10:45 AM - 12:15 PM
[STT39-P13] Reduction of error components in ALOS-2 InSAR analysis-Application of Noise2Noise deep learning-
Keywords:InSAR, Noise2Noise, Machine Learning
Mapping surface displacement is important for monitoring subsidence, earthquake, or volcanic activity. InSAR analysis estimates surface displacement from the phase difference of microwave observed by a satellite. The estimated displacement by InSAR analysis contains error components, such as the phase delay in troposphere and decorrelation noise. However, existing methods have sometimes had difficulty in completely eliminating these noises. Recently, a new method called Noise2Noise deep learning has been proposed in the field of image analysis. Unlike conventional deep learning, this new method uses only noisy images to estimate noise distribution of images, and it has been shown its ability to remove noise with high accuracy when using photographs. In this study, we examined the effectiveness of the Noise2Noise deep learning for the noise removal of InSAR surface displacement images using pseudo and observed differential interferograms.
For pseudo differential interferograms, we first simulated tropospheric noise and decorrelation noise to surface displacement, and then added these noise components. We assumed a circular subsidence for surface displacement, and used a mathematical model of fractal surface for tropospheric noise. For decorrelation noise, a random noise followed by Gaussian distribution was used. We then applied Noise2Noise deep learning to the constructed pseudo images. In the analysis using pseudo images, we changed the number of images used in the learning and compared the accuracy of noise removal using peak signal to noise ratio and root mean square error (RMSE). Subsequently, we applied the Noise2Noise deep learning to the ALOS-2/PALSAR-2 images taken for Kujukuri, Chiba, where ground subsidence has been reported owing to gas and groundwater pumping. We used 30 SAR data observed from January 15, 2015 to September 30, 2021 for InSAR analysis and made all 435 pairs of differential interferograms. We finally applied Noise2Noise deep learning to these differential interferograms.
In the analysis of 3000 pseudo images, the RMSE between image without noise and processed image for noise removing was 4.12 mm. Considering general accuracy of InSAR analysis, we concluded that the RMSE value of this study suggested that both tropospheric noise and decorrelation noise were successfully removed. In the analysis of the interferograms for Kujukuri area, we successfully removed the decorrelation noise components. However, the displacement components in area with little noise were also removed at the same time. We considered that this is owing to the fact that the learning was performed based on two images that did not have the same level of displacement components in the deep learning process. In order to improve the accuracy of noise reduction, we will correct the noise components caused by the ionosphere at the preparation stage of the training image. In conclusion, this study demonstrated the effectiveness of the Noise2Noise deep learning for the InSAR analysis from the result of pseudo images and real images.
For pseudo differential interferograms, we first simulated tropospheric noise and decorrelation noise to surface displacement, and then added these noise components. We assumed a circular subsidence for surface displacement, and used a mathematical model of fractal surface for tropospheric noise. For decorrelation noise, a random noise followed by Gaussian distribution was used. We then applied Noise2Noise deep learning to the constructed pseudo images. In the analysis using pseudo images, we changed the number of images used in the learning and compared the accuracy of noise removal using peak signal to noise ratio and root mean square error (RMSE). Subsequently, we applied the Noise2Noise deep learning to the ALOS-2/PALSAR-2 images taken for Kujukuri, Chiba, where ground subsidence has been reported owing to gas and groundwater pumping. We used 30 SAR data observed from January 15, 2015 to September 30, 2021 for InSAR analysis and made all 435 pairs of differential interferograms. We finally applied Noise2Noise deep learning to these differential interferograms.
In the analysis of 3000 pseudo images, the RMSE between image without noise and processed image for noise removing was 4.12 mm. Considering general accuracy of InSAR analysis, we concluded that the RMSE value of this study suggested that both tropospheric noise and decorrelation noise were successfully removed. In the analysis of the interferograms for Kujukuri area, we successfully removed the decorrelation noise components. However, the displacement components in area with little noise were also removed at the same time. We considered that this is owing to the fact that the learning was performed based on two images that did not have the same level of displacement components in the deep learning process. In order to improve the accuracy of noise reduction, we will correct the noise components caused by the ionosphere at the preparation stage of the training image. In conclusion, this study demonstrated the effectiveness of the Noise2Noise deep learning for the InSAR analysis from the result of pseudo images and real images.