2:15 PM - 2:30 PM
[U03-03] A Destriping Method for Infrared Remote Sensing Images Based on Gaussian Mixture Model
Keywords:HY-1E, PMRIS, Destriping, GMM
The new-generation ocean color observation satellite (HY-1E), launched in November 2023, is equipped with a Programmable Medium Resolution Imaging Spectrometer(PMRIS), which composed of five camera sub-modules, each with a field of view of 14.4 degree. The five modules employ field-of-view stitching for push-broom imaging. Its characteristics of wide swath and programmable channels enable it to be widely applied in various fields such as ecological diversity monitoring and disaster assessment. The channel mode can effectively monitor indicators such as chlorophyll concentration, suspended substance, and oil spills; meanwhile, the hyperspectral mode can classify and identify phytoplankton and harmful algae.However, in infrared remote sensing mode, the non-uniform response of the detector leads to noticeable stripe noise in the images, significantly degrading image quality and affecting subsequent quantitative research. To address this issue, this paper proposes an improved moment matching de-striping method based on the Gaussian Mixture Model (GMM), which effectively removes stripe noise from the infrared images captured by the HY-1E PMRIS.
Compared to traditional Moment Matching (MM), Histogram Matching (HM), and Wavelet-Fourier Joint Filtering (WFFT) methods, the proposed method in this paper achieves superior performance in corrected images, with the maximum stripe coefficient of all detection elements being better than 0.0857%, the average stripe coefficient better than 0.0019%, and the median stripe coefficient better than 0.0025%. In terms of both visual quality and quantitative metrics, the proposed method outperforms the three compared algorithms. Experimental results demonstrate that the proposed method effectively enhances the radiometric quality of the infrared images captured by the HY-1E PMRIS, enabling it to better serve quantitative inversion applications of ocean color parameters.
Using the infrared images from the HY-1E PMRIS as experimental data, the proposed method was compared with traditional MM, HM, and WFFT. The results are shown in Figure 1. Although traditional MM and HM can effectively remove stripes, they simultaneously alter the grayscale information of the original data, causing severe distortion of ground objects (red circles in Figure 1(b) and (c)). This is because the grayscale statistical information of each column in the data differs significantly from that of the entire dataset, making it difficult to satisfy the prerequisite assumptions of traditional MM and HM. The result of the WFFT method exhibits noticeable banding artifacts (indicated by arrows in Figure 1(d)). In contrast, the proposed method treats the grayscale distribution of local window regions as a reference, effectively preserving the radiometric characteristics of the original image.
Figure 2 shows the column mean curves corresponding to the results in Figure 1. The column mean curve of the original noisy image exhibits significant fluctuations. By comparing the column mean curves of the results processed by different methods, it can be observed that the column mean curve of the proposed method effectively eliminates the abnormal fluctuations present in the original data.
Figure 3 presents the scatter plot of stripe coefficients, revealing that both HM and the proposed method significantly reduce the stripe levels compared to the other two methods. Table 1 provides the statistical metrics of the stripe coefficients. The proposed method achieves a maximum stripe coefficient of less than 0.0857%, an average stripe coefficient of less than 0.0019%, and a median stripe coefficient of less than 0.0025% for all detection elements in the processed image. These metrics collectively demonstrate excellent performance, indicating that the proposed method effectively removes stripe noise.
Compared to traditional Moment Matching (MM), Histogram Matching (HM), and Wavelet-Fourier Joint Filtering (WFFT) methods, the proposed method in this paper achieves superior performance in corrected images, with the maximum stripe coefficient of all detection elements being better than 0.0857%, the average stripe coefficient better than 0.0019%, and the median stripe coefficient better than 0.0025%. In terms of both visual quality and quantitative metrics, the proposed method outperforms the three compared algorithms. Experimental results demonstrate that the proposed method effectively enhances the radiometric quality of the infrared images captured by the HY-1E PMRIS, enabling it to better serve quantitative inversion applications of ocean color parameters.
Using the infrared images from the HY-1E PMRIS as experimental data, the proposed method was compared with traditional MM, HM, and WFFT. The results are shown in Figure 1. Although traditional MM and HM can effectively remove stripes, they simultaneously alter the grayscale information of the original data, causing severe distortion of ground objects (red circles in Figure 1(b) and (c)). This is because the grayscale statistical information of each column in the data differs significantly from that of the entire dataset, making it difficult to satisfy the prerequisite assumptions of traditional MM and HM. The result of the WFFT method exhibits noticeable banding artifacts (indicated by arrows in Figure 1(d)). In contrast, the proposed method treats the grayscale distribution of local window regions as a reference, effectively preserving the radiometric characteristics of the original image.
Figure 2 shows the column mean curves corresponding to the results in Figure 1. The column mean curve of the original noisy image exhibits significant fluctuations. By comparing the column mean curves of the results processed by different methods, it can be observed that the column mean curve of the proposed method effectively eliminates the abnormal fluctuations present in the original data.
Figure 3 presents the scatter plot of stripe coefficients, revealing that both HM and the proposed method significantly reduce the stripe levels compared to the other two methods. Table 1 provides the statistical metrics of the stripe coefficients. The proposed method achieves a maximum stripe coefficient of less than 0.0857%, an average stripe coefficient of less than 0.0019%, and a median stripe coefficient of less than 0.0025% for all detection elements in the processed image. These metrics collectively demonstrate excellent performance, indicating that the proposed method effectively removes stripe noise.