Two-stage fusion strategy and residual shrinkage capsule network for SAR image change detection

ABSTRACT

Acquiring precise change map for synthetic aperture radar (SAR) images poses challenges due to the absence of labelled datasets and the presence of speckle noise interference. This paper presents a two-stage fusion strategy coupled with capsule network classification detection method. Initially, to enhance the structure of change areas within the difference image (DI), we propose an energy ratio (ER) algorithm based on the mean ratio (MR). Subsequently, employing a two-stage fusion strategy integrating multiplication and discrete wavelet transform (DWT), we fuse the DIs generated by the three algorithms to effectively complement change information across distinct DIs. In addition, a saliency detection algorithm is introduced to mitigate noise interference and balance the number of samples. Hierarchical fuzzy c-means clustering (HFCM) is then used to pre-classify the salient fusion difference image (SFDI) and derive pseudo-labels. Finally, we apply the residual shrinkage capsule network (RSCN) to reclassify pre-classification results and obtain the change detection outcome, capable of extracting the multi-scale refinement features and adaptively learning discriminant information from changed and unchanged regions. Experimental evaluations on various SAR image datasets demonstrate the exceptional detection performance of the proposed method.

KEYWORDS:

• Synthetic aperture radar (SAR)
• energy ratio (ER)
• two-stage fusion strategy
• residual shrinkage capsule network (RSCN)

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

This work was supported by the National Natural Science Foundation of China (grant number 61901297).