COMPUTER VISION APPLIED TO DIGITAL SEISMOLOGICAL DATA

Abstract

Recent advancements in remote sensing technology suggest that satellite based earth observation has great potential for providing and updating very high resolution spatial information timely and efficiently. Damage detection and assessment. before and afier disasters have attracted significant attentions among researchers and practitioners of disaster management. Earthquakes are one of the most devastating natural hazards on the earth that can neither be predicted nor be prevented by today's technology. However, the extent of the damage caused by such a catastrophic event can be efficiently assessed in order to reduce the losses of lives and various infrastructures. Change detection is one of the most preliminary steps in various computer vision tasks and widely used technique in remote sensing to detect earthquake or tsunami induced changes. There are basically two main approaches for change detection, supervised and unsupervised. The supervised method of change detection requires a ground truth in order to derive appropriate training set for learning of the classifiers. However, obtaining ground truth is a tedious, time consuming and expensive task. Whereas, the unsupervised change detection is performed by finding absolute differences between two images which has the advantages of not requiring any ground truth data therefore it is easy. fast and less expensive. However, the main problems lie on the fact that the difference image is further binarized by thresholding to classify the objects and the background Thresholding is usually performed empiricall\ or manually by trial-and-error basis which affects both the accuracy and reliability of the change detection process. Therefore. proper selection of threshold is critical which should not be either very high or too low. Various noises will be detected if the value of threshold is too low, while structural changes will be neglected when, the threshold is too high. The correct threshold value relies mainly on the factors. such as noise and illumination which can't be controlled manually and severely aflCcts the structure or geometry of the objects in the change detection process. Further, there exist major limitations in dealing with very high spatial resolution optical remote sensing satellite images as they are huge datasets (several terabytes in size) and highly complex hence, highly redundant. In turn, cost increases to very high value to stole, process and transmit such huge data sets to various stations due to limitations of memory and bandwidtli iv The high complexity is mainly due to presence of vegetation, clouds, shadows, occlusion and misregistration of the very high resolution remote sensing images. A large number of theories have been documented in various research volumes to alleviate such problems. One of the solutions to such problems is edge detection. Edges are to a certain degree invariant to change in illumination and viewpoint. Besides providing good spatial accuracy edge detection leads to significant reduction of image data. In addition, they are convenient from computation (or, processing). transmission and storage viewpoint. Hence. edge maps are more robust for feature extraction than intensity images for change detection under varying illumination. Therefore, objectives of this research work are mainly to study and propose the edge based techniques for unsupervised change detection by the use of very high spatial resolution satellite imageries in order to detect change and to estimate damage caused by an earthquake! tsunami. in this research work, two edge detection techniques have been proposed on the basis of entropy optimization algorithms. The first algorithm has been proposed based on fuzzy logic concept which utilizes the maximization of conditional entropies. The proposed fuzzy based algorithm overcomes some of the limitations of most basic classical algorithms for example Sobel. Roberts. Prewitt. LoG and Canny. The proposed fuzzy based technique has the advantages of being less complex and faster response as compared to some of the standard fuzzy techniques in literature for example. Pal and King algorithm. However, the noise and illumination variations are not balanced out fully by the first technique. Therefore, another technique has been proposed which is being robust against such problems. The second algorithm utilizes the concept of independent component analysis which is based on the optimization of negentropy by FastICA algorithm which inherently utilizes intensity or contrast normalization process as the preprocessing step by centering and whitening (or sphering) the data Lising principle component analysis which overcomes the limitation of illumination variation of the first algorithm. Another limitation of the first algorithm for noise removal has been overcome to higher extent by second algorithm by using the concept of sparse coding. instead of using change detection based on absolute differences between the two imageries this research ork proposes edge differencing based change detection techniques corresponding to pre- and post-earthquake/ tsunami disaster images. The direct comparison between the two edge maps is possible only when the two images are exactly same. However, practically this is not possible due to the presence of noise. varying illumination, shadows. CIOLIdS. occlusion and misregistration on either imagery. Therefore, images have been made free from such problems befbre edges are extracted from the to imageries. The proposed technique first preprocesses the satellite images to remove any sensor noise by median filtering which has the ability to retain edges whilst suppressing the noise. Then. the images have been decorrelated in order to obtain the infbrrnation regarding high reflectance chlorophyll which helps us to sharply highlight the manmade and natural features. The effect of vegetation, clouds. shados of clouds or hills and uneven illumination have been suppressed by using the Normalized Difference Vegetation Index (NDVI) to decorrelated images. By thresholding the NDVI images, spatial distribution of vegetation has been clearly located and extracted out with rest of the features such as bai'e soil, water bodies, buildings. bridges, road networks, canal banks, water tanks etc. Then, the multitemporal images have been fine registered pixelwise by using monornodal image registration technique to make the two images exactly similar. The edges of the pre- and post earthquake! tsunami images have been obtained individually by the proposed algorithms. Then, both the edge images have been registered pixelwise by using monomodal edge registration technique in order to establish match/ difference between them. A match between edge pixels (segments or connected components) have been determined from the two edge maps depending on overlap between them either, partially or fully Mismatched components have been considered as changed regions. In order to estimate change/ damage quantitatively, this research work also proposes some statistical measures based on match' difference between edge densities of the two edge maps The accuracy of the proposed fUzzy based technique has been determined by comparing the performance to that of the Pal-King's technique. The performance measures based on similarity! disparity of two edge maps have also been carried out using modern qua!it\ perf'orniance measure techniques such as'F.dge Based Sirnilarit\ Index (ESSIM)' and 'Feanuic Based Similarity Index (FSIM' FSlM) The very high resolution DigitalGlohe (QuickBtrd) and (eoFye (lKONOS) optical satIIiL imageries have been used as data sets to study Banda Aceh and Gleebruk areas of Surnati,i (Indonesia) which had experienced vast damage caused by a powerful earthquake followed h\ a terrifying tsunami on 26 Dec. 2004. The results have been simulated using Matlab programming which demonstrate the suitahilit, effectiveness and superiority of the proposed research work.