Improved Recoloring of SAR Satellite Images

Sukrut Oak, Young Chol Song, Da Sun

Abstract:

Synthetic Aperture Radar (SAR) images, while invaluable for various remote sensing applications, lack the intuitive visual information provided by color. Our project seeks to enhance the quality of SAR image colorization, a process of assigning colors to grayscale SAR images, by employing a combination of image classification, segmentation, and coloring techniques. By utilizing an ensemble of colorization methods and selecting the optimal approach based on an improved understanding of scene structure and the distinct performance strengths of different models, we demonstrate that our method achieves more accurate and realistic colorization of SAR images. Our results underscore the potential of our approach in incorporating more context for SAR imaging in the enhanced visual interpretation and application of various environmental and surveillance tasks.

We utilize a subset of the SEN12MS-CR dateset, focusing on the San Francisco Bay Area. We first classify the SAR images into five terrain categories by using an instruction refined GPT-4o model. We then evaluate four recolorization models (cGAN, CNN, NL, and LR) via NRMSE, SAM, and Q4 metrics on their performance for each terrain category. When testing our terrain-based approach, we use the
optimal model for the terrain category and compare the results of our terrain-based approach to the best performing model presented in Shen et al.