Development of post classification change detection algorithms for multispectral imagery

Abstract

Multispectral image processing is developing technology in the field of remote sensing. A multispectral image is a multi-band image comprising of 4 10 bands, each having different wavelength. These different wavelengths are captured by corresponding sensors. Each band has its own distinct information about the earth s surface. One of the important research areas in remote sensing is change detection of multispectral images. The thesis aims on developing novel algorithm for post classification change detection in the multispectral LANDSAT 7 images captured in 7 bands at different temporal points. It extracts useful information needed for environmental applications. The proposed algorithm has two major stages: Segmentation and Classification. For segmentation, two different methods such as Hybrid Clustering and Proximal Splitting algorithms have been proposed. In classification, Multiclass Support Vector Machine and Game Theory algorithms have been employed. The four possible combinations of segmentation and classification algorithms followed by image differencing method have been applied on multispectral images to find out changes occurred in HANOI and BAOLOC region of Vietnam city. Hybrid Clustering is the combination of Graph cut method and K-means clustering. Initial clustering is formed by K-means method in order to identify widely varying intensities. Then, slowly varying intensity details are preserved by Graph cut method. These details provide additional information about the boundaries between different land cover classes. Hybrid Clustering Based Segmentation method iteratively deforms the contour and it has the ability to jump over local minima and provide more global results. It is well suited for high dimensional data sets especially for multispectral images. newline