Remote Sensing Image Change Detection and Location Based on Dynamic Level Set Model
Yunkai Liu,Yuanxiang Li,Yongshuai Lu,Jiawei Liu
DOI: https://doi.org/10.1007/978-981-15-1773-0_22
2020-01-01
Abstract:Image change detection is established for extracting changed regions in multiple images of the same scene captured at different times. Recent research has demonstrated that the change detection methodologies using satellite images, such as multi-temporal visible light remote sensing image and synthetic aperture radar (SAR) image, are particularly useful for damage assessment after various disasters, e.g., earthquakes, fires, floods, and landslides. The level set method, because of its implicit handling of topological changes and low sensitivity to noise, is one of the most effective unsupervised change detection techniques for satellite images. The signed pressure force function (SPF) improved the performances of conventional level set methods through including two grayscale parameters, i.e., the average pixel intensity inside and outside the contour, respectively. However, the mean of region pixel intensity is not a good indicator in case that the images are inhomogeneities grayscale, e.g., confused-edge objects in satellite images. In order to address this problem, we propose a novel model, denoted as dynamic SPF (D-SPF) model, which can dynamically learn a discriminative indicator for distinguishing the pixels inside or outside the contour. Specifically, the principle of maximizing entropy between the regions inside and outside the contour is used to learn the K distinguish parameters, which help to guide the segmentation contour to end on the object’s edge. The experiments are conducted on a public satellite image dataset, i.e., ERS, which contain 670 SAR and 670 optical images; each image covers approximately 150 m2 area including forests, lakes, and cities, etc. These images are challenging due to the inhomogeneities of landforms and unknown natural disasters. The experimental results demonstrate that D-SPF model reduces almost 30.4% missed detection rate on the optical images and 41.2% missed detection rate on the SAR images in comparison with SPF and obtains the best detection performances in ERS dataset.