Introduction
Recent advancements in AI-based weather forecasting, exemplified by models like Pangu-Weather and GraphCast, have revolutionized global meteorological predictions. However, these models predominantly rely on ERA5 reanalysis data, which features a temporal resolution of 6 hours and a spatial resolution of 0.25°. This granularity is insufficient for "nowcasting"—the critical 0–6 hour window required to predict fast-evolving, small-scale convective events like flash floods or severe thunderstorms.
To address this gap, we introduce DaYu, a novel transformer-based model designed to forecast high-resolution cloud images directly from raw geostationary satellite data. By bypassing reanalysis datasets, DaYu achieves a 0.05° spatial resolution, providing the high-frequency observations necessary for effective disaster prevention and maritime safety.
The DaYu Architecture
DaYu (named after the legendary Chinese figure who tamed the Great Flood) employs a Transformer-based autoregressive cascaded architecture. The design is optimized to handle the massive scale of raw satellite imagery (2401 × 2401 pixels) while maintaining computational efficiency.
Key Components:
- Spatio-Temporal Encoder-Decoder: The model uses convolutional patch embedding and merge blocks to compress high-resolution images into manageable feature maps. This structure effectively encodes temporal dynamics between consecutive frames.
- Spatial-Affine Transformer: To capture global patterns without prohibitive computational costs, DaYu utilizes a modified attention mechanism. A convolutional kernel reduces the dimensionality of Key ($K$) and Value ($V$) vectors, allowing the model to process large feature maps efficiently.
- Two-Stage Training: The model is first pre-trained to predict a single 30-minute step and then fine-tuned autoregressively to generate forecasts up to 12 hours (24 steps).
Datasets and Training
DaYu was trained on data from the Advanced Himawari Imager (AHI) aboard the Himawari-8/9 satellites. The dataset covers the Asia-Pacific region at 0.05° spatial and 0.5-hour temporal resolution. The model utilizes eight infrared (IR) channels (6.9 $\mu$m to 13.3 $\mu$m), enabling continuous 24/7 monitoring. Training utilized data from 2016–2018, while testing was conducted on 2023 data, representing a massive volume of high-frequency observations.
Performance and Results
DaYu demonstrates exceptional skill in maintaining structural details over long lead times. It achieves a Pearson Correlation Coefficient (PCC) above 0.9 for the first 3 hours and remains above 0.7 at the 12-hour mark.
Key Insights:
- Detail Retention: Unlike many AI models that produce blurry outputs due to MSE loss, DaYu uses residual learning to focus on frame-to-frame differences, preserving sharp edges in cloud structures.
- Channel Sensitivity: Performance is highest on the 6.9 $\mu$m channel (mid-tropospheric water vapor), which is less susceptible to low-level noise. The 11.2 $\mu$m channel proves most challenging due to surface-level variability and diurnal cycles.
- Mesoscale Accuracy: In case studies involving events like Super Typhoon Bolaven, DaYu accurately predicted the position and spiral structure of the typhoon eye 12 hours in advance, capturing details often missed by global models.
Conclusion and Future Work
DaYu represents a significant step toward high-resolution, observation-driven weather nowcasting. While the model currently forecasts brightness temperatures, future work will focus on linking these cloud images to specific meteorological variables such as precipitation and wind speed. Additionally, the team aims to further optimize the computational resources required for full-disk, high-resolution processing to make real-time deployment more accessible.
Summary updated