Precipitation is a weather phenomenon that significantly affects human lives. Extreme or heavy precipitation can lead to natural disasters. Therefore, a precise, accurate and timely estimate of approaching extreme precipitation can help avoid severe economic losses and assure the safety of human lives and property.
How Precipitation is Observed
Weather radars like Z_R relations, use unique algorithms to observe precipitation. The data collected from the radar echoes series can be utilized for precipitation forecasting for the next two hours and can give information on changes and development of precipitation, which can help in decision-making about possible effects of precipitation.
However, the uncertainty in spatiotemporal properties of the precipitation development process results in complications in forecasting its change and movement trends. Accurate prediction of upcoming changes in radar echoes is fundamental to improving the accuracy of precipitation prediction.
Convective Weather Nowcasting
The forecast of extinction, evolution, development, and occurrence of thunderstorms and catastrophic convective weather for the upcoming few hours is known as convective weather nowcasting. It is critical for meteorological disaster mitigation and prevention.
A weather radar is a primary tool for up to two hours of convective weather nowcasting. The current operational nowcasting techniques include identifying and tracking thunderstorms and radar data-based auto-extrapolation forecasting technologies such as optical flow, cross-correlation, and single centroid methods.
The applicability of conventional extrapolation techniques based on radar echoes is restricted to a single unit with robust radar echoes and a narrow range, and they only employ shallow-level feature information from the radar image. As a result, these techniques are not accurate in forecasting widespread precipitation.
Tracking Radar Echoes by Cross-Correlation (TREC)
The updated algorithm of Tracking Radar Echoes by Cross-correlation (TREC) treats the echo variation as linear. However, the formation and extinction of a convective process in the real atmosphere cause the strength and shape of radar echoes to change in a highly complicated manner.
These conventional approaches do not use past radar readings very often. As a result, the forecast accuracy no longer suffices for high-precision prediction, and the forecast leading time is often less than an hour.
Deep Learning Algorithms
Deep learning has recently been used to evaluate, associate, remember, learn, and infer ambiguous situations. As a result, deep learning applications have made great progress in environment changes, nowcasting, image identification, and other disciplines.
However, all current deep learning algorithms for extrapolation always suffer from blur; that is, as the forecast leading time rises, the severity of the forecast echoes' dispersion worsens, leading to blurring. Therefore, a critical problem that has to be resolved in the current forecast operational applications is how to decrease the expected echo's blur while increasing forecast accuracy.
The Study Area Chosen For This Research
The study area that the researchers have focused on in this research is Guangdong Province in southern China: a subtropical monsoon climate region. Intense airflow and frequent interactions between this region's hot and cold flows provide strong convective weather and copious amounts of precipitation.
This region has plenty of meteorological disasters, including cold waves, cold dew wind, drought, typhoons, rainstorms and floods, strong convection (strong wind, strong thunderstorm, tornado, hail), and low temperature and rain. Among these disasters, rainstorms and tropical cyclones are the most intense and frequent. These disasters cause severe losses to the country's economy. Improving nowcasting technology is imperative for this region.
How the Research was Conducted
In this study, researchers have proposed a radar echo forecast technique based on the Temporal and Spatial Generative Adversarial Network (TSGAN), which extracts input radar images' spatiotemporal features through a 3D convolution and attention mechanism module and utilize a dual-scale generator and a multi-scale discriminator to restore the predicted echoes' complete information.
The key benefit of the suggested technique is that it efficiently reduces the blur of the expected echoes while improving the predictions of the echo details and assuring the accuracy of the forecast findings.
The prediction performance of the TSGAN algorithm on intense convective rainfall was tested using over 80,000 instances in 2021. However, due to page limitation, only two processes were selected for echo extrapolation forecasts visualization in this study, i.e. the typhoon process on 8 October 2021 and the squall line process on 4 May 2021. Furthermore, the results from the TSGAN method were compared with several other methods, including PredRNN V2, PredRNN, ConvGRU, and the optical flow.
Significant Findings of the Study and Future Outlooks
The two scenarios tested show that the TSGAN technique can more accurately anticipate the location and shape of radar echoes while preserving rich spatial features. However, despite the clear benefits of the TSGAN technique in forecasting spatial details, the CSI's (critical success index) comprehensiveness and grid-to-grid calculation method prevent a direct correlation between an increase in spatial details and an increase in the CSI score.
To further refine the algorithm and enhance its practical applicability, future research shall use more diverse categories of weather processes as test cases and incorporate the concept of the spatial neighborhood.
Reference
Xunlai Chen, Mingjie Wang, Shuxin Wang, Yuanzhao Chen, Rui Wang, Chunyang Zhao and Xiao Hu (2022) Weather Radar Nowcasting for Extreme Precipitation Prediction Based on the Temporal and Spatial Generative Adversarial Network. Atmosphere. https://www.mdpi.com/2073-4433/13/8/1291/htm
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.