“DRISHTI : SAR Image Colorization System By Using GAN and CNN”

Abstract

This work involves the design of a machine learning–driven, generative model for the colorization of Synthetic Aperture Radar (SAR) images. SAR images, though valuable for remote sensing due to their all-weather and day–night imaging capability, are mostly grayscale and difficult to interpret. To address this, a Multidomain Cycle-Consistency Generative Adversarial Network (MC-GAN) will be developed to translate grayscale SAR images into realistic, terrain-specific colored images. The system integrates mask vectors to represent terrain types and applies multidomain classification loss to ensure correct and consistent colorization across farmlands, deserts, urban, and rural areas. L1 cycle-consistency loss enables the use of unpaired SAR–optical datasets, overcoming the limitations of traditional methods like pix2pix that require paired data. An application framework will provide visualized outputs, enabling easier interpretation for non-experts and supporting applications such as terrain classification, remote sensing data fusion, and visualization. The expected result is more realistic and interpretable SAR imagery, reduced dependency on paired datasets, and enhanced accuracy in remote sensing analysis with maximum efficiency and reliability.