Change detection in multitemporal satellite images using multiscale bilateral filter and sift flow

Abstract

In the past few decade the number of satellites orbiting earth has increased exponentially. These satellites vary in purpose and application ranging from communications, security, and military to climate monitoring and many more. This increment in the number of satellites observing earth's surface led to an abundance in remotely sensed data and satellite images. Furthermore, the recent development of new and advanced technology in the sensor field made images obtained thought satellites more reliable and accurate than ever. High-resolution satellite images provide a level of resolution unparalleled before. Change detection is process to establish and indicate the difference between two items. Multitemporal satellite images are a set of at least two satellite images of the same geographical area taken over different times. This makes change detection in multitemporal satellite images the process of establishing and indicting the difference between a set of at least two satellite images taken over different times. Change detection in remotely sensed images has many applications which range from climate change, disaster prediction, weather monitoring to urban planning and agricultural land change. The field of change detection in satellite images is growing rapidly due to necessity and need for efficient, reliable, fast and accurate change detection algorithms. However despite the abundance of literature facing these issues, truly reliable and accurate change detestation methods are scarce and even when they exist they tend to be slow and complex. The goal of this thesis is to introduce an accurate, robots, fast and unsupervised change detection algorithm. One of the most common methods of change detection are change detection algorithms that are based on the difference image. These algorithms typically use the difference of the two temporal images for further feature extraction and clustering. This makes them simple, fast and usually quite accurate. This makes building a difference image based algorithm a very enticing idea. Hence, the first algorithm proposed here, is an algorithm based on the difference image analysis. The bilateral filter is used to further enhance the difference image. This yields better features that can be used for clustering. The proposed algorithm checks many of the goals that were set in the beginning. It is accurate, fast and unsupervised. However, it underperforms in the robustness category. The problem with difference image based algorithms is that they fail on big real images, they work better on small images has a consistency in its change. This brings us to our second method; this method is based on image matching using dense correspondence. xxiii More specifically SIFT flow algorithm is utilized for change detection. The SIFT flow algorithm is more commonly used in video processing field. It uses scale invariant feature transform (SIFT) to match pixels in any two given images. SIFT flow is highly robust and extremely fast. A way to use this algorithm in change detection instead of its natural use is proposed: assuming that the two images have been previously registered. The SIFT flow between these two images will reflect change. False detections can be curbed by applying a threshold on the SIFT flow intensity. This method is simple in concept but extremely effective in application. The results of using SIFT flow as change detection algorithm proves to be good since it works under unideal conditions rather correctly. However the change map resulting from SIFT flow tents to be a little bit exaggerated and general. The two algorithms proposed so far are a perfect match for each other. The bilateral filter based algorithm finds small detail extremely accurately but underperforms when it comes to large regions of change. The SIFT flow based algorithm on the other hand locates the big areas of change correctly but struggles to find small details. This makes them a perfect match to each other, by applying SIFT flow first we can provide the bilateral based algorithm with the kind of data it likes, bordered areas of consistent change. By combining the two algorithms we can find accurate results on any scale of an image under any circumstance with high speed and efficiency, in other words we take the best of two worlds. The bilateral filter is first tested by presenting artifact to an image. The resulting image is considered the changed image. Then, the bilateral based algorithm is applied over these two images. The result is quantified by comparing the obtained change map to the artifact. Results using this approach show that the bilateral filter based algorithm performs much better than its peers in every aspect. Later, the SIFT flow algorithm and bilateral based algorithm are tested on more realistic data set. The results are presented in the form of change and heat maps. Change maps of the bilateral filter based method are rich in detail, however, they lack accuracy over large areas. SIFT flow based change maps are accurate but lack enough detail. On the other hand, the change maps obtained by combining the two algorithms are both rich in details and accurate.