An AI-Based Query System for Historical Weather Radar Reflectivity Retrieval Using a Convolutional Auto-encoder
Abstract The weather patterns in Taiwan have been more changeable in recent years. Taking rainfall as an example, natural disasters that lead to major damage are often caused by heavy rainfall in a short period of time. Therefore, how to effectively predict changes in weather patterns has become an important issue for the development of disaster prevention technology. In this paper, an AI-based query system using a convolutional auto-encoder is developed for historical weather radar reflectivity retrieval that is expected to forecast heavy rainfall soon. When the weather radar receives a current radar reflectivity, meteorologists can employ the query system to quickly search for the historical radar reflectivity with the weather characteristics similar to the one to be queried. The experimental results reveal that under the given MSE threshold of 58,000, the average pass rate achieves 83.9% for the ranked top 10 similar results from 33,461 historical records via 4,319 query tests in total. Without considering the model loading time of 60 seconds, the execution time of each query requires 6 seconds for only listing query results or 8 seconds for including saving resulting maps in pseudo-colour.