| 基于机器学习算法的土壤有机质
质量比估算 |
| |
| 引用本文： | 白婷,丁建丽,王敬哲.基于机器学习算法的土壤有机质
质量比估算[J].排灌机械工程学报,2020,38(8):829-834. |
| |
| 作者姓名： | 白婷 丁建丽 王敬哲 |
| |
| 作者单位： | 新疆大学资源与环境科学学院,新疆 乌鲁木齐 830046;新疆大学绿洲生态教育部重点实验室,新疆 乌鲁木齐 830046;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,新疆 乌鲁木齐 830046;新疆大学资源与环境科学学院,新疆 乌鲁木齐 830046;新疆大学绿洲生态教育部重点实验室,新疆 乌鲁木齐 830046;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,新疆 乌鲁木齐 830046;新疆大学资源与环境科学学院,新疆 乌鲁木齐 830046;新疆大学绿洲生态教育部重点实验室,新疆 乌鲁木齐 830046;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,新疆 乌鲁木齐 830046 |
| |
| 摘 要： | 为快速高效地估测干旱、半干旱地区土壤有机质(soil organic matter, SOM)质量比,提出了一种结合竞争适应重加权法(CARS)和随机森林(RF)的估测模型.以内陆干旱区艾比湖流域为研究区,测定土壤高光谱反射率和SOM质量比,经预处理后,利用CARS对原始光谱(R)、一阶导数(R′)、吸光度(log(1/R))及吸光度一阶导数［log(1/R)］′4种光谱变量的可见-近红外光谱进行筛选,并结合RF算法,建立全谱段RF模型与CARS-RF模型.结果表明,基于CARS方法对光谱进行变量筛选后,得出4种光谱变量的优选变量集个数分别为35,26,34和121;在4种光谱变量中,R′和［log(1/R)］′的SOM估测模型精度较高,以［log(1/R)］′为基础数据获得的模型精度最高;CARS-RF模型精度优于全谱段RF模型,模型验证集决定系数(R2)、均方根误差(RMSE)、相对分析误差(RPD)分别为0.881,6.438 g/kg和2.177.该研究在预处理的基础上通过变量优选,应用较少的变量个数获得较高的估测精度,为干旱、半干旱区SOM高光谱估测提供了适宜高效的方法.
|
| 关 键 词： | 土壤有机质 高光谱 光谱变量 竞争适应重加权法 随机森林 |
| 收稿时间： | 2018-08-09 |
Estimation of soil organic matter content based on machine learning |
| |
| BAI Ting,DING Jianli,WANG Jingzhe.Estimation of soil organic matter content based on machine learning[J].Journal of Drainage and Irrigation Machinery Engineering,2020,38(8):829-834. |
| |
| Authors: | BAI Ting DING Jianli WANG Jingzhe |
| |
| Institution: | 1. College of Resources & Environmental Science, Xinjiang University, Urumqi, Xinjiang 830046, China; 2. Key Laboratory of Oasis Ecology, Xinjiang University, Urumqi, Xinjiang 830046, China; 3. Key Laboratory of Smart City and Environment Modelling of Higher Education Institute, Xinjiang University, Urumqi, Xinjiang 830046, China |
| |
| Abstract: | To estimate soil organic matter(SOM)quickly and efficiently, a combined estimation model of the competitive adaptativereweighted sampling(CARS)and random forests(RF)was developed. The Ebinur Lake Basin was selected as a study area, and the soil hyperspectral reflectance and SOM content were measured. After data pre-processing, the visible-near infrared spectra of the four spectral variables, the original spectrum(R), the first derivative(R′), absorbance(log(1/R))and the first derivative of absorbance(［log(1/R)］′)were screened with CARS method, and the full-spectrum RF and CARS-RF models were further developed with RF algorithm. Results indicated that after screening the variables with the CARS method, the number of preferred variable sets for the four spectral variables was 35, 26, 34, and 121, respectively; between the four spectral variables, the R′ and ［log(1/R)］′ showed higher accuracy in the estimation of SOM, and the model accuracy based on ［log(1/R)］′ is the highest; the accuracy of the CARS-RF model is better than that of the full spectrum RF model, and the verification set coefficient(R2), root mean square error(RMSE)and relative analysis error(RPD)for the CARS-RF model were 0.881, 6.438 g/kg and 2.177, respectively. It can be concluded that based on the data pre-processing, this study can provide a more suitable and efficient method for the estimation of arid and semiarid lands SOM by using variable preferred method and with less variables. |
| |
| Keywords: | soil organic matter hyperspectral spectral variables CARS algorithm random forests |
| 本文献已被 CNKI 万方数据 等数据库收录！ |
| 点击此处可从《排灌机械工程学报》浏览原始摘要信息 |
| 点击此处可从《排灌机械工程学报》下载免费的PDF全文 |
|
