All plots are produced by the modules under src/eruption_forecast/plots/ plus src/eruption_forecast/label/label_plots.py. The pipeline auto-renders the most useful plots; you can also invoke each helper directly for ad-hoc figures.

All of these are re-exported from eruption_forecast.plots:

from eruption_forecast.plots import (
 plot_tremor, plot_forecast, plot_significant_features,
 plot_roc_curve, plot_label_distribution, apply_nature_style,
)

Multi-panel band-decomposed tremor plot. One panel per band/method (RSAM/DSAR/entropy).

from eruption_forecast.plots import plot_tremor
fig = plot_tremor(tremor_df, methods=["rsam", "dsar", "entropy"])

Auto-rendered when fm.calculate(plot_daily=True, save_plot=True):

{station_dir}/tremor/figures/{nslc}_{YYYY-MM-DD}.png

Auto-rendered by fm.train(..., plot_features=True).

Three-panel forecast figure consumed by PredictionModel.forecast():

┌────────────────────────────────────────────────────┐
│ Panel 1 Consensus max-envelope prediction │
│ + probability, with threshold line │
├────────────────────────────────────────────────────┤
│ Panel 2 Per-classifier predictions overlaid │
│ with the consensus envelope │
├────────────────────────────────────────────────────┤
│ Panel 3 Per-classifier probabilities overlaid │
│ with the consensus envelope │
└────────────────────────────────────────────────────┘

Auto-rendered by fm.predict(plot_threshold=0.7, plot_pdf=True, eruption_dates=[...]):

{station_dir}/prediction/figures/forecast_{basename}.png
{station_dir}/prediction/figures/forecast_{basename}.pdf

Key kwargs forwarded via **plot_kwargs from fm.predict(...):

To re-render a forecast plot from the persisted CSV:

from eruption_forecast.plots import plot_forecast_from_file
fig = plot_forecast_from_file(
 "output/VG.OJN.00.EHZ/result_all_model_predictions_2025年07月27日_2025年08月22日.csv",
 eruption_dates=["2025-08-02"],
)
fig.savefig("forecast.png", dpi=200, bbox_inches="tight")

These are the aggregate plots EvaluationModel.evaluate(plot_aggregate=True) renders per classifier:

Auto-rendered at:

{station_dir}/evaluation/{kind}/classifiers/{classifier}/figures/*.png

ClassifierComparator.plot_all() adds cross-classifier figures under evaluation/{kind}/comparison/figures/ — see Evaluation Workflow → Cross-Classifier Comparison.

Debug-friendly bar plot showing the positive/negative class balance across the labelled window range:

from eruption_forecast.plots import plot_label_distribution
fig = plot_label_distribution(label_df)
fig.savefig("labels.png", dpi=150)

Useful when tuning window_step and day_to_forecast — flips the imbalance immediately visible.

The current evaluation flow has SHAP plotting stubbed via evaluate(plot_shap=True) — the call is accepted but emits a warning instead of rendering. Per-seed SHAP plots will return once the follow-up rebuilds them from the (y_proba, y_pred, y_true) matrices persisted by MetricsEnsemble. When that happens, remember to pass plot_size=None to shap.plots.beeswarm so SHAP does not override the pre-created figsize.

Every figure goes through apply_nature_style() from plots/styles.py, which sets:

• Serif fonts and small-format figure sizes consistent with Nature/Science columns
• The Okabe–Ito palette by default; a sequential brewer palette for diverging signals
• mpl.rc("pdf", fonttype=42) so saved PDFs keep editable text
• nature_figure(width_in, height_in) helper for one-line publication-quality sizing

Override per call:

import matplotlib.pyplot as plt
from eruption_forecast.plots import apply_nature_style
apply_nature_style()
fig, ax = plt.subplots(figsize=(6, 3))
ax.plot(df.index, df["consensus_eruption_probability"])
fig.savefig("custom.png", dpi=300, bbox_inches="tight")