摘要
氮、磷是维持草原生态健康的重要元素,其含量变化及分布特征,对草原放牧、草地生态管理具有重要的意义。以内蒙古锡林郭勒草原矿区为研究对象,采用多元逐步回归筛选与N/P比最为密切的高光谱波段,建立了基于高光谱反演草原矿区表层土壤N/P的BP神经网络模型。结果表明:1)通过多元逐步回归筛选出的70个波段,其中包括2个可见光波段和68个红外光波段;2)BP神经网络模型随着隐含层层数的增加,训练数据和测试数据的拟合优度先增大后减小,且隐含层层数为3时最大,分别为R2tr=0.872 7和R2t=0.884 2(P<0.000 1);训练数据均方根误差先减小后增大,隐含层层数为3时最小,RMSEtr=0.060 0,测试数据均方根误差在隐含层层数为3时,RMSEt=0.088 2;3)节点数为6、3和10的3层隐含层BP神经网络模型在拟合和预测土壤N/P时的效果较好,可用于草原表层土壤N/P的快速预测。
Nitrogen and phosphorus are among the most important elements to sustain the health of grasslands.Changes in their distribution characteristics are vitally significant to grazing and ecosystem management.In this study,Xilingol grassland mining area in Inner Mongolia was selected as the area of interest.The closest hyperspectral bands with N/P ratio in the surface soil of the grassland were filtered using multiple stepwise regression,and a BP neural network based on hyperspectral inversion of this ratio was established.Results indicated:1)70 bands were selected through multiple stepwise regression that included two visible wavelengths and 68 infrared lights;2)with an increase in the number of hidden layers of the BP neural network model,the goodness-of-fit of the training and test datasets increased and then decreased.When the number of hidden layers was 3,the maximum values reached were R 2 tr=0.872 7 and R 2 t=0.884 2(P<0.000 1);The root mean square error of the training data decreased at first and then increased.When the number of hidden layers was 3,it reached minimum values of RMSE tr=0.060 0 and RMSE t=0.088 2;3)The BP neural network with 3 hidden layers,comprising of 6,3,and 10 nodes,provided a good fitting and forecasting of the soil N/P ratio and can be used to predict this ratio effectively in the surface soil of grasslands.
作者
卢志宏
刘辛瑶
常书娟
杨胜利
赵薇薇
杨勇
刘爱军
LU Zhihong;LIU Xinyao;CHANG Shujuan;YANG Shengli;ZHAO Weiwei;YANG Yong;LIU Aijun(Tongren College,Tongren 554300,Guizhou,China;Inner Mongolia Agricultural University,Hohhot 010018,Inner Mongolia,China;Inner Mongolia Institute of Grassland Survey and Planning,Hohhot 010051,Inner Mongolia,China)
出处
《草业科学》
CAS
CSCD
2018年第9期2127-2136,共10页
Pratacultural Science
基金
内蒙古自治区自然科学基金"基于高光谱的草原矿区土壤光谱特征与重金属定量反演研究"(2017BS0310)
内蒙古自治区自然科学基金"典型草原退化植被的光谱特征研究"(2015BS0321)
内蒙古自治区科技创新引导奖励资金"天空地协同草原生态监测及预警体系构建与示范"(20170519)
关键词
高光谱
多元逐步回归
BP神经网络
N/P比
hyperspectral estimation
multivariate statistical analysis
BP neural network
N/P ratio
作者简介
卢志宏(1982-),男(蒙古族),内蒙古赤峰人,讲师,博士,主要从事生态学研究。E-mail:luzhihong305@163.com;通信作者:杨勇(1984-),男,内蒙古乌兰察布人,助理研究员,博士,主要从事草原生态修复和草原生态监测研究。E-mail:yangyong606@gmail.com;通信作者:刘爱军(1965-),女,内蒙古赤峰人,研究员,博士,主要从事草原生态监测研究。E-mail:liuaj_81@163.com。
