Advanced Deep Learning Methods for Converting Single-Polarized SAR Images to Fully Polarized Representations and Corresponding Freeman-Durden and H/α Feature Decompositions

Document Type : Original Article

Authors

University of Tehran

10.22059/eoge.2024.372687.1147

Abstract

Nowadays, the utilization of satellite and aerial imagery for understanding and interpreting land features holds significant importance across various fields. Consequently, numerous algorithms are employed to enhance image quality. Notably, fully polarized images demonstrate superior capability in land surface feature classification compared to single polarized ones. However, acquiring these images is often challenging due to limitations. Due to the limitations in obtaining these images, it is often difficult to access these images or the results of decomposition algorithms such as Freeman-Durden and H/α ̅ which are applied to fully polarized images. This study utilized the UNet++ deep learning model to reconstruct complete polarized images and evaluate them alongside the results of aforementioned analysis algorithms. Evaluation metrics such as Mean Squared Error (MSE) and Mean Absolute Error (MAE) were employed to assess model performance and compare original and reconstructed images. These findings underscore the considerable potential of deep learning algorithms for image reconstruction, offering a viable solution to the constraints associated with accessing complete polarized images.

Keywords

Main Subjects