Environment Unsupervised Change Detection using Change Point Analysis in SAR Images

Nowadays, Earth observation (EO) technology became an indispensable tool to help environmental monitoring, as well as their changes, for natural resources management, urban planning and development, water management and land use planning. In particular, radar EOs, unlike the optical ones, can be collected regardless of illumination and weather conditions. Multitemporal polarimetric synthetic aperture radar (PolSAR) images are useful source of information for detection and mapping the environmental changes, especially in wide areas, during the day and night and all weather conditions. Change detection methods can identify the change or no change conditions in land covers using the time series observations. In this paper a method is proposed for change detection in SAR remote sensing images. This method is based on the Change Point Analysis. The cumulative frequency of difference image, which contains the environmental changes, normally follows a specific class of statistical distribution. Gaussian mixture model is one of the most suitable models for Change Point Analysis. This model can efficiently estimate the parameters of mixture distribution. The intersection point of two distributions is a change point, which can be seen as a threshold. This threshold is then used to separate the change and no change classes. The proposed method is implemented and analyzed using three SAR data sets. The analytical evaluations of the final change maps from two of these data sets with reference data had the Kappa coefficients of 90% and 96% respectively. The other data set contained the multitemporal PolSAR images and had been acquired over an agricultural area. The changes in these images were enough reliable to be connected to the agricultural activities, such as crop growing stages and harvesting, based on an available crop map. Finally, the method was evaluated against the Otsu method, as one of the best threshold estimation methods, and the results showed the superiority of the proposed method, e.g. 2% better in term of kappa coefficient. . As a result, the proposed method, can be efficiently employed for land cover change detection and monition in natural resources management.