institute, Xinjiang University, Urumqi 800046, China
Abstract:
【Objective】Soil salinity is an important indicator for evaluation of soil salinization in arid regions. It is of great practical significance to grasp real-time information about salinization to disaster prevention, ecology stabilization and harmonization of human-land relationship in this area. 【Method】A total of 400 soil samples were collected from the Weigan River-Kucha River Delta Oasis in the Xinjiang Uygur Autonomous Region of China in October 2017, and prepared, separately, with distlled water into suspensions, 5:1 in ratio, from which soil extracts were obtained for analysis of electrical conductivity (EC) and VIS-NIR (visible-near infrared) spectral reflectances in the laboratory. Based on the obtained data and the Monte Carlo cross validation (MCCV), 364 samples were determined to be valid. After the raw spectrum reflectances (R) were pre-processed with differential, absorbance (Abs), continuum removal (CR) and three others, 21 spectal indices were selected and established. Then based on the 21 optimal spectral indices, EC was assessed using the back propagation neural network (BPNN), support vector machine (SVM), and extreme learning machine (ELM), compared to the partial least squares regression (PLSR), and validated with its root mean square error (RMSE), determination coefficient (R2) and ratio of performance to interquartile range (RPIQ). 【Result】Results show: the 21 optimal spectral indices based on R and its pretreated data are R_NDSI(R2006, R2257), R_DSI(R1882, R2010), R_RSI(R2006, R2257), R_FD_NDSI(R1376, R2142), R_FD_DSI(R1376, R2142), R_FD_RSI(R1913, R2142), R_SD_NDSI(R416, R2470), R_SD_DSI(R894, R1373), R_SD_RSI(R689, R2355), ABS_NDSI(R2005, R2168), ABS_DSI(R2006, R2257), ABS_RSI(R2006, R2168), ABS_FD_NDSI(R876, R2490), ABS_FD_DSI(R1376, R2123), ABS_FD_RSI(R1913, R2142), ABS_SD_NDSI(R1081, R1725), ABS_SD_DSI(R858, R1374), ABS_SD_RSI(R709, R2355), CR_NDSI(R2119, R2261), CR_DSI(R2119, R2261), and CR_RSI(R2119, R2261), among which R_FD-RSI (R1913, R2142) is the optimal (r= 0.649) one. Compared with the PLSR, the machine learning algorithm (MLA) could significantly improve accuracy of the model, with the R2 increased by 34.55%. Among the three models using the machine learning algorithm, ELM was the best (R2 = 0.884, RMSE = 3.071 mS?cm-1, RPIQ = 2.535). 【Conclusion】In this study, different spectral pretreatment methods were used to obtain 21 optimal spectral indices. In constructing the spectral indices in this study, besides considering the remote sensing mechanism, it is advisable to explore in depth more implicit information. Compared with the traditional linear model, the MLA-based soil EC estimation model is obviously higher in accuracy. All the findings in this study may serve as a scientific reference for quantitative estimation of soil salinity in arid regions.
Key words:  Spectral  Soil electrical conductivity  Spectral pretreatment  Spectral index  Machine learning