Retrieval of spinach crop parameters by microwave remote sensing with back propagation artificial neural networks: A comparison of different transfer functions

详细信息	查看全文 | 推荐本文 |

• 作者：Rajendra Prasad^a ; ^{rprasad.app@itbhu.ac.in} ; ^{rprasad1@rediffmail.com} ; A. Pandey^a ; K.P. Singh^b ; V.P. Singh^c ; R.K. Mishra^d ; D. Singh^e
• 关键词：Back propagation artificial neural network ; Transfer function ; Scattering coefficient ; Biomass ; Leaf area index
• 刊名：Advances in Space Research
• 出版年：2012
• 期刊代码：6_02731177
• 类别：et
• 出版时间：1 August, 2012
• 卷：50
• 期：3
• 页码：363-370
• 文件大小：1059 K

摘要

Back propagation artificial natural network (BPANN) is a well known and widely used machine learning methodology in the field of remote sensing. In this paper an attempt is made to retrieve the spinach crop parameters like biomass, leaf area index, average plant height and soil moisture content by using the X-band scattering coefficients with BPANN at different growth stages of this crop. The maturity age of this crop was found to be 45 days from the date of sowing. After 45 days from the date of sowing, this crop was cut at a certain height for production. Then, it is a point of interest to investigate the microwave response of variation in production. Significant variations in all the crop parameters were observed after cutting the crop and consequently made the problem more critical. Our work confirms the utility of BPANN in handling such a non-linear data set. The BPANN is essentially a network of simple processing nodes arranged into different layers as input, hidden and the output. The input layer propagates components of a particular input vector after weighting these with synaptic weights to each node in the hidden layer. At each node, these weighted input vector components are added. Each hidden layer computes output corresponding to these weighted sum through a non-linear/linear function (e.g. LOGSIG, TANSIG and PURLIN). These functions are known as transfer functions. Thus, each of the hidden layer nodes compute output values, which become inputs to the nodes of the output layer. At nodes of output layer also a weighted sum of outputs of previous layer (hidden layer) are obtained and processed through a transfer function. Thus, the output layer nodes compute the network output for the particular input vector. In this paper, output nodes use linear transfer function. Different transfer functions e.g. TANSIG, LOGSIG and PURELIN were used and the performance of the ANN was optimized by changing the number of neurons in the hidden layers. The present analysis suggests the need of critical analysis of the BPANN in terms of selection of the best transfer function and other network parameters for the better results.