SimMod Rainfall Prediction (Radar Reflectivity DBZH → 10-min Ground Precipitation) — Group Project Baseline Demo

Project Overview

This is a baseline demo pipeline for our SimMod group project. The goal is end-to-end validation: load radar files → match timestamps → extract simple features → train baseline regression → save plots/logs. It is not meant to be a final model, just a reproducible baseline.

Project Goal

• Input: Weather radar reflectivity DBZH snapshots over Cologne region, stored as HDF5 *.hd5 files under radar/
• Target: 10-minute ground precipitation measured at stations, provided as CSV labels
• Task: Supervised regression (predict precipitation intensity from radar information)

What is Implemented (Current Baseline)

• Recursively scans radar/ for files ending with *.hd5
• Parses timestamps from radar filenames
• Matches each precipitation timestamp to the nearest radar scan (within ±10 minutes)
• Reads DBZH values from the radar HDF5 structure and converts them to dBZ using gain/offset metadata
• Extracts simple global radar features from each radar frame:
  • mean DBZH
  • max DBZH
  • 90th percentile DBZH
• Trains a baseline regression model (LinearRegression)
• Saves a run log and diagnostic plots into outputs/

Data Layout (Local)

Raw radar data is large and therefore not meant to be committed to GitHub. Users should keep radar/ locally and the repo should stay lightweight.

Expected local folder structure:

radar/
  2021/
    07/
      07/
        .../*.hd5
    08/
      .../*.hd5
    ...
  2022/
    01/
      .../*.hd5
    ...
data/
  precip_10min_long_balanced.csv
  Station_Features_Lookup.csv
outputs/
  [created at runtime]

CSV Files

• data/precip_10min_long_balanced.csv: Label dataset with 10-minute precipitation values and timestamps (used as target y)
• data/Station_Features_Lookup.csv: Station lookup / metadata mapping table used by the pipeline

How to Run the Demo

From the repo root:

python --version
pip install -r requirements.txt
mkdir -p outputs
python src/run_demo.py > outputs/run_log.txt 2>&1
tail -n 60 outputs/run_log.txt

Output Files

The demo saves outputs into outputs/:

• outputs/run_log.txt: Run summary (radar file count, paired sample count, MAE/RMSE, etc.)
• outputs/fig_radar_example.png: Example radar DBZH image (sanity check)
• outputs/fig_true_vs_pred.png: Scatter plot of true vs predicted precipitation (baseline model)
• outputs/fig_error_hist.png: Error histogram (pred - true)

Baseline Results

Metrics are sanity-check values, not final performance. The baseline uses only global summary statistics, so it is expected to be limited.

Success at this stage means:

• the pipeline runs end-to-end
• timestamp matching works reliably
• feature extraction is consistent
• outputs are reproducible

Notes / Next Steps

Potential improvements for the next iteration:

• Better spatial features around station locations (pooling/aggregation, not global stats)
• Multi-elevation / vertical structure features
• Stronger baselines (Ridge/Lasso, tree-based models, neural networks)
• Proper time-based train/val/test split to avoid leakage
• Scale evaluation to longer spans (full year / multi-year)

Contributors

• Beshoy Hanna Ayaad Labib (blabib@smail.uni-koeln.de)
• Fariba Yazdanjooei (fyazdanj@smail.uni-koeln.de)
• Mohammed Fawaz Nawaz (mnawaz@smail.uni-koeln.de)
• Pelin Su Kaplan (pkaplan@smail.uni-koeln.de)
• Raj Nandini Singh (rsingh8@smail.uni-koeln.de)
• Udayan Mishra (umishra@smail.uni-koeln.de)

License

Educational / university course project, no commercial use intended.

Citation / References

Radar reflectivity (DBZH) and station precipitation data are course-provided datasets. If the instructors provide an official dataset citation/link, it will be added here.

Known Limitations / Assumptions

• Only global radar summary statistics (mean/max/p90), ignores spatial structure
• Radar → precipitation mapping is highly non-linear and depends on physical factors not represented in baseline
• Nearest-neighbor timestamp matching with fixed ±10 minute window may introduce mismatch
• Results depend on locally available radar files and paired sample count

Reproducibility Notes

• Dependencies listed in requirements.txt
• Keep radar/ out of GitHub
• Store run outputs in outputs/ and keep outputs/run_log.txt for exact run record