A novel 3D bilateral filtering algorithm with noise level estimation assisted by multi-temporal SAR

The bilateral filter is widely employed in the field of image denoising due to its flexibility and efficiency. It calculates the weights of neighboring pixels based on both spatial and grayscale distances from the pixel to be denoised. By incorporating the information of neighboring pixels through a weighted average, it reduces the disparity between the target pixel and its neighbors, achieving the goal of denoising. However, the extensive imaging range of SAR, coupled with low spatial resolution and the complexity of surface features, results in significant variations in the information expressed by each pixel within the kernel. Consequently, relying solely on neighboring pixel information for denoising can introduce a considerable amount of extraneous data into the target pixel, reducing image contrast and blurring edge contours. Additionally, because the noise levels in pixels of SAR images vary, the uniform filtering approach of the bilateral filter may lead to a degree of information loss in the filtered pixels. Ultimately, while the bilateral filter performs well in addressing additive noise, it is less effective against the multiplicative noise common in SAR images, further diminishing its filtering efficacy. To address these issues, we have developed the 3D bilateral filtering algorithm with noise level estimation assisted by multi-temporal SAR(3D-NLE-BF). This algorithm begins by evaluating the noise content of pixels to be denoised based on their temporal and spatial stability, classifying them into strong noise, weak noise, and noise-free pixels. Given the higher similarity of pixels along the temporal axis in multitemporal SAR data, the algorithm capitalizes on this feature to ensure that denoised pixels contain more useful information. Taking into account the characteristics of multitemporal SAR, the algorithm incorporates range-weight, spatial-weight, confidence-weight, and time-weight, designing corresponding filtering kernels for both strong and weak noise pixels. To verify the superiority of the algorithm, we selected Bilateral, NLM, Kuan, Lee, Lee-Enhanced, and Lee-Sigma as comparison algorithms. Real and simulated SAR denoising experiments were designed, and the denoising results were evaluated using ENL, SSI, PSNR, and QIUI, achieving favorable evaluation results. This demonstrated the effectiveness and general applicability of the algorithm proposed in this paper.