Object-Based Change Detection in Urban Areas Using Multitemporal High Resolution SAR Images with Unsupervised Thresholding Algorithms

Abstract

With the recent launches of optical and SAR systems that are capable of producing images in very high resolution, the quantification of temporal changes can be achieved with unprecedented level of details. However, very high resolution data presents new challenges and difficulties such as the strong intensity variations within land cover classes thus the noisy appearance of change map generated by pixel-based change detection. This has led to the development of object-based approaches that utilize image segmentation. For unsupervised change detection, on the other hand, automatic thresholding algorithms provided a simple yet effective technique to produce a binary change map. Thresholding techniques have been used successfully for pixel-based change detection using medium resolution SAR images. They have also been used for object-based change detection using high resolution optical imagery. However, they have not been tested in the context of object-based change detection using high resolution SAR images. Therefore, this chapter investigates the potential of several thresholding techniques for object-based unsupervised detection of urban changes using high resolution SAR images. To avoid the creation of sliver polygons, the multidate image segmentation strategy is adopted to produce image objects that are spectrally, spatially, and temporally homogeneous. A change image is generated by comparing objects multitemporal mean intensities using the modified ratio operator. To threshold the change image and generate a binary change map, three thresholding algorithms, i.e., the Kittler-Illingworth algorithm, the Otsu method, and the outlier detection technique, are tested and compared. Two multitemporal datasets consisting of TerraSAR-X images acquired over Beijing and Shanghai are used for evaluation. Quantitative and qualitative analyses reveal that the three algorithms achieved similar results. The three algorithms achieved Kappa coefficients around 0.6 for the Beijing dataset and 0.75 for the Shanghai datasets. The analysis also reveals the limitation of the mathematical comparison operator in accentuating the difference between the changed and the unchanged class, thus calls for the development of more sophisticated object-based change image generation mechanisms capable of reflecting all types of changes in the complex urban environment.