Hurricane wind radius (R34) — 6h vs 1h LSTM with grouped SHAP

Predicts hurricane 34-kt wind radius (R34) at 6-hour best-track points and compares two models, then attributes each model's skill to three physics-based feature groups via grouped SHAP.

• 6h model (r34_6h.csv): every timestep is a labeled best-track observation.
• 1h model (r34_1h.csv): the full 1-hour sequence is fed to the LSTM, but the loss and evaluation are computed only at the 6h points (raw_data==True). The 5 intermediate hourly ERA5 steps provide sub-6h context. The 6-hourly best-track observations are fed only at the 6h points — intermediate steps are masked to zero and flagged by an obs_flag channel — while sporadic hourly-ERA5 gaps (~0.1%) are linearly interpolated within each storm. Standardization statistics come only from real values (best-track obs from the 6h points, ERA5 from all hours; obs_flag stays 0/1). Both models are evaluated on the same valid 6h test storms.

Feature groups (for grouped SHAP)

50 features in three physics groups (8 / 24 / 18 features; the grouping was recovered from the Excel header colours). A separate obs_flag channel — the 51st model input — gates the masked obs and is in no SHAP group. Group colours below match the SHAP bar charts (colour-blind-safe):

Group	Meaning
🟪 Location & Motion	DIST2LAND, USA_LAT/LON, STORM_SPEED/DIR, era_min_lat/lon, distance_dev
🟧 Intensity & Inner-core	USA_WIND/PRES/SSHS, uv_max, rmax*, u850/v850, warm_core, era_min_pressure + *_mslp_center twins
🟩 Large-scale Environment	u200/v200, shear, rh500 + *_mslp_center twins

SHAP is computed per group (exact permutation Shapley over all 3!=6 orderings), because storm-center and *_mslp_center features are ~0.99 correlated and per-feature SHAP would split their importance. Each group is revealed as a whole — from a baseline (the per-feature mean over labelled 6h steps) to its real values — in every ordering, and its averaged marginal change in predicted R34 is the group's Shapley value; importance is the mean |SHAP| in km at the 6h test points. Only forward passes are used, so it applies directly to the masked LSTM.

Run

Everything is in a single self-contained script. Edit only the CONFIG block at the top (the three data/output paths) — nothing else.

pip install -r requirements.txt        # torch, numpy, pandas, scikit-learn, matplotlib
python r34_6h_vs_1h.py

On Colab:

!pip install torch numpy pandas scikit-learn matplotlib
# set DATA_6H / DATA_1H / OUTPUT_DIR in the CONFIG block to your Drive paths, then:
!python r34_6h_vs_1h.py

Outputs (outputs/)

File	Description
scatter_6h.png, scatter_1h.png	Observed vs predicted R34 (with R²/RMSE/MAE)
shap_6h.png, shap_1h.png	Grouped-SHAP importance of the three feature groups
predictions_6h.csv, predictions_1h.csv	Test-storm predictions at labeled 6h points
shap_6h.csv, shap_1h.csv	Grouped-SHAP values used in the bar plots
metrics_summary.csv	R²/RMSE/MAE comparison table

A comparison summary (R²/RMSE/MAE for both models) is printed to the console.