(SAR) can actively obtain surface information, has a wide image coverage, is less affected by natural conditions, and can conduct all-weather and all-day ground reconnaissance. Change detection based on SAR images can obtain target change information in the same airspace and different time domains. It plays an important role in both military and civilian fields, and can provide support for emergency and rapid decision-making by relevant national departments. SAR image contains rich multi-dimensional and multi-domain information, and its processing can improve image utilization. With the development of SAR, difference map generation in SAR image change detection plays a key role in subsequent processing. Spatial domain filtering takes into account the correlation between pixels and their neighbors, and directly denoises the space of pixels in the image. The frequency domain low-pass filtering is the operation of the image in the frequency domain, reducing the sharp edge contour part and highlighting the smooth part. The existing difference map construction method mainly focuses on spatial domain filtering, which cannot retain the change information well, and has less consideration for the frequency domain filtering method. In the difference map generation, only a single spatial domain filtering method is used, ignoring the information in the frequency domain of the image. In order to improve the model generalization ability and detection accuracy of SAR image change detection, we propose a SAR image change detection method based on dual-domain filtering.

Firstly, we filter the original SAR image in the spatial domain, and filter the dual temporal SAR image in different ways. We construct a logarithmic ratio operator after the adaptive median filter, and we construct a difference operator after the mean filter. Then, Laplace fusion algorithm is used to fuse the difference map in the spatial domain and synthesize the feature information of different difference operators. Afterwards, the fused image is transformed into the frequency domain for low-pass filtering in the frequency domain. Finally, the change detection result graph is obtained by using clustering algorithm.

In order to verify the effectiveness of the proposed method, four datasets of Bern, Ottawa, San Francisco, and Yellow River are used for experiments. In the hidden line elimination experiment, the difference operator proposed in this paper is used to improve the accuracy of the basic algorithm, which has significantly improved in the objective indicators (Figs. 9-16). It can be seen that the noise points in the detection results after the dual-domain filtering are the least, and the degree of detail in the detection results is well preserved, and the number of missed detections and false alarms are reduced to varying degrees, which are relatively balanced. The results of the two clustering algorithms are close (Fig. 17). At the same time, in order to compare the performance of the registration methods proposed in this paper, we use the existing five algorithms to test on the four experimental datasets in this paper (Fig. 18). The proposed Dual Domain-K method is the best in Bern and Ottawa datasets. Compared with the depth learning method, the accuracy of Dual Domain-K method is lower, but the calculation time cost is greatly reduced, and the accuracy is also guaranteed. Finally, the influence of the experimental parameters manually adjusted in this method on the result indicators is given (Fig. 19, Table 10 and Table 11). Four sets of data from Bern, Ottawa, San Francisco, and Yellow River are used for experimental verification, and the experimental results show the effectiveness of differential images after dual-domain filtering in the clustering.

This paper mainly studies the high noise problem of difference operator in SAR image change detection. The feature representation of difference operator is studied. Under the condition of ensuring certain accuracy, the calculation time of depth learning algorithm is reduced, and the operation efficiency is improved. A new change detection algorithm is proposed, which deals with difference operators in frequency domain. We fuse the features of different operators in the spatial domain, and use Laplace for fusion to retain the features in the spatial domain to the greatest extent. Then, Fourier transform is used to transform the SAR difference operator to the frequency domain for low-pass filtering, and the main part of the transform is retained. Finally, experiments are carried out with real SAR image data, and change detection results with high accuracy are obtained. Several groups of experimental data show that, compared with other methods, the proposed method has strong robustness on different datasets and can quickly generate binary mapping of change detection results.