Robust approaches for image change detection (ICD) are essential for a range of large-scale applications. However, the uncertainties
involved in such approaches have not been fully addressed. To investigate this problem, this paper proposes a new
approach for change detection from multi-temporal very high resolution (VHR) satellite imagery based on uncertainty detection
and management. First, two GeoEye-1 images of Giza urban area (Egypt), acquired in 2009 and 2019, have been geographically
co-registered and their histograms have been matched. Second, a set of feature attributes have been generated from the coregistered
images. Third, the support vector machine (SVM) algorithm has been adopted to classify the data into four classes:
building, tree, road, and ground. In this regard, the co-registered images along with the generated attributes have been applied as
input data for the SVM to calculate the probability of each pixel belonging to each class. After that, the probability images for
both epochs have been compared to model the uncertainty of changes. The uncertainty places are then evaluated to estimate their
likelihood of being change or no change. Finally, the obtained results have been compared with manually digitized change
detection map. Compared with using the widely used post-classification comparison (PCC) approach, the results suggest that (1)
the proposed method has improved the overall accuracy of change detection by 13%; (2) the class-accuracies have been improved
by 35.63%; and (3) the achieved accuracies for the proposed approach are less variable. Whereas the standard deviation (SD) of
the accuracies obtained for the proposed approach is 6.80, the SD of those obtained for the PCC approach is 35.50. |
