GANs for Satellite-Based Next-Frame Prediction of Dust

Sand and dust storms (SDS) pose significant environmental and socioeconomic challenges by disrupting human well-being, ecosystems, transportation routes, agricultural operations, and infrastructure stability. Prediction of SDS events is essential to minimize their harmful effects. This research enhances SDS predictions through access to data from the European Organization for the Exploitation of Meteorological Satellites (EUMETSAT), including Meteosat Second Generation (MSG) and its Spinning Enhanced Visible and Infra-Red Imager (SEVIRI) RGB dust detection dataset. The data set comprises 96 daily images at 15-minute intervals, providing high-resolution temporal data essential for training and validating a machine learning-based prediction model. A custom-built next-frame prediction model, inspired by the Pix2Pix architecture, was developed and optimized to operate efficiently with lower computational power. The modified technique reduces GPU memory demands alongside faster training durations (approximately 16.17 minutes in total, with each epoch taking 192 seconds) without compromising image quality during generation. Extensive pre-processing steps in the model protect spatial information, leading to better SDS prediction accuracy levels. The optimal model generated images with 32.59 dB PSNR, 0.8722 SSIM, and 0.9985 FSIM, showing a close visual match for the original frame structures. Satellite-based SDS prediction technologies show their value for early warnings through reliable tools to support disaster management in regions susceptible to SDS occurrences. The novel integration of superior satellite imagery with advanced machine learning algorithms results in more precise forecasts and better environmental risk management planning. Using weather data in real-time operational scenarios presents computational efficiency issues and limitations when generalizing to different atmospheric conditions. Future research efforts will focus on model scalability improvements along with real-time application performance optimization and developing methodology extension capabilities beyond atmospheric prediction.