ClimaX via Radar

Regional radar-based rainfall forecasting using the ClimaX transformer with multi-variable radar inputs.

Problem

Accurate rainfall forecasting from radar data requires modeling highly non-linear spatiotemporal dynamics across multiple atmospheric variables.
Conventional numerical and heuristic methods struggle to jointly capture local convective structures and large-scale weather patterns, especially for longer forecast horizons.

This project investigates whether ClimaX, a transformer-based climate model, can be adapted for regional radar-based rainfall prediction.

Approach

Data

We used Korean radar composite data with the following characteristics:

• Temporal resolution: 10-minute intervals
• Spatial resolution: 100 m
• Original grid size: 2305 × 2881
• Projection: Lambert Conformal Conic (LCC)
• Target region: Seoul metropolitan area

Radar variables:

• HSP: Radar-estimated rainfall intensity (mm/hr)
• HSR: Rainfall derived from radar reflectivity
• HSR_dBZ: Raw radar reflectivity (dBZ)

Data preprocessing steps:

• Values ≤ −20000 treated as missing
• Rainfall values scaled by dividing by 100
• Reflectivity converted using the Marshall–Palmer Z–R relationship
• Spatial cropping and downsampling to 100 × 100 grids
• Dataset-wide normalization
• Construction of a climatology map for ClimaX conditioning

Days with missing radar channels were excluded.

Model

We adopted RegionalClimaX, a grid-to-grid transformer model originally designed for climate variables.

Model configuration:

• Patch size: 4 × 4
• Embedding dimension: 1024
• Transformer layers: 8
• Attention heads: 16
• Positional embeddings enabled

Each radar variable is tokenized independently and aggregated via cross-variable attention, allowing the model to learn dependencies between reflectivity and rainfall intensity.

Training

• Total samples: 188,071
• Time span: June 2021 – December 2023
• Train / validation / test split: 70% / 20% / 10%
• Batch size: 32
• Epochs: 10
• Forecast horizon: up to 72 hours
• Training from scratch on a single GPU
• Runtime: approximately 40 minutes per epoch

Multiple configurations were explored, including single-variable and multi-variable inputs.

Evaluation

Performance was evaluated using:

• Test accuracy
• Mean squared error
• Qualitative inspection of spatial rainfall patterns

Key observations:

• Multi-variable inputs improved accuracy over single-variable baselines
• Large-scale rainfall structures were captured reliably
• Fine-grained convective details were often smoothed
• Patch-wise artifacts appeared in longer-horizon forecasts

Notes and Lessons Learned

• ClimaX effectively models structured, multi-variable geophysical data
• Multi-variable conditioning is critical for radar-based forecasting
• Patch tokenization limits fine-scale detail in long-horizon predictions
• Future improvements may include:
  • Multi-scale patching
  • Temporal attention refinement
  • Hybrid convolution–transformer front-ends

This project served as a foundation for further exploration of transformer-based weather and climate forecasting using radar observations.