A Study of Neural Network Classification of Jers-1/Ops Images
Yuttapong Rangsaneri, Punya Thitimajshima, and Somying Promcharoen
Department of Telecommunications Engineering,
Faculty of Engineering
King Mongkut's Institute of Technoloyg
Ladkrabang, Bangkok 10520, Thailand
E-mail: kryuttha@kmitl.ac.th,
ktpunya@kmitl.ac.th
Abstract
In this paper, the multiplayer perceptron (MLP) neural network using the back-propagation (BP) algorithm is studies for the classification of multispectral images. The network architecture is made up of three layers: the input layer, one hidden layer and the output layer. The number of input nodes is specified by the dimension of the image to be categories. A simulation has been made and the experimental results on JERS-1/OPS image are given in comparison with the Gaussian maximum likelihood classifer.
1.Introduction
The maximum likelihood algorihm based on Gaussian probability distribution functions is considered to be the best classifier in the sense of obtaining optimal classification rate. However, the application of neural network to the classification of satellite image is increasingly emerging. Without any assumption about the probabilistic model to be made, the networks are capable of forming highly non-linear decision boundaries in the feature space and therefore they have the potential of outperforming a parametric Bayes classifier when the feature statistics deviate significantly from the assumed Gaussian statistics.
In this paper one of neural network, the multiplayer perceptron (MLP) model using the back-propagation (BP) algorithm, will be studies to classification the four categories of Japanese Earth Resource Satellite/Optical Sensore (JERS-1/OPS) . The fundamental of this kind of neural network is described in first section. In the second section, the data of JERS-1/OPS is discussed, and we describe about the network and parameter selection in the following section. Finally, the classification result are given and compared to those obtained by the Gaussian maximum likelihood classification.
2 MLP model and BP algorithm
The multiplayer perceptron (MLP) model using the back-propagation (BP) algorithm is one of the well-known neural network classifiers which consist of sets of nodes arranged in multiple layers with connections only between node in the adjacent layers by weights. The layer where the inputs information are presented is known as the input layer. The layer where the processed information is retrieved is called the output layer. All layers between the input and output layers are known hidden layers. A schematic of a layers MLP model is shown in Fig. 1.
Fig 1. Schematic of a 3-layer MLP model.
For all nodes in the network , except the input layer nodes, the total input of each node is the sum of weighted outputs of the nodes in the previous layer. Each node is activated with the input to the node and the activation function of the node [1]. In Fig 2, node computations is shown.
Fig 2. Node computations
The input and output of the node I (except for the input layer) in a MLP mode, according to the BP algorithm [1] [2], is :
Input : Xi = SWijOj + bi (1)
Output: Oi = f(Xi) (2)
Where
W
ij : the weight of the connection from node I to node j
B
i : the numerical value called bias
F : the activation function
The sum in eq. (1) is over all nodes J in the previous layer. The output function is a nonlinear function which allows a network to solve problems that a linear network cannot solve [3]. In this study the Sigmoid function given in eq. (3) is used to determine the output state.
F(Xi) = 1/(1+exp(-Xi) (3)
Back-propagation (BP) learning algorithm is designed to reduce an error between the actual output and the desired output of the network in a gradient descent manner. The summed squared error (SSE) is defined as:
Where p index the all training patterns and i indexes the output nodes of the network.
O
pi and T
pi denote the actual output and the desired output of node, respectively when the input vector p is applied to the network.
A set of representative input and output patterns is selected to train the network. The connection weight Wij are adjusted when each input pattern is presented. All the patterns are repeatedly presented to the network until the SSE function is minimized and the network "learns" the input patterns. Applications of the gradient descent method [3] yields the following iterative weight update rule :
Dwij (n+1) =
h(diOi+
aDwij(n) (5)
Where
D: the learning factor
a: the momentum factor
di: the node error, for output node I is then given as
di = (ti-Oi)Oi(1-Oi) (6)
The node error at an arbitrary hidden node is
For details BP algorithm including derivation of the equation see[1][2].
3 JERS-1/OPS image data
The JERS-1/OPS (Japanese Earth Resource Satellite / Optical Sensors) image data which used to training and testing the classification of the neural network is the area of Chantaburi city, Thailand. This image data consists of three bands: band 1 (0.52-0.60 mm), band 2 (.63-0.69 mm) and band 3 (0.76-0.86mm), and was taken on December 12, 1996. the image size is 256 x 256 pixels. Band 2 of this image is shown in fig. 3.
The aim of the classification with the neural network was to distinguish between the four categories : water, urban, vegetation and bare soil. The result was to be compared with the Gaussian maximum likelihood classification.
Two small sets of pixels were chosen to be the training and the testing sets of both methods (the neural method and the maximum likelihood method) are shown in Table 1.
Fig.3. Band 2 of JERS-1/OPS image.
Table 1. The training pixels and the testing pixels which used to classify the four categories.
| Category |
Train pixels |
Test pixels |
| 1. Water |
217 |
1,023 |
| 2. Urban |
104 |
487 |
| 3. Vegetation |
135 |
469 |
| 4. Bare Soil |
92 |
255 |
| Total |
548 |
2,334 |
