Unsupervised change detection monitoring by feature change extraction using bi-temporal high resolution polarimetric SAR images

Document Type : Original Article


School of Surveying and Geospatial Engineering, College of Engineering, University of Tehran


Synthetic aperture radar (SAR) sensors are microwave active systems which represent a major tool for Earth observation. The completed information lying in the polarimetric channels represents a possibility for better detecting changes in different applications. In the literature, the log-ratio operator is applied to the original SAR image. In this paper, due to the use of full polarimetric images, first the coherency matrix, polarimetric decomposition, segmentation and data analysis features are extracted respectively, then the log-ratio and difference operators are applied to the extracted features. The use of decomposition increases the detection power due to extraction of single, double and volume bounce components. The aim of this work is proposing a framework for change detection in multi-temporal multi-polarization SAR data. In the novel representation, multi-temporal SAR images are employed to compute log-ratio polarimetric features. After pre-processing data, the coherency matrix, polarimetric decomposition, segmentation, and data analysis features are extracted. Then, the log-ratio and difference operators are applied to the features and create change maps using two unsupervised classification methods. The input of unsupervised classification is a stack of log-ratio features. Finally, the t1t2 (changes from epoch1 to epoch 2) and t2t1 (changes from epoch 2 to epoch 1) change maps, that are classification outputs, are fused. This representation is employed to design a novel unsupervised change detection approach for separating an unchanged class and two changed classes. The proposed approach is validated on a pair of UAVSAR data (L-band) acquired in Oakland, California, between the period 2010 to 2017. In the both groups of changes, the t1t2 and t2t1, coherency based feature combination achieves the best result with an overall accuracy of 87% and Kappa of 74%. Considering all changes (both t1t2 and t2t1), coherency based feature combination yields the best result with an overall accuracy of 86% and Kappa of 79%. As is clear from the evaluation results, the log-ratio operator has shown far better results among the two log-ratio and difference operators. However, the best option is the simultaneous use of the both operators so that the noise and error of the log-ratio operator can be reduced using the difference operator. According to the final results, it can be concluded that the coherence matrix is a better feature for detecting changes compared to other features.