Deep Learning for Precipitation Nowcasting: A Benchmark and A New Model (1706.03458v2)

Abstract: With the goal of making high-resolution forecasts of regional rainfall, precipitation nowcasting has become an important and fundamental technology underlying various public services ranging from rainstorm warnings to flight safety. Recently, the Convolutional LSTM (ConvLSTM) model has been shown to outperform traditional optical flow based methods for precipitation nowcasting, suggesting that deep learning models have a huge potential for solving the problem. However, the convolutional recurrence structure in ConvLSTM-based models is location-invariant while natural motion and transformation (e.g., rotation) are location-variant in general. Furthermore, since deep-learning-based precipitation nowcasting is a newly emerging area, clear evaluation protocols have not yet been established. To address these problems, we propose both a new model and a benchmark for precipitation nowcasting. Specifically, we go beyond ConvLSTM and propose the Trajectory GRU (TrajGRU) model that can actively learn the location-variant structure for recurrent connections. Besides, we provide a benchmark that includes a real-world large-scale dataset from the Hong Kong Observatory, a new training loss, and a comprehensive evaluation protocol to facilitate future research and gauge the state of the art.

• The paper presents the TrajGRU model, which learns dynamic location-variant connections to improve precipitation nowcasting over ConvLSTM and other baselines.
• The paper establishes a comprehensive benchmark using the HKO-7 dataset and balanced loss functions to rigorously evaluate model performance across rain intensity thresholds.
• The paper demonstrates TrajGRU's superior performance in synthetic experiments and real-world radar forecasting, highlighting its potential for enhanced public safety and operational efficiency.

Deep Learning for Precipitation Nowcasting: A Benchmark and A New Model

Abstract

Precipitation nowcasting is essential for various public services, ranging from rainstorm warnings to flight safety. This paper introduces a benchmark and a new deep learning model, Trajectory GRU (TrajGRU), for precipitation nowcasting, highlighting the shortcomings of existing ConvLSTM models in capturing location-variant transformations. The proposed TrajGRU model actively learns state-to-state connection structures, providing improved performance over ConvGRU, Dynamic Filter Network (DFN), and Convolutional Neural Network (CNN) baselines. The benchmark includes a large-scale dataset from the Hong Kong Observatory and a comprehensive evaluation protocol, emphasizing the importance of balanced loss functions for training models with imbalanced rainfall data.

Introduction

Precipitation nowcasting, which predicts short-term rainfall intensity based on radar echo maps and other observational data, is vital for real-world applications such as road safety, flight safety, and citywide rain alerts. Traditional techniques rely on optical flow methods, which estimate cloud movements and predict future radar maps using semi-Lagrangian advection. These methods, however, do not utilize the vast amount of existing radar data.

Recent advancements have seen the application of deep learning models like Convolutional Long Short-Term Memory (ConvLSTM) in nowcasting, substantially outperforming traditional methods. Nevertheless, challenges remain, including the location-invariant nature of ConvLSTM's convolutional recurrent structure, which fails to model location-variant motion patterns like rotation. This paper addresses these issues by proposing the TrajGRU model that actively learns location-variant recurrent connections and a benchmark for systematic evaluation using large-scale datasets.

Model and Benchmark

1. Trajectory GRU Model (TrajGRU):
   • TrajGRU revises the recurrent structure by incorporating learned location-variant structures for state-to-state connections, thus capturing motion patterns more effectively than ConvGRU.
   • The model introduces flow fields to represent learned connection structures, allowing it to be more efficient and flexible than fixation structures in ConvGRU.
2. Benchmark:
   • The HKO-7 dataset, comprising radar echo data from 2009 to 2015, with real-time, large-scale, and fine-grained scenarios, provides a substantial foundation.
   • Two testing protocols (offline and online) and balanced loss metrics (B-MSE and B-MAE) ensure comprehensive evaluation, emphasizing higher rainfall impact areas. Balanced loss functions are particularly noted for their consistency in evaluating models across multiple rain-rate thresholds.

Experiments and Results

1. Synthetic MovingMNIST++ Experiment:
   • First, to validate the effectiveness of TrajGRU, the synthetic MovingMNIST++ dataset includes random rotations, scale changes, and illumination changes.
   • TrajGRU consistently outperforms ConvGRU, DFN, and CNNs, demonstrating its superiority in capturing spatiotemporal correlations with fewer parameters.
2. Evaluation on HKO-7 Benchmark:
   • Both offline and online settings validated various models, including simple baselines (last frame predictions), optical flow-based methods (ROVER), and deep learning models (TrajGRU, ConvGRU, CNNs).
   • TrajGRU achieves the best performance overall, with significant improvements in high rainfall thresholds, indicating its potential for real-world application. Online fine-tuning consistently enhances model performance, emphasizing the necessity of adapting models dynamically.

Implications and Future Directions

The research highlights several practical and theoretical implications: - Practical: Improved nowcasting performance can significantly enhance public safety and operational efficiency in weather-sensitive industries. - Theoretical: The TrajGRU model's ability to learn dynamic connection structures opens avenues for improving other spatiotemporal tasks. The benchmark sets a new standard for evaluating precipitation nowcasting models, advocating for balanced loss measures and online learning strategies.

Future developments may include applying TrajGRU in visual object tracking and video segmentation, further refining dynamic connection learning mechanisms, and integrating the model into operational nowcasting systems to validate its practical utility.

Conclusion

This paper makes a significant contribution by proposing a novel deep learning model, TrajGRU, and establishing a comprehensive benchmark for precipitation nowcasting. TrajGRU's ability to model location-variant correlations and the benchmark's emphasis on balanced evaluation protocols set a new direction for future research, paving the way for more accurate and reliable rainfall predictions.