Precipitation nowcasting using transformer-based generative models and transfer learning for improved disaster preparedness

Abstract

Due to the rapidly changing climate conditions, precipitation nowcasting poses a daunting challenge because it is impossible to make accurate short-term forecasts due to the rapid fluctuations in weather conditions. There are limitations to traditional methods of forecasting precipitation, such as the use of numerical models and radar extrapolation, when it comes to providing highly detailed and timely forecasts. With the help of contemporary machine learning (ML) models, including deep neural networks, transformers and generative models, complex precipitation nowcasting tasks can be performed in an efficient way. To address this critical task and enhance proactive emergency disaster management, we propose an innovative method based on transformer-based generative models for precipitation nowcasting. Our study area is the Soyang Dam basin in South Korea, located upstream of the Han River, characterized by a monsoon climate with approximately 1200 mm of annual precipitation. To develop a precipitation nowcasting model, radar composite data from 10 weather radars across South Korea is used. By utilizing radar reflective data in order to train our model, we are able to effectively predict future precipitation patterns, thus mitigating the risk of catastrophic weather conditions caused by heavy rainfalls. This dataset covers reflectivity data from 2018 to 2022, with a spatial resolution of 1km over a 960 × 1200 grid. Normalization using the min–max scaler method is applied to this reflectivity data, which is then transformed into grayscale images for uniform comparison. We enhance performance effectively by employing transfer learning with pre-trained Transformer models. Initially, we train the model using a comprehensive dataset. Subsequently, we fine-tune it for precipitation nowcasting using radar reflective data. This adaptation improves the accuracy of rainfall forecasting by capturing crucial features. Leveraging prior task knowledge through transfer learning not only enhances prediction accuracy but also increases overall efficiency. In terms of predictive accuracy, extensive experimental results demonstrate that our transformer-based nowcasting model outperforms related approaches, including conditional generative adversarial networks (cGANs), U-Net, convolutional long short-term memory (ConvLSTM), pySTEP. As a result of this research, disaster preparedness and response will be greatly improved through improved weather prediction.

Keywords

• Precipitation nowcasting
• Transformer-based generative model
• Radar reflective data
• ConvLSTM
• cGAN
• U-net