Exploring Deep Feature Loss in Generative Adversarial Network for Satellite Image Resolution Enhancement

Satellite imaging is an essential tool that is commonly utilized for agricultural development, environmental monitoring, urban planning, and many other applications. However, the poor resolution of satellite images makes it difficult to extract fine-grained features and prevents them from being used effectively. Several techniques, including unsupervised and supervised learning models, have been developed to address the challenges. In this paper, an efficient resolution improvement approach is proposed by tweaking the adversarial loss functions of a Generative Adversarial Network using deep neural network features. A lightweight MobileNet model is incorporated as backbone to improve adversarial loss functions for generating a high-resolution (HR) image from a low-resolution (LR) image. The proposed model is trained on pairs of LR and HR image datasets to map the LR image to its corresponding HR image. Then, extensive experiments are carried out to obtain HR images that are compared to some state-of-the-art models. The experimental results also investigated losses, indicating the effectiveness and improvement of the suggested approach for boosting satellite image resolution.