Search

Quantitative precipitation nowcasting (QPN) can help to reduce the enormous socioeconomic damage caused by extreme weather. The QPN has been a challenging topic due to rapid atmospheric variability. Recent QPN studies have proposed data-driven models using deep learning (DL) and ground weather radar. Previous studies have primarily focused on developing DL models, but other factors for DL-QPN have not been thoroughly investigated. This study examined four critical factors in DL-QPN, focusing on their impact on forecasting performance. These factors are the deep learning model (U-Net, as well as a convolutional long short-term memory, or ConvLSTM), input past sequence length (1, 2, or 3 h), loss function (mean squared error, MSE, or balanced MSE, BMSE), and ensemble aggregation. A total of 24 schemes were designed to measure the effects of each factor using weather radar data from South Korea with a maximum lead time of 2 h. A long-term evaluation was conducted for the summers of 2020–2022 from an operational perspective, and a heavy rainfall event was analyzed to examine an extreme case. In both evaluations, U-Net outperformed ConvLSTM in overall accuracy metrics. For the critical success index (CSI), MSE loss yielded better results for both models in the weak intensity range (≤ 5 mmh-1), whereas BMSE loss was more effective for heavier precipitation. There was a small trend where a longer input time (3 h) gave better results in terms of MSE and BMSE, but this effect was less significant than other factors. The ensemble by averaging results of using MSE and BMSE losses provided balanced performance across all aspects, suggesting a potential strategy to improve skill scores when implemented with optimal weights for each member. All DL-QPN schemes exhibited problems with underestimation and overestimation when trained by MSE and BMSE losses, respectively. All DL models produced blurry results as the lead time increased, while the non-DL model retained detail in prediction. With a comprehensive comparison of these crucial factors, this study offers a modeling strategy for future DL-QPN work using weather radar data. [ABSTRACT FROM AUTHOR]

Quantitative precipitation nowcasting (QPN) can help to reduce the enormous socioeconomic damage caused by extreme weather. The QPN has been a challenging topic due to rapid atmospheric variability. Recent QPN studies have proposed datadriven models using deep learning (DL) and ground weather radar. Previous studies have primarily focused on developing DL models, but other factors for DL-QPN have not been thoroughly investigated. This study examines four critical factors in the DL-QPN, focusing on their effect on forecasting performance. The four key factors were the prediction design (single, recursive, and multiple predictions), deep learning model (U-Net and convolutional long short-term memory; ConvLSTM), input past sequence length (60 and 120 min), and output future sequence length (60 and 120 min). Twelve schemes were designed to measure the effects of each factor using weather radar data from South Korea. A long-term evaluation for 2018-2020 was conducted from an operational perspective, and a summer heavy rainfall event was analyzed to examine the extreme case. In both cases, U-Net demonstrated the best performance in the critical score index (CSI) with a multiple prediction design. U-Net performed better with shorter input sequences (i.e., 60 min) than with longer input sequences (i.e., 120 min), whereas ConvLSTM did not demonstrate a clear CSI difference over the input sequence length. Future sequence length did not have a significant effect on model performance. All the DL-QPN schemes exhibited underestimation and blurry results as the lead time increased. Through sensitivity analysis, U-Net was highly dependent on the most recent input time (i.e., 0 past minutes), whereas ConvLSTM was more sensitive to various time steps. With an explicit comparison of crucial factors, this study can provide a modeling strategy and contingency plan for future DL-QPN using weather radar data. [ABSTRACT FROM AUTHOR]

Numerical weather prediction models and probabilistic extrapolation methods using radar images have been widely used for precipitation nowcasting. Recently, machine-learning-based precipitation nowcasting models have also been actively developed for relatively short-term precipitation predictions. This study was aimed at developing a radar-based precipitation nowcasting model using an advanced machine-learning technique, conditional generative adversarial network (cGAN), which shows high performance in image generation tasks. The cGAN-based precipitation nowcasting model, named Rad-cGAN, developed in this study was trained with the radar reflectivity data of the Soyang-gang Dam basin in South Korea with a spatial domain of 128 × 128 pixels, spatial resolution of 1 km, and temporal resolution of 10 min. The model performance was evaluated using previously developed machine-learning-based precipitation nowcasting models, namely convolutional long short-term memory (ConvLSTM) and U-Net. In addition, Eulerian persistence model and pySTEPS, a radar-based deterministic nowcasting system, are used as baseline models. We demonstrated that Rad-cGAN outperformed reference models at 10 min lead time prediction for the Soyang-gang Dam basin based on verification metrics: Pearson correlation coefficient (R), root mean square error (RMSE), Nash–Sutcliffe efficiency (NSE), critical success index (CSI), and fraction skill scores (FSS) at an intensity threshold of 0.1, 1.0, and 5.0 mm h -1. However, unlike low rainfall intensity, the CSI at high rainfall intensity in Rad-cGAN deteriorated rapidly beyond the lead time of 10 min; however, ConvLSTM and baseline models maintained better performances. This observation was consistent with the FSS calculated at high rainfall intensity. These results were qualitatively evaluated using typhoon Soulik as an example, and through this, ConvLSTM maintained relatively higher precipitation than the other models. However, for the prediction of precipitation area, Rad-cGAN showed the best results, and the advantage of the cGAN method to reduce the blurring effect was confirmed through radially averaged power spectral density (PSD). We also demonstrated the successful implementation of the transfer learning technique to efficiently train the model with the data from other dam basins in South Korea, such as the Andong Dam and Chungju Dam basins. We used the pre-trained model, which was completely trained in the Soyang-gang Dam basin. Furthermore, we analyzed the amount of data to effectively develop the model for the new domain through the transfer learning strategies applying the pre-trained model using data for additional dam basins. This study confirmed that Rad-cGAN can be successfully applied to precipitation nowcasting with longer lead times and using the transfer learning approach showed good performance in dam basins other than the originally trained basin. [ABSTRACT FROM AUTHOR]

Numerical weather prediction models and probabilistic extrapolation methods using radar images have been widely used for precipitation nowcasting. Recently, machine-learning-based precipitation nowcasting models have also been actively developed for relatively short-term precipitation predictions. This study aimed to develop a radar-based precipitation nowcasting model using an advanced machine learning technique, conditional generative adversarial network (cGAN), which shows high performance in image generation tasks. The cGAN-based precipitation nowcasting model, named Rad-cGAN, developed in this study was trained with a radar reflectivity map of the Soyang-gang Dam region in South Korea with a spatial domain of 128 × 128 km, spatial resolution of 1 km, and temporal resolution of 10 min. The model performance was evaluated using previously developed machine-learning-based precipitation nowcasting models, namely convolutional long short-term memory (ConvLSTM) and U-Net, as well as the baseline Eulerian persistence model. We demonstrated that Rad-cGAN outperformed other models not only for the chosen site but also for the entire domain across the Soyang-gang Dam region. Additionally, the proposed model maintained good performance even with lead times up to 80 min based on the critical success index at the intensity threshold of 0.1 mm h-1, while RainNet and ConvLSTM achieved lead times of 70 and 40 min, respectively. We also demonstrated the successful implementation of the transfer learning technique to efficiently train model with the data from other dam regions in South Korea, such as the Andong and Chungju Dam regions. We used pre-trained model, which was completely trained in the Soyang-gang Dam region. This study confirms that Rad-cGAN can be successfully applied to precipitation nowcasting with longer lead times, and using the transfer learning approach it shows good performance in regions other than the originally trained region. [ABSTRACT FROM AUTHOR]