The quality of satellite image matching directly affects the accuracy and reliability of the subsequent block adjustment accuracy, which in turn affects the generation of products such as digital orthophoto maps (DOMs) and digital elevation models (DEMs). The traditional scale-invariant feature transform (SIFT) algorithm based on image gradient features performs poorly in handling nonlinear radiation differences, and existing phase congruency matching methods have difficulty in simultaneously handling nonlinear radiation differences and geometric differences. For example, the radiation-variation insensitive feature transform (RIFT) algorithm has difficulty in handling large scale differences; the histogram of absolute phase consistency gradient (HAPCG) algorithm has a general matching effect for large rotation problems, and the histogram of orientated phase congruency (HOPC) algorithm requires more accurate geographic reference information. There are three challenges for satellite image matching with multiple phases, multiple views, and radiation differences. Traditional Gaussian linear scale space construction methods lead to image edge blur and loss of detail in the process of building image pyramids; traditional phase congruency methods have difficulty in extracting repeatable and robust feature points, and traditional random sample consensus (RANSAC) algorithm often fails to address the high rate of gross errors in the image matching process. We proposed a phase congruency satellite image matching method based on anisotropic filtering, which could further improve the accuracy and number of correctly matched points in satellite images with significant nonlinear radiation differences and geometric differences.

In light of the challenges posed by satellite matching images with varying phases, views, and radiation differences, we proposed a satellite image matching method based on anisotropic filtering and phase congruency. Firstly, anisotropic filtering was used to establish the nonlinear scale space of the image, and then the phase congruency model was used to calculate the maximum moment map at each scale. Secondly, feature points were extracted using the block-based Shi-Tomasi algorithm on the maximum moment map at each scale, and then the Log-Gabor filter was used to establish the amplitude response at multiple scales and orientations and calculate the maximum amplitude response and its corresponding orientation index. Then, in polar coordinates, feature descriptor construction was accelerated based on OpenMP parallel computing, followed by image matching and mismatch elimination. The proposed method further enhances the matching effect of satellite images with significant nonlinear radiation and geometric differences.

Due to significant nonlinear radiation and scale differences between satellite images taken at different time, the matching performance of the SIFT algorithm is poor, and the experimental data in group F does not yield correctly matched point (NCM) pairs. Similarly, the matching performance of the RIFT algorithm is mediocre, as the significant scale differences in the satellite images taken at different times result in fewer point pairs being matched. The matching performance of HAPCG algorithm is better than that of the RIFT algorithm, as it also utilizes a nonlinear scale space construction method, providing a certain level of robustness to scale differences. However, the method proposed in this paper achieves the best matching performance, being able to match a sufficient number of point pairs in agricultural, urban, and mountainous areas. Particularly, for satellite images taken at different time (groups D-F), as shown in Figs. 16-18, the proposed method outperforms the HAPCG algorithm, even when there are certain angular rotation differences between the images. Furthermore, the matching performance of these four matching methods on the six experimental datasets is quantitatively analyzed, including the statistical data for the NCMs and the root mean square error (RMSE), as shown in Table 1.

In response to the poor matching effects of satellite images with multiple phases, multiple views, and radiation differences, we propose a phase congruency image matching method based on anisotropic filtering. By utilizing anisotropic filtering to establish a nonlinear scale space for images, we propose an improved feature descriptor construction method for the phase congruency model and implement feature descriptor acceleration construction based on OpenMP parallel computing. The proposed method has demonstrated significant advantages in terms of NCMs compared with existing matching algorithms, particularly excelling in handling weak texture, repetitive texture, non-coincident time phases, and nonlinear radiation differences. In the future, we will explore how to integrate cutting-edge technologies such as deep learning to further enhance the robustness and applicability of the matching method.