Unsupervised Change Detection of Satellite Images Using Local Gradual Descent

In this paper, we propose a novel technique for unsupervised change detection of multitemporal satellite images using Gaussian mixture model (GMM), local gradual descent, and
$k$
-means clustering. Data distribution of the difference image is first modeled by bimodal GMM with “changed” and “unchanged” components. The neighborhood data around each pixel form a sample and are modified by the so-called local gradual descent matrix (LGDM), values of which are descending from center toward outside. LGDM visits each sample and causes small variations in pixel values of the sample in an attempt to shift the sample toward the correct Gaussian component center in the feature space. Thus, LGDM decides how much modification to the current sample is necessary for true categorization of the current pixel by later
$k$
-means. The motivation behind the proposed approach is twofold. First, a general method that could efficiently explore both local and global changes for unsupervised change detections is needed. Second, unsupervised change detection methods generally use nonsystematic selections of system parameters. Hence, a parameter selection method without using the ground truth image is required for unsupervised methods. The proposed change detection method is tested for both optical and advanced synthetic aperture radar satellite images and compared with the recent works based on the same input set. The proposed method outperforms the others qualitatively and quantitatively.