Vous êtes sur la page 1sur 2

Landslide Detection Using Image Processing and Neural Network

Ganesh H
Abstract Landslides occur in many regions of the

world. They are mainly in mountainous regions. We have a high chance of having a landslide in a region which contains a slope or steep descent. The landslide is caused due to many causes like earthquake, typhoons and other natural calamities. The presence of an automation technique for the regions which have possibility of landslide will help reduce the damage caused due to it. In this paper we concentrate on predicting a landslide region using artificial neural network. The multi-layer perceptron (MLP) model of artificial neural network is used. I. INTRODUCTION

Aerial photography of a region contains vital details of mountains and valley portions. This can be used to extract information on steep portions of the mountain. Landslide occurrence is high in such regions. Traditional method of classifying an area is landslide prominent or not is done by survey of the region. This is time consuming. Artificial neural network (ANN) approach gives a novel solution to this by classifying a region using the satellite image of the region under consideration. ANN offers many methods to solve this problem. I will be considering MLP with back propagation learning algorithm to train the ANN. Learning by a network is accomplished by adjusting the weights to produce appropriate output through training examples fed to the network (Zurada, 1992). MLP type neural network model utilized in this was developed using MATLAB R2010a by Mathworks. III. LEARNING IN MLP

Landslides are a natural phenomenon for the dynamic balance of the surface. The factors for landslide include land topography, climate, earthquake, hydrology and human activities. This occurs in regions where the geology is high fractured. It also depends on frequency of heavy rainfalls which imposes more stress to the earth. Hence there is massive movement of earth such as landslide, clumping, and mudflows. Landslide is common in Himalayas during the rainy season. There loss of life and damage to property. This poses a threat to settlements in that region and infrastructure that needs to be provided. A human can find out the possible areas for landslide occurrence but is time consuming. An artificial neural network implemented by using back-propagation learning rule MLP, has the capability to do the same.

Learning algorithm of ANN includes important parameter of learning rate (alpha or ). In general valid range for alpha is between 0.0 and 1.0. For a small value of alpha, the numerical convergence of the algorithm is better but the training time is greatly increased. For larger values of 0.6 and above there is a risk of divergence. So an optimum value in the range of 0.001 to 0.1 is used. The transfer function for the hidden layer is linear and for the output layer is hardlimitter. The satellite image is converted to grayscale image. The prominent features of the image containing the ridges and valley points are extracted. This is fed to the ANN which separates regions into landslide possible and not possible regions The network is trained using a set of test samples with output known. The weights are adjusted accordingly. The weights are then frozen.

II.

OVERVIEW AND MULTI-LAYER PERCEPTRON(MLP)

Landslides cause approximately 100 deaths a year and property damage of about US$4 billion, and pose threats to settlement and transportation structures. Slope cutting and deforestation leads to landslide in many cases.

This can now perform the required classification of classifying a region to be landslide prominent region or not. The learning curve for this is as shown in diagram below

b Dept. of Civil Eng., NCTU & ERL, ITRI, Hsin-Chu 311, TAIWAN. JKLiu@itri.org.tw [2] Zurada, J. M., 1992. Introduction to Artificial Neural Systems, West Pub. Co. [3]Simon Haykin, Neural Networks,1999, Prentice Hall India Private Limited.

VIII.

BIOGRAPHIES

Ganesh. H is studying B.E at R.V College of Engineering.

IV.

APPLICATION

This finds application in determining whether a region has probability of having landslide from the satellite photograph of the region. It helps reduce time in finding regions with landslide faster than traditional methods. This helps prevent loss to life and damage to property.

V.

CONCLUSION

The high resolution image produces better output compared with low resolution images. The data from only an image is not enough. The inclusion of topographical data along with the image will help to get better output which is on power with that of a human making the same judgement. The learning rates in the range 0.001 to 0.2 finds convergence and learning rate of 0.5 and above have a risk of divergence. The algorithm has a good of scope for improvement.

VI.

ACKNOWLEDGEMENT

I would like to thank Prof. Sivaramakrishnan Narayanan, Professor, R.V College of Engineering for his help in understanding MLP. VII. REFERENCES

[1] LANDSLIDE FEATURES INTERPRETED BY NEURAL NETWORK METHOD USING A HIGH-RESOLUTION SATELLITE IMAGE AND DIGITAL TOPOGRAPHIC DATA K. T. Chang a and J. K. Liu b a Dept. of Civil Eng., MUST, No. 1, Hsin-Hsing Rd., Hsin-Fong, Hsin-Chu 304, TAIWAN. ktchang@must.edu.tw