ClusterCast: Advancing Precipitation Nowcasting with Self-Clustering GANs

By Soham NandiReviewed by Susha Cheriyedath, M.Sc.May 9 2024

In an article published in the journal Nature, researchers introduced ClusterCast, a novel generative adversarial network (GAN) framework for precipitation nowcasting. They addressed the challenge of capturing diverse precipitation patterns by employing a self-clustering approach within a generator network.

By learning to label precipitation types and predict future radar frames based on these labels, ClusterCast achieved more accurate representation learning. Experimental results demonstrated its effectiveness in generating non-blurry future frames and outperforming benchmarks in precipitation nowcasting tasks.

Background

The field of precipitation nowcasting, crucial for weather prediction and disaster management, has seen significant advancements with the advent of deep learning (DL) techniques. However, existing DL models for precipitation nowcasting encounter challenges such as data blurring with increasing forecast lead time, limiting their predictive capabilities. While some models, like adversarial extrapolation neural network (AENN) and deep generative models of rainfall (DGMR), have attempted to address this issue using GAN-based approaches, they still face limitations in capturing diverse precipitation patterns accurately.

Most notably, these models often rely on learning a single latent representation of precipitation, which may not effectively capture the nuanced characteristics of different precipitation types. Additionally, there is a risk of mode collapse in GAN-based models, where the generator network fails to capture various precipitation scenarios accurately.

To address these challenges, this paper introduced ClusterCast, a novel GAN framework for precipitation nowcasting. ClusterCast leveraged a self-clustering approach within a generator network to capture the high-dimensional distribution of disparate precipitation types effectively. By utilizing hierarchical resolutions and self-supervised labels, ClusterCast aimed to learn more robust and adaptable representations of precipitation, thus overcoming the limitations of previous models.

Additionally, ClusterCast incorporated ensemble forecasting techniques to further improve predictive performance and mitigate uncertainty in radar measurements. Overall, this paper aimed to fill the gaps in existing precipitation nowcasting models by proposing a more effective and versatile approach for learning heterogeneous representations of precipitation and enhancing predictive accuracy in various precipitation scenarios.

Self-supervised Precipitation Nowcasting Framework

The self-clustered generator architecture proposed in this paper for precipitation nowcasting leveraged hierarchical convolutional gated recurrent unit (ConvGRU) cells and discriminators to capture space-time patterns effectively. The generator generated future radar frames based on a Gaussian distribution clustered among sampled points, while the discriminators ensured accurate discrimination between real and generated sequences.

To address the challenge of collapsing in GANs, the framework incorporated a self-clustering sub-network that classified precipitation types based on hierarchical resolution features. By initializing Gaussian distributions for each precipitation type, the model alleviated distribution variations, enhancing predictive accuracy. The hierarchical ConvGRU cells decoded input latent states and resolution features to generate future output frames. The framework's unified approach simultaneously optimized parameters for both regression and clustering tasks, improving stability and performance.

Additionally, the self-supervised learning process employed traditional clustering techniques like k-means, principle component analysis (PCA), and linear-based methods to enhance generator performance and provide robust representations of precipitation distribution. The discriminators, based on the DGMR discriminator structure, distinguished between real and fake spatio-temporal features in radar sequences using hinge loss.

Ensemble prediction systems further improved forecasting accuracy by considering observation uncertainties. This ensemble system, comprising 64 members, including control, applied random perturbations to address uncertainties from observation errors. The ensemble mean forecasts served as the final prediction output. Through these approaches, the framework achieved more accurate and stable precipitation nowcasting, overcoming challenges such as mode collapse and uncertainty in radar measurements.

Experimental Evaluation of Precipitation Nowcasting Models

The experiments were conducted to evaluate the effectiveness of the proposed approach for two-hour precipitation prediction, comparing it with four baseline models: Rainymotion, convolutional long short-term memory (ConvLSTM), trajectory gated recurrent unit (TrajGRU), and DGMR. The dataset comprised radar reflectivity data from South Korea, obtained from the Korea Meteorological Administration, covering a 1024 square kilometers area with a temporal resolution of five minutes.

Training data ranged from 2012 to 2019, while verification and test datasets spanned 2020 and 2021, respectively. Radar data underwent downscaling to a resolution of four kilometers and was sampled systematically to ensure representation across various precipitation intensities. The proposed generative model utilized the Adam algorithm for optimization, with experiments conducted using a batch size of 16.

Rainfall intensity estimation was performed using the radar reflectivity factor (Z)- rainfall rate (R) relationship, with constants calibrated for the Korean climate. Evaluation metrics included mean squared error (MSE), peak-signal-to-noise ratio (PSNR), characteristic stability index (CSI), fractions skill score (FSS), equitable threat score (ETS), and Heidke skill score (HSS), providing a comprehensive analysis of model performance.

Additionally, the lead time performance of the benchmark models was compared against ClusterCast to assess predictive accuracy over time. Through these experiments, the effectiveness of the proposed approach for precipitation nowcasting was thoroughly evaluated, highlighting its potential for real-world applications in weather forecasting.

Results and Insights in Precipitation Prediction

The researchers focused on assessing the performance of ClusterCast. Results from comprehensive analyses revealed that ClusterCast demonstrated flexibility across various precipitation types and exhibited robustness over time by learning distinct distributions for each precipitation type. Visualization techniques, mapping high-dimensional features into two and three dimensions, provided insights into the spatial and temporal characteristics of precipitation, aiding in self-supervised labeling.

The main results showcased ClusterCast outperforming comparison models in predicting two-hour precipitation, particularly excelling in MSE after 60 minutes. Despite initial setbacks attributed to sharpness pursuit and distribution shifts, GAN-based models adapted effectively over time, maintaining robust performance.

Ablation studies validated the effectiveness of clustering components, with k-means clustering combined with an encoder achieving superior performance, particularly for heavy rain events. Visualizations of clustered results highlighted logical patterns in precipitation types, supporting the efficacy of multi-latent space learning. Ensemble approaches further enhanced prediction reliability, mitigating uncertainty associated with sensitive precipitation events.

Conclusion

In conclusion, ClusterCast introduced a breakthrough in precipitation nowcasting, surpassing existing models by effectively capturing diverse precipitation patterns. Through self-clustering and ensemble techniques, it achieved robust predictions, addressing challenges like mode collapse and data blurring. Experimental results validated its superiority in two-hour precipitation forecasting, offering insights into spatial and temporal characteristics.