DETECTION OF TSUNAMI INDUCED CHANGES FROM SATELLITE IMAGES USING NEURAL NETWORKS

Abstract

A niajor threat to human lives and property are natural disasters. With the increasing urbanization and environmental degradation, the vulnerability to such events has increased manifolds in the recent years. One of the most catastrophic natural disasters is tsunami. Tsunamis are wave train, or series of waves, generated in a body of water by an impulsive disturbance that vertically displaces the water column. Tsunami is an unavoidable event and it affects large areas including both the local coast as well as distant coasts causing enormous amount of daniage to both human lives and property. Proper preparedness and effective disaster management helps in reducing the extent of damage and destruction caused due to tsunami. Disaster management cycle consists of preparedness, mitigation and recovery phase. All these phases require correct information about the disaster struck area. This information is used in recovery tasks like development of evacuation plans for the population, developing route maps etc. Prior to the availability of remote sensing data the only source of information for disaster planners was field surveys. The limitation of these surveys is that they involve a large amount of effort and time causing delay in information delivery. This delay in information leads to increased amount of fatalities and losses as the recovery tasks are delayed. Satellite images have proved to be extremely useful in recovery and rehabilitation of the disaster prone area. In this thesis, three novel methods using medium resolution satellite images and neural networks are developed for the identification of tsunami induced changes. Neural network based methods perform 1 more accurately as neural networks are useful when the feature space is complex as in case of satellite images. Also, these methods work rapidly due to their concurrent operation. The change information obtained from these methods will be helpful for emergency response and recovery operations. The first method presents, a supervised change detection method based on Probabilistic Neural Network (PNN) and normalized neighborhood ratio. The PNN based method has a fast training process and guaranteed optimal classification performance. The proposed method first produces a difference image using a novel normalized neighborhood ratio (NNR) approach. The NNR overcomes the limitations of the conventional methods used for computing the difference image and is less sensitive to speckle noise. This difference image is then fed as input to the PNN. The estimator function of the PNN finds the probability density functions (pdf), the pdfs are used to compute likelihood ratio. The pixels of the difference image are then assigned to change and unchanged classes based on the log likelihood ratio test. This results in a change map highlighting changed and classified images are then used to identify the changes that have occurred. The pre and post tsunami images are classified into three classes including water, vegetation and urban area. The image showing changes shows three extra class representing change information.Pre and post tsunami images of a 60 km long coastal strip of Japan including four cities Soma, Watari, Natori and Iwanuma are used. Landsat 7 ETM + pre tsunami image and EOI ALl post tsunami image acquired on 9Ih May 2003 and 291h March 1, 2011 respectively are used for the study. The accuracy assessment shows that the results are quite satisfactory.