东篱科研大数据发现系统（DRDS）

位置：成果数据库 > 期刊 > 期刊详情页

基于人工神经网络的土壤有机质含量高光谱反演

ISSN号：0564-3929
期刊名称：土壤学报
时间：0
页码：391-397
语言：中文
分类：S151.9[农业科学—土壤学;农业科学—农业基础科学]
作者机构：[1]南京信息工程大学遥感学院,南京210044, [2]中国科学院南京土壤研究所,南京210008
相关基金：国家重点基础研究发展规划项目（G20000779）、国家自然科学基金重大项目（30590381）和江苏省青蓝工程资助
相关项目：生态系统水碳氮循环过程对全球变化的响应与适应机制

关键词：高光谱, 土壤有机质, 逐步回归, 神经网络, Hyper-spectrum , Soil organic matter （SOM） , Multiple linear stepwise regression （MLSR） , Artificial neural network （ANN）

中文摘要：

研究了土壤有机质含量与土壤高光谱之间的关系，在对原始光谱进行了预处理分析后，运用多元线性逐步回归法（MLSR）和人工神经网络法（ANN）建立了土壤有机质含量的反演模型，并对模型进行了验证。结果表明：人工神经网络所建立的反演模型普遍优于回归模型，网络集成模型优于单个BP网络模型，网络集成是提高反演模型准确性与稳定性的有效途径。网络集成模型为最优模型，总均方根误差为1．31，可以用于土壤有机质含量的快速测算。

英文摘要：

Abstract Historically, soil quality and function used to be assessed through routine soil chemical and physical analysis in the lab. Standard procedures for measuring soil properties are rather complex, costly and time-consuming. A rapid economical soil analytical technique is needed as there is a great demand for larger amounts of good quality, inexpensive soil data available for use in environmental monitoring, modeling and precision agriculture. In this paper possibility of predicting soil organic matter （SOM） content from measured reflectance spectra is studied using multiple linear stepwise regression （MLSR） and artificial neural network （ANN）. After pre-processing of the primitive spectrum, some hyper-spectral models for predicting SOM are built up with the aid of MLSR and ANN, and verified by a validation set. Performance of these two adaptive methods is compared in order to examine linear and non-linear relationship between soil reflectance and SOM content. Results show that to a certainty, both methods have some potential for application in estimating SOM. Performance indexes from both methods suggest ANN models are better than regression models, and the BP integrated model is better than the single BP model. Integrating the ANN subnets is a valid method for improving accuracy and stability of SOM retrieval. The ANN integrated model with the root mean square error （RMSE） of 1.31 is the best model in this research, which can be used in rapid acquisition of SOM content.

同期刊论文项目