A three-class change detection methodology for SAR-data based on hypothesis testing and Markov Random field modelling
This study presents a new automatic change detection process chain based on bi-temporal co-registered and calibrated Sentinel-1 level-1 Interferometric Wide Ground Range Detected C-band synthetic aperture radar intensity imagery. The whole processor contains three main components: firstly, a preprocessing step is used to perform geometrical and radiometrical calibration. Secondly, an automatic coarse detection step is applied based on a statistical hypothesis test to obtain an initial classification. Thirdly, a post-classification step is introduced to optimize the initial classification result in the form of minimizing a global energy function defined on a Markov Random Field. In this study, a graph-cut algorithm is applied iteratively to solve the global optimization problem. At each iteration, the data energy function for the current classification is set by the logarithmic probability density function. The relevant parameters are estimated by the method of logarithmic cumulants. Supplemental data is presented to explain the formulae used in this study. Experiments are performed using a flood event which occurred in 2015 along the coastline of Greece near Kavala region and the Evros River at the border between Greece and Turkey. The proposed method shows a satisfying classification result with overall accuracy above 95% and kappa coefficient () above 0.87.