Precipitation nowcasting with radar data for evaluating multiple horizons using U-Net-based algorithm in Eastern...

Area of Science:

• Meteorology
• Artificial Intelligence
• Earth Science

Background:

• Severe meteorological events, particularly intense rainfall, are increasing globally, posing significant risks to safety, well-being, and economic sectors like agriculture and mining.
• Effective emergency alerts are critical for mitigating losses and preventing fatalities during extreme weather events.
• High temporal resolution precipitation forecasting tools, such as meteorological radars and satellites, are essential for timely warnings.

Purpose of the Study:

• To evaluate a multi-horizon precipitation nowcasting approach using a U-Net machine learning architecture.
• To investigate the optimal amount of past meteorological radar data (input horizons) required for accurate short-term precipitation forecasts.
• To determine the impact of increasing input data duration on nowcasting performance.

Main Methods:

• Utilized a U-Net deep learning architecture for spatial-temporal precipitation nowcasting.
• Employed meteorological radar data from the Eastern Amazon region.
• Evaluated the performance of nowcasting models with varying durations of past precipitation data as input (e.g., 60 minutes vs. 120 minutes).
• Assessed forecast accuracy using metrics such as Root Mean Square Error (RMSE) and Critical Success Index (CSI).

Main Results:

• Increasing input data beyond 60 minutes negatively impacts short-term precipitation nowcasting performance.
• Using 120 minutes of input data instead of 60 minutes led to a significant performance decrease: 17.60% increase in RMSE and 7.18% decrease in CSI for short-term forecasts.
• The study identified 60 minutes of past radar data as the optimal input duration for accurate short-term precipitation nowcasting.

Conclusions:

• The optimal duration of input data is crucial for accurate spatial-temporal precipitation nowcasting.
• Excessive input data can degrade the performance of U-Net based nowcasting models for short-term forecasts.
• These findings provide essential guidance for developing more effective precipitation nowcasting systems, enhancing decision-making during severe weather events.