Published: 1 Feb 2026
Accurate short-term rainfall prediction — known as precipitation nowcasting — plays a crucial role in many sectors, including flood prevention, transportation planning, agriculture, and disaster management.
However, predicting precipitation is challenging because rainfall patterns evolve dynamically across both space and time.
In this work, we introduce a deep learning framework that models rainfall prediction as a spatiotemporal graph problem, enabling the model to learn how precipitation evolves across regions and over time.
Rainfall is not an isolated phenomenon.
Two key processes drive precipitation evolution:
Spatial interactions
Storm systems move across geographical regions
Rainfall in one location influences neighboring areas
Temporal dynamics
Rain intensity evolves over time
Past precipitation strongly affects future rainfall patterns
Traditional deep learning approaches often focus on only one of these aspects. This research instead proposes a model designed to capture both spatial and temporal dependencies simultaneously.
To model these relationships, the study reformulates precipitation forecasting as a spatiotemporal graph sequence problem.
In this representation:
each geographic region is represented as a graph node
rainfall maps are treated as high-dimensional tensors stored at each node
connections between nodes represent spatial relationships between regions
This graph structure allows the model to learn how precipitation patterns propagate across space while also tracking how they evolve through time.
A central component of the proposed model is a dual-stream attention mechanism.
Instead of combining spatial and temporal information directly, the architecture learns them separately through two specialized streams.
The spatial stream models relationships between different regions at the same time step.
This mechanism allows the model to learn how precipitation in one location affects neighboring regions, capturing spatial dependencies between weather systems.
The temporal stream models relationships within the same region across different time steps.
Since precipitation patterns evolve continuously, understanding how past rainfall affects future rainfall is essential for accurate predictions.
Temporal attention allows the model to capture these dynamic temporal dependencies.
After extracting spatial and temporal patterns independently, the architecture combines them using a gated fusion module.
This fusion mechanism enables the model to:
integrate spatial and temporal information
balance their relative importance
produce more accurate precipitation forecasts
By explicitly separating and then integrating these dependencies, the model captures complex rainfall dynamics more effectively.
Another important contribution of the work is its ability to process high-dimensional precipitation maps directly within graph nodes.
Many graph neural network models require node features to be flattened into one-dimensional vectors, which can lead to information loss.
In contrast, the proposed model operates directly on 3-dimensional precipitation tensors, preserving the spatial structure of radar data and enabling richer feature extraction.
The proposed model is evaluated using seven years of hourly precipitation data from the Copernicus Emergency Management Service.
The dataset includes precipitation maps across multiple regions in Europe, represented as nodes in a spatial graph.
The model is compared against several existing precipitation nowcasting approaches, including:
Persistence baseline
SmaAt-UNet
RainNet
Across multiple prediction horizons and graph configurations, the proposed architecture consistently demonstrates improved forecasting performance.
Beyond predictive accuracy, the study also analyzes the learned spatial and temporal attention patterns.
These visualizations provide insights into:
which regions influence each other during precipitation events
how past rainfall contributes to future predictions
seasonal variations in spatial dependencies
Such analyses help improve the interpretability of deep learning models for meteorological applications.
Precipitation nowcasting is inherently a spatiotemporal problem, requiring models to understand both the movement of weather systems across space and their evolution through time.
By introducing a dual-stream attention architecture combined with graph neural networks, this work provides a powerful framework for capturing these complex dependencies.
The results demonstrate that explicitly modeling spatial and temporal interactions can significantly improve rainfall forecasting performance, highlighting the potential of graph-based deep learning methods for next-generation weather prediction systems.
Vatamány, L., & Mehrkanoon, S., Graph Dual-stream Convolutional Attention Fusion for precipitation nowcasting, Engineering Applications of Artificial Intelligence.