• •

基于RBF神经网络的土壤铬含量空间预测

1. 1.中山大学地球环境与地球资源研究中心, 广东 广州 510275
2.华南农业大学信息学院, 广东 广州 510642
3.广东省土地利用与整治重点实验室, 广东 广州 510642
4.华南农业大学公共管理学院, 广东 广州 510642
• 收稿日期:2012-03-26 修回日期:2012-07-01 出版日期:2013-01-20 发布日期:2013-01-20
• 作者简介:

作者简介：陈飞香（1978-）,女,讲师,博士研究生,研究方向为地理信息系统应用与土地资源。E-mail:chfx@scau.edu.cn

• 基金资助:
国家自然科学基金项目（40971125）、广东省科技计划项目（2011B020313020）资助

Spatial Prediction of Soil Properties by RBF Neural Network

Fei-xiang CHEN1,2,3, Jia-chang CHENG2,3(), Yue-ming HU2,3, Yong-zhang ZHOU1, Yuan ZHAO2,3, Jia-chun YI4

1. 1.Center for Earth Environment and Resources, Sun Yat-Sen University, Guangzhou,Guangdong 510275,China
2.College of Informatics, South China Agricultural University, Guangzhou,Guangdong 510642,China
3.Guangdong Province Land Use and Remediation of the Key Laboratory, Guangzhou, Guangdong 510642,China
4.College of Public Management, South China Agricultural University, Guangzhou, Guangdong 510642, China
• Received:2012-03-26 Revised:2012-07-01 Online:2013-01-20 Published:2013-01-20

Abstract:

The key problem in precision agriculture research is how to use fewer samples to reflect the distribution regular pattern of farmland information and to use scientific interpolation method to interpolate and estimate farmland. Based on the soil chromium (Cr) content areas with large differences in Zengcheng City as an experimental base, 200 soil samples were collected by random sampling method. According to GB/T17137-1997 in China, the flame atomic absorption spectrometric method was used for the determination of a variety of chromium in soil. By Create subsets function in ArcGIS 9.3, four kinds of layout program were set,which were 200 sample points, 150 samples points , 100 samples points and 50 samples points . In accordance with the ratio of 4:1, the four of sample sets were divided into training dataset and test dataset, which were used to train the neural network, testing the accuracy of the interpolation results. Then RBF neural network method were used in soil Cr content interpolatin in the previous three date sets, and their error were analyzed and forecasted using the variance of the RMS (root mean square) error. In order to highlight the contrast, the previous corresponding Kriging interpolation maps and Kriging interpolation results of RMS error were gotten when Kriging interpolation method for data of the corresponding treatment was used. The results show that, in the case of fewer sample points, the interpolation result of RBF neural network was smaller than the root mean square error of traditional Kriging interpolation method, which are 0.003, 0.009 and 0.008, respectively. It was found that RBF neural network method was more accurate. However, when it was applied to the date set which were only 50 samples, whether the RBF neural network or Kriging interpolation method, numerical root mean square error of the prediction results was great, which were 0.179 and 0.128, respectively. Hence, it is hard to obtain the right result. Compared to the traditional method of statistical interpolation, RBF neural network method could overcome the smoothing effect with good self-learning features and strong non-linear computing power. Especially in the case of fewer sample points, the effect of spatial prediction was relatively good. Then, it can be concluded that RBF neural network method is applied more broadly and it is enough when it is used for the interpolation of the data sampling points.

• S153.6+1/TP183