Unsupervised Techniques for Change Detection in Multispectral Images

Abstract

The changes on earth’s surface are increasing continuously. Various factors are responsible for these changes including anthropogenic and non-anthropogenic. These transformations or changes affect biogeochemical cycles and energy balance of the earth, which causes the climate change. Continuing change in earth’s surface and their effects in climate change motivate to monitor the surface of earth frequently at a global scale. Moreover, timely and accurate observation of these changes is extremely desirable for understanding the interactions and relationships between the natural phenomena and human, which helps to manage resources in a better way. However, monitoring through field survey is time consuming and it is also inconvenient when the changes occurs for very large areas or at inaccessible places. Nowadays, advanced remote sensing technologies provide great opportunity to monitor earth’s surface effectively. Regular acquisition of remotes sensing images over the same geographical regions provides valuable information on earth’s surface dynamic. However, manual monitoring of these images is not feasible. Therefore, development of automatic techniques are essential, which can monitor land surface transitions with minimal human involvement. To detect changes, many change detection techniques have been introduced, which analyze the multitemporal remote sensing images to generate the change map. This thesis addresses detection of changes on land surface by analyzing the multispectral images. The main objective of this thesis is the development of unsupervised change detection techniques using multispectral images for binary detection of changes. Here, different unsupervised methods are proposed for detection of changes on earth’s surface. At first, an unsupervised technique is developed by utilizing the Otsu’s thresholding algorithm. By incorporating local threshold for each pixel, this proposed technique avoids the problem of global thresholding that may generate inaccurate results in many cases. Local threshold is calculated by using Otsu’s technique that is applied on concatenated patch of each pixel of multitemporal images. The derived local threshold is combined with an inter-image based threshold to avoid the effects of external factors i.e. different atmospheric and sunlight conditions. This technique achieves significant improvement by utilizing interimage information and inter-local neighborhood information. Next, a novel unsupervised technique based on binary descriptor is proposed to detect changes on multitemporal multispectral satellite images, which makes use of large number of binary comparisons Abstract to improve the performance. In this technique, features are extracted in terms of binary vectors by applying local binary similarity pattern (LBSP) descriptor that has a good resistance to illumination variation. Hence, this descriptor is helpful to tackle the illumination variation problem in the multitemporal satellite images. As binary descriptor based technique works on the patch combinations of multitemporal images, hence, this proposed framework introduces two techniques for binary change detection corresponding to different patch combinations. Computing threshold for each of the local neighborhood pixel through inter-image information improves the performance of this technique significantly. Subsequently, a novel unsupervised technique is introduced to avoid the postprocessing operation, which is generally the limitations of earlier reported change detection techniques. The performance of this technique is improved by simultaneous use of two thresholds and generation of initial indicators. Simultaneously use of two thresholds, which are derived by automatic generated initial indicators, improves the performance significantly. Finally, the thesis focuses on manifold learning based unsupervised technique to deal with the nonlinearity problem of multitemporal images. This method consists of feature extraction and clustering stages. An orthogonal manifold learning technique is proposed here to extract the features, which is able to address the nonlinearity that occurs due to nonnormalizable radiometric differences in multitemporal images. The clustering stage proposes radial basis function based clustering. The proposed orthogonal manifold learning based discriminant features and clustering technique achieves the significant improvement of change detection performance. Experimental results are conducted on multispectral satellite images to demonstrate the effectiveness of all the proposed methods. In this thesis, used datasets are collected from different sensors of Landsat such as Thematic Mapper (TM), Enhanced Thematic Mapper Plus (ETM+), and Operational Land Imager (OLI).