RN-Net: A Deep Learning Approach to 0-2 Hour Rainfall Nowcasting Based on Radar and Automatic Weather Station Data

Fuhan Zhang, Xiaodong Wang, Jiping Guan, Meihan Wu, Lina GuoSensors 2021 · OpenReview

TL;DR

This paper introduces RN-Net, a novel deep learning model that fuses radar and automatic weather station data to achieve high-resolution, short-term rainfall nowcasting (0-2 hours). RN-Net significantly outperforms traditional numerical weather prediction models, offering a more accurate and timely forecast for critical applications like disaster prevention and daily planning.

What Problem Does This Paper Address?

Accurate and timely rainfall nowcasting is crucial for disaster preparedness, agriculture, and daily life. Traditional Numerical Weather Prediction (NWP) models struggle with "spin-up" issues, leading to inaccurate short-term forecasts, especially for rainfall. Existing machine learning and deep learning approaches often lack the necessary spatiotemporal resolution for effective large-area nowcasting. The challenge lies in integrating diverse data sources (radar and ground-based stations) to capture the complex dynamics of precipitation at a fine scale and within a short lead time.

Methods

The core of this work is the Rainfall Nowcasting Network (RN-Net), a dual-input, dual-encoder recurrent neural network. RN-Net addresses the temporal resolution mismatch between radar (6-minute intervals) and automatic weather station (AWS) data (10-minute intervals, interpolated to 30-minute for this study) by using separate encoders for each data type.

  • Dual Encoders: Two distinct encoders process the radar echo data and the interpolated AWS rainfall data. These encoders extract spatiotemporal features from their respective inputs.
  • Fusion Module: A Convolutional Neural Network (CNN)-based fusion module combines the features extracted by the two encoders. This allows the model to leverage the complementary information from both data sources.
  • Rainfall Predictor: A predictor module, also structured with recurrent layers and upsampling units, takes the fused features and generates the future rainfall grid data.
  • TrajGRU: The paper utilizes TrajGRU as the recurrent unit within the encoders and predictor. TrajGRU is an advancement over ConvGRU, featuring a dynamic connection structure that generates optical flow to better capture complex motion patterns in spatiotemporal data.

The model is trained end-to-end using PyTorch with an Adam optimizer and L2 regularization.

Datasets

The study uses a dataset from the southeastern coastal area of China.

  • Radar Echo Data: Doppler radar mosaic data covering a 240x240 grid with a 5 km spatial resolution and 6-minute temporal resolution. Data from May to September in 2018 and 2019 were used.
  • Rainfall Data: Interpolated rainfall data from 13,655 automatic weather stations. The spatial resolution matches the radar data (5 km), and the temporal resolution was converted to 30 minutes after interpolation.

The dataset was split into training (207 days), validation (29 days), and testing (57 days) sets.

Key Insights and Ideas

  • Data Fusion is Key: Combining radar and ground-based rainfall data is crucial for accurate nowcasting. Radar provides broad spatial coverage, while AWS data offers direct, ground-truth measurements.
  • Addressing Temporal Mismatch: The dual-encoder architecture effectively handles the different temporal resolutions of radar and AWS data, preventing information loss or distortion.
  • TrajGRU's Advantage: The use of TrajGRU, with its dynamic connection structure, appears to be more effective than standard ConvLSTM for capturing complex spatiotemporal rainfall patterns.
  • High Spatiotemporal Resolution is Achievable: The model demonstrates that it's possible to achieve high-resolution (5 km, 30 min) rainfall nowcasting, which is a significant improvement over previous methods.

Key Results

  • Superior Performance: RN-Net achieved the highest Threat Scores (TS) and Critical Success Indices (CSI) across various thresholds (0.25 mm, 1 mm, 2.5 mm) and lead times (0.5 h, 1 h, 2 h) compared to other deep learning methods and the WRF model.
  • Significant Improvement over WRF: For a 0.25 mm threshold, RN-Net's TS at 0.5h, 1h, and 2h were nearly 4x, 3x, and 3x higher than WRF, respectively. CSI also showed similar substantial improvements.
  • Accuracy at Different Lead Times: RN-Net demonstrated strong performance even at 2-hour lead times, with a TS of 0.435 and CSI of 0.371 (at 0.25 mm threshold).
  • Qualitative Improvements: Visualizations showed RN-Net producing more accurate and less prone to "false alarms" than the WRF model, which often exhibited large-scale spurious rainfall predictions.

Limitations and Future Work

  • Generalization: The authors acknowledge that RN-Net's generalization ability might be limited. This is attributed to the dataset only containing summer and autumn data from a specific region in Southeast China, and rainfall being influenced by various factors like topography and climate.
  • Accuracy Improvement: While significantly better than WRF, the paper suggests that RN-Net's rainfall intensity nowcasting accuracy still has room for improvement, especially when compared to specialized precipitation intensity nowcasting models.
  • Future Directions:
    • Expand Dataset: Increase the scale and geographical coverage of the dataset.
  • More Input Data: Incorporate additional meteorological data sources to provide richer spatiotemporal features.
  • Novel Networks: Explore other advanced deep learning architectures.

Summary updated

Abstract

Precipitation has an important impact on people’s daily life and disaster prevention and mitigation. However, it is difficult to provide more accurate results for rainfall nowcasting due to spin-up problems in numerical weather prediction models. Furthermore, existing rainfall nowcasting methods based on machine learning and deep learning cannot provide large-area rainfall nowcasting with high spatiotemporal resolution. This paper proposes a dual-input dual-encoder recurrent neural network, namely Rainfall Nowcasting Network (RN-Net), to solve this problem. It takes the past grid rainfall data interpolated by automatic weather stations and doppler radar mosaic data as input data, and then forecasts the grid rainfall data for the next 2 h. We conduct experiments on the Southeastern China dataset. With a threshold of 0.25 mm, the RN-Net’s rainfall nowcasting threat scores have reached 0.523, 0.503, and 0.435 within 0.5 h, 1 h, and 2 h. Compared with the Weather Research and Forecasting model rainfall nowcasting, the threat scores have been increased by nearly four times, three times, and three times, respectively.
BibTeX
@inproceedings{zhang2021rn,
  author = {Fuhan Zhang and Xiaodong Wang and Jiping Guan and Meihan Wu and Lina Guo},
  title = {{RN-Net: A Deep Learning Approach to 0-2 Hour Rainfall Nowcasting Based on Radar and Automatic Weather Station Data}},
  booktitle = {Sensors},
  year = {2021},
  url = {https://openreview.net/forum?id=IFFWkegYpDp},
}
Figure