Wednesday, 31 January 2024
Hall E (The Baltimore Convention Center)
One of the key elements about weather forecasting is to search similar events from the past data and to compare them with the current weather condition for similar case-based reasoning. However, it is rarely used for forecasters despite its usefulness because of the sheer volume of data.
This study proposes a tool of Artificial Intelligence (AI) / machine learning (ML) based “quick” retrieval of similar weather events using imagery from geostationary satellites. This imagery is taken from the fixed area at different times to observe diverse weather events, and widely used in meteorological research. In this study, we use two different geostationary satellites, which are COMS and GK2A, because they had different operating periods. They also have three different channels which are Infrared, Short wavelength infrared and Water vapor. Each channel can complement each other because objects to be observed by each channel are different.
Our tool for searching similar weather events consists of two modules: an image vectorization module, and a search module. An image vectorization module including Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) is used to extract and vectorize the spatiotemporal features of satellite imagery. Search module uses the NN-Descent algorithm to produce a searchable K-Nearest Neighbors (KNN) graph format. To speed up similar weather events retrieval, the KNN graph is ultimately transformed into an optimal form by removing unnecessary edges.
To help forecasters work, we have implemented it as User Interface (UI). The UI provides diverse weather services like displaying satellite imagery and weather maps provided by the Korea Meteorological Administration. In addition, we are continuously working to add additional features and functions to assist forecasters in their work.
This study proposes a tool of Artificial Intelligence (AI) / machine learning (ML) based “quick” retrieval of similar weather events using imagery from geostationary satellites. This imagery is taken from the fixed area at different times to observe diverse weather events, and widely used in meteorological research. In this study, we use two different geostationary satellites, which are COMS and GK2A, because they had different operating periods. They also have three different channels which are Infrared, Short wavelength infrared and Water vapor. Each channel can complement each other because objects to be observed by each channel are different.
Our tool for searching similar weather events consists of two modules: an image vectorization module, and a search module. An image vectorization module including Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) is used to extract and vectorize the spatiotemporal features of satellite imagery. Search module uses the NN-Descent algorithm to produce a searchable K-Nearest Neighbors (KNN) graph format. To speed up similar weather events retrieval, the KNN graph is ultimately transformed into an optimal form by removing unnecessary edges.
To help forecasters work, we have implemented it as User Interface (UI). The UI provides diverse weather services like displaying satellite imagery and weather maps provided by the Korea Meteorological Administration. In addition, we are continuously working to add additional features and functions to assist forecasters in their work.

