Comparison of segmentation inputs for object-based unsupervised change detection between very-high-resolution bi-temporal images

Due to a high level of spatial complexity and heterogeneity in very-high-resolution (VHR) imagery, an object-based unsupervised change detection approach instead of a pixel-based approach is generally conducted for VHR imagery. Segmentation methods that subdivide an image into meaningful homogeneous regions and organize them into image objects corresponding to ground entities have been developed. However, few studies have focused on determining common boundaries of objects in bi-temporal images for the change detection. Therefore, we investigated and compared segmentation inputs, with the goal of minimizing inconsistent boundaries between bi-temporal images, to obtain an optimal object-based unsupervised change detection result. For this purpose, simple linear iterative clustering (SLIC) was selected as a segmentation approach due to its computational efficiency. We carried out SLIC-based segmentation for bi-temporal images with five different segmentation inputs: i) T1 image, ii) T2 image, iii) difference image of the bi-temporal images, iv) principal component analysis-based image, and v) intersection image. After defining the segmentation results according to each input, the average pixel value of the band in each segment was calculated and allocated to the segment to process the change detection further. Change vector analysis (CVA) was implemented to conduct the unsupervised change detection. Bi-temporal Kompsat-2 satellite images were used to generate a study site for implementing the experiments. From the experiments, we demonstrated that using one image or using a difference image for generating the segmentation map produces better change detection results than the results obtained by the intersection of two segmentation results.