引用本文:国佳欣,赵小敏,郭 熙,徐 喆,朱 青,江叶枫.基于PLSR-BP复合模型的红壤有机质含量反演研究[J].土壤学报,2020,57(3):636-645. DOI:10.11766/trxb201904160060
GUO Jiaxin,ZHAO Xiaomin,GUO Xi,XU Zhe,ZHU Qing,JIANG Yefeng.Inversion of Organic Matter Content in Red Soil Based on PLSR-BP Composite Model[J].Acta Pedologica Sinica,2020,57(3):636-645. DOI:10.11766/trxb201904160060
【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 215次   下载 140 本文二维码信息
码上扫一扫!
分享到: 微信 更多
基于PLSR-BP复合模型的红壤有机质含量反演研究
国佳欣1,2 , 赵小敏1,2 , 郭 熙1,2, 徐 喆1,2 , 朱 青1,2 , 江叶枫1,2
1. 江西农业大学国土资源与环境学院,南昌 330045;2. 江西省鄱阳湖流域农业资源与生态重点实验室,南昌 330045
摘要:
对红壤地区土壤有机质进行快速预测,以满足智慧农业与精准施肥的需要。以江西省奉新县北部为研究区域,采用1 km × 1 km标准格网划分研究区进行采样,共得到红壤样本248个。对土壤光谱进行了包含分数阶导数在内的3种数学变换方法,将经过P=0.01显著性检验的波段用于模型的构建,选用偏最小二乘回归(PLSR)和BP神经网络建立土壤有机质含量预测模型。结果表明:当对红壤光谱数据进行1.5阶导数变换后再使用PLSR-BP复合模型对土壤有机质含量进行预测时的结果为最优,训练集R2=0.89,RMSE=4.68 g∙kg-1,验证集R2=0.87,RMSE=5.55 g∙kg-1,RPD=2.75。1.5阶导数对红壤光谱数据的变换能够更好地突出与有机质相关的特征信息,有助于其含量预测。PLSR-BP复合模型预测精度优于单一模型,能够较好地预测红壤有机质含量,为精准农业快速监测红壤有机质含量提供了新的途径。
关键词:  红壤  有机质  分数阶导数  偏最小二乘回归  BP神经网络
基金项目:国家重点研发计划项目(2017YFD0301603)、江西省赣鄱英才“555”领军人才项目(201295)
Inversion of Organic Matter Content in Red Soil Based on PLSR-BP Composite Model
GUO Jiaxin1,2,ZHAO Xiaomin1,2,GUO Xi1,2,XU Zhe1,2,ZHU Qing1,2,JIANG Yefeng1,2
1. College of Land Resources and Environment,Jiangxi Agricultural University,Nanchang 330045,China;2. Key Laboratory of Poyang Lake Watershed Agricultural Resources and Ecology of Jiangxi Province,Nanchang 330045,China
Abstract:
【Objective】 The purpose of this study is to explore how to rapidly predict soil organic matter in red soil so as to meet the needs of smart agriculture and precision fertilization. 【Method】 This paper took the northern part of Fengxin County in the Northwest Jiangxi Province as its research area and used the 1 km×1 km standard grid method to divide the study area for soil sampling. A total of 248 red soil samples were collected and dried for spectral measurement. Three different mathematical transformation methods, including fractional order derivatives, were used to analyze the soil spectra. In the tests the 350~399 nm and 2 451~2 500 nm bands were removed because they were very susceptible to environmental noises. And noises in the remaining bands were removed with Daubechies(DB) wavelet. Then samples were collected from the pretreated spectral bands at 10 nm intervals to form a 205-band so as to reduce data dimensions and data redundancy. The 800~1 000 nm band, which was liable to the impact of iron oxide, was ruled out of the experiment. The bands used to construct the model were filtered by the P=0.01 significance test. A model was built up with the partial least squares regression (PLSR) in combination with BP neural network for prediction of soil organic matter content. And the model was tested. 【Result】 Results show that the prediction using the PLSR-BP composite model was the best after the soil spectral data was transformed with the 1.5 order fractional derivative, with R2=0.89 and RMSE=4.68 g∙kg-1 for the training dataset and R2=0.87, RMSE=5.55 g∙kg-1 and RPD=2.75 for the validation dataset. 【Conclusion】 The transformation of red soil spectral data with the 1.5 order fractional derivative better highlights characteristics of organic-matter-related information, which is helpful for prediction of organic matter contents. And the PLSR-BP composite model is higher than any single models in prediction accuracy, and can be used to predict organic matter content in red soil very well. So it can also serve as a new approach to predicting quickly organic matter content in red soil for precision agriculture.
Key words:  Red soil  Organic matter  Fractional order derivative  Partial least squares regression  BP neural network