TL-GAN: Transfer Learning with Generative Adversarial Network Model for Satellite Image Resolution Enhancement

Satellite imagery is essential in many sectors, including remote sensing, environmental monitoring, and urban planning. However, the low-resolution of satellite images makes retrieving fine-grained features difficult and limits their usefulness. Several resolution enhancement methods, including iterative algorithms and deep network-based algorithms, were widely employed to overcome this problem. In this paper, a hybrid model based on a deep neural network is designed to overcome the problem of low-resolution satellite images. To improve resolution, the proposed hybrid network combines a Generative Adversarial Network (GAN) and a pre-trained DenseNet model. The leading benefit of the technique is that GAN learns to generate high-resolution images with realistic and visually appealing details through adversarial training, and the incorporated DenseNet model improves the accuracy of the generated GAN images. To validate the effectiveness of the suggested approach, experimental assessments are designed utilizing a variety of satellite image datasets. The proposed hybrid network produced significant improvements in visual quality and detail preservation in super-resolved satellite images. Finally, the algorithm was compared to the state-of-the-arts method for better evaluation, and the experimental findings demonstrated quantitative and qualitative improvements. Experimental result showed that in terms of peak signal to noise ratio and structure similarity index measurement, the proposed technique achieved 36.28 dB and 0.8304 accuracy, respectively.

I. Introduction

Satellite imaging has proven to be a vital source of data in a variety of applications, including remote sensing, climate change modeling, environmental monitoring, weather forecasting, land management, and urban planning by providing valuable insights into Earth’s surface and enabling data-driven decision-making [1]. Satellite images are frequently acquired using remote sensing tools onboard satellites, which record electromagnetic radiation at various wavelengths, including visible, infrared, and microwave, depending on the sensor employed. The obtained data is then processed to generate computerized images of the Earth’s surface or atmospheric conditions. The resolution of satellite imagery varies depending on the sensor and satellite platform used, with high-resolution satellite images preferred for identifying objects and features on Earth’s surface. Capturing high-resolution satellite images is essential for accurate data analysis and interpretation. However, due to a variety of constraints, such as limitations in capturing devices or transmission channels and the costly expense of high-resolution imaging, satellite images are frequently obtained at lower resolutions, compromising their utility and hampered decision-making processes. Image resolution improvement approach, which uses iterative algorithms or deep neural network-based techniques, is commonly employed to overcome these challenges [2]–[4] without altering the existing hardware.