AI-Based Retrieval and Prediction of Weather Imagery Using Geostationary Satellite Data

One of the critical aspects of weather forecasting is identifying similar past weather events from historical data and comparing them with current conditions. However, the sheer volume of accumulated data has made it challenging for forecasters to effectively utilize this approach.

This study proposes a tool that leverages imagery from geostationary satellites to rapidly retrieve similar weather events and employs artificial intelligence (AI) and machine learning (ML) techniques to provide short-term forecasts for the next six hours. The tool utilizes data from two geostationary satellites, COMS and GK2A, which have observed a range of meteorological phenomena over their operational periods.

The proposed tool consists of an image vectorization module and a search and prediction module. The image vectorization module employs Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN) to extract and vectorize the spatiotemporal features of satellite imagery. The search and prediction module utilizes the NN-Descent algorithm to optimize the K-Nearest Neighbors (K-NN) graph, enabling rapid retrieval of similar events and generating short-term forecasts. Additionally, the ultra-short-term prediction model incorporates adversarial learning, wherein a generator and discriminator model are employed in tandem to enhance prediction accuracy. This model analyzes the past 6 to 12 hours of satellite imagery to produce precise forecasts for the next six hours.

The tool is implemented as a user interface (UI), providing access to various meteorological data, including satellite imagery and weather maps. This interface is designed to support forecasters in performing their tasks more efficiently, with ongoing development to introduce additional features that further augment its utility.