Poster Title:Super-Resolution of Out-of-Domain Satellite Sensory Imagery for Effective Cloud Cover Analysis
Full Name:Hanan Gani
Affiliation / Institution:Mohamed Bin Zayed University of Artificial Intelligence
Co-Author
Full Name:Salman Khan
Affiliation / Institution:Mohamed Bin Zayed University of Artificial Intelligence
Co-Author 2
Full Name:Guy Pulik
Affiliation / Institution:
Co-Author 3
Full Name:Daniel Rosenfeld
Affiliation / Institution:
Abstract:Cloud seeding, a technique used to artificially induce rainfall, plays a crucial role in reducing regional temperatures and contributes to efforts against global warming. The UAE has been using this method to address its scarcity of rainfall and lower overall temperatures. While cloud seeding is effective, it comes with high operational costs due to the difficulty of identifying the ideal clouds for seeding. Heavy cloud cover and thin, dusty clouds are generally unsuitable for seeding, making it essential to locate optimal cloud formations. Currently, satellite imagery is employed for this purpose, with specific wavelengths reflected by cloud cover analyzed to identify potential seeding locations. However, the resolution of the SEVIRI satellite images used for this analysis is low, making it challenging to accurately pinpoint the best clouds for seeding. In this paper, we present the first-ever initiative to apply artificial intelligence (AI) to the problem of identifying cloud formations suitable for seeding. While AI has proven effective in many natural vision tasks, it struggles with satellite imagery due to differences in the data domains. Even with training, existing AI models do not generalize well to this context. To overcome these challenges, we propose a novel approach that generates high-resolution (HR) images from low-resolution (LR) satellite data without sacrificing diversity or perceptual quality. Our method is based on diffusion models, inspired by non-equilibrium thermodynamics, which gradually add random noise to the data and then learn to reverse the process, reconstructing desired data samples from the noise. Specifically, our diffusion-based super-resolution (SR) model learns a bi-directional deterministic mapping between input noise and the generated HR image, guided by a well-trained teacher diffusion model. We trained this model on a dataset of 8,000 pairs of HR images from the VIIRS satellite and corresponding LR images from SEVIRI. Remarkably, our approach performs well even in zero-shot settings, where the model is applied without prior training on the specific dataset. Compared to existing deep learning Super-resolution approaches, our method shows superior performance and generalization. This makes it highly adaptable for real-world use, even in unfamiliar conditions. The proposed method has the potential to significantly improve the accuracy of identifying optimal cloud seeding locations in the UAE, leading to more efficient use of energy and resources while enhancing the effectiveness of cloud seeding operations.