Object-based change detection of VHR imagery based on extension of various pixel-based methods

Change detection (CD), one of the main applications of multitemporal satellite images, is an indicator that directly reflects changes in human activity. Very high resolution (VHR) multitemporal images are useful in CD on the Earth's surface and have abundant information sources, enabling more precise CD analysis. The CD is largely divided into pixel-based change detection (PBCD) and object-based change detection (OBCD). In this study, various PBCD methods were combined with a segmentation result to conduct the OBCD of VHR images. The used PBCD techniques in this study are change vector analysis (CVA), iteratively reweighted-multivariate alteration detection (IR-MAD), and principal component analysis (PCA) and K-means clustering from which the binary CD results (i.e., change and no-change) were derived. In order to expand PBCD to OBCD, major voting technique was applied to binary CD results within each object of segmented image that was created using eCognition software. In this process, registration noise (RN) was combined to further improve CD accuracy. The accuracy evaluation of the proposed CD method was conducted using manually digitized reference map. According to the accuracy evaluation, OBCD results that were made by combining all PBCD were the most accurate with 0.393 f1-score. The OBCD results generated by CVA on the side showed the lowest accuracy with 0.301 f1-score.