Algorithm to Estimate Scalloping & Banding in Scansar Images and Deep Learning Based Descalloping Technique

An algorithm is designed and implemented to estimate the scalloping or banding effects in wide swath ScanSAR images and an Autoencoder based deep learning method is attempted to remove the scalloping from the images which is an independent method and do not use SAR imaging parameters. Various SAR images are analysed to demonstrate the effectiveness of the proposed method. We have used both simulations and the Radar Imaging Satellite (RISAT-1) SAR image for computing the depth of scalloping. The estimation is based using periodic variation (inherent pattern of ScanSAR because of scan mechanism and SAR processing) of Scalloping/Banding noise present in ScanSAR images. The estimated depth of scalloping from the simulated dataset is observed around 4.50 dB & 7.11 dB whereas in RISAT-1 SAR data (uncorrected) acquired over the Amazon Rainforest in HH and HV image is 7.29 dB & 8.42 dB respectively. The deep learning-based technique is proposed to remove scalloping and interscan banding noise from an image, especially SAR images. For developing deep learning-based image denoising solution, large-scale training dataset is required, here we propose scalloping dataset. We had simulated 19,125 images with different scalloping noises for training and testing of our Autoencoder neural network model. We have used many images to evaluate the performance of our neural network model. These include some simulated images and original RISAT-1 image. Experiments on RISAT-1 and simulated images shows good performance of our proposed method which can further be improved utilizing realistic datasets.