Advanced Precipitation Retrieval over Korea: Adaptive Error Reduction of GPM IMERG Products through Multi- Source Data AI Modeling

Abstract

In recent years, global warming has accelerated due to the increase in greenhouse gas concentrations in the atmosphere, causing a variety of extreme weather events around the world. These extreme weather events are causing increasing economic damage around the world, which has led to a growing interest in extreme weather. Extreme weather is defined as extreme weather events that are significantly above or below normal. Typical extreme weather events include heat waves, cold waves, heavy rainfall, droughts, typhoons, wildfires, Arctic Sea ice loss and sea level rise, and come in many forms. Climate change-induced changes in high pressure in the North Pacific are having a major impact on seasonal rainfall patterns in East Asia. In particular, the strengthening and expansion of high-pressure areas can lead to more intense weather events in East Asia, and Korea is also experiencing an increase in damage from heavy rainfall. In addition, precipitation measurement is becoming increasingly important as climate change leads to more erratic precipitation patterns. Satellite-based remote sensing data have the advantage of providing regular and consistent data over a wide area, compensating for spatial limitations. In particular, the introduction of microwave sensors has enabled more direct and accurate measurements of precipitation, making them an essential tool in modern meteorology. In particular, the Integrated Multi-satellite Retrievals for Global Precipitation Measurement (IMERG) provides high spatial and temporal resolution and is an important input for research in a variety of fields. The objective of this thesis is to improve the IMERG precipitation data by using multi- source data for the Korean summer season. To improve the IMERG precipitation data, we first collected Korean point data and analysed precipitation-related climate data to understand the changes and characteristics of Korean precipitation patterns. Second, the accuracy of microwave-based precipitation products was compared. IMERG, a representative precursor of multi-satellite and microwave remote sensing precipitation products, and the Korean ground-based weather radar Constant Altitude Plan Position Indicator (CAPPI) were evaluated against Automated Synoptic Observing System (ASOS). This was done to assess the advantages and disadvantages of each output compared to ground observations. In this study, two models are proposed to improve the precipitation performance of IMERG. The first model develops a machine learning based precipitation model that uses optical information from the Geo-KOMPSAT-2A/Advanced Meteorological Imager (GK- 2A/AMI) to compensate for the limitations of IMERG's physical algorithm. The second model uses additional meteorological variables to design a multilayer perceptron (MLP) and machine learning model. The meteorological variables used here are K-Index, 500 hPa U-Wind and 500 hPa Geopotential Height. The K-Index indicates the instability of the atmosphere and is used to predict heavy rainfall, while the U-Wind is a variable that captures the intensity of the typhoon. The 500 hPa Geopotential Height is a key indicator for identifying rainy season fronts in Korea, and this study compared the performance of the model using this data as input to the researcher and the case where the researcher made the distinction. After comparing CAPPI and IMERG, we found that IMERG performed well in light precipitation, while CAPPI performed better in heavier precipitation. Based on this, we developed a model that reflects the characteristics of CAPPI to improve the performance of IMERG. The precipitation improvement model using GK-2A/AMI optical information is a model that has been verified to reflect the characteristics of Korean precipitation. Compared with IMERG, the performance improvement was not significant, but compared with ASOS, it showed stability in temporal and spatial correlation. In addition, the model using optical satellite information and meteorological factors together significantly improved performance compared to the model using optical satellite information alone. In particular, the 500 hPa geopotential height model was designed by the researchers rather than learned directly by the model. These results can be used as a basis for improving the accuracy of weather forecasts and are expected to provide a basis for the design of future satellite-based precipitation output models in a number of ways.