monsoonbench 0.1.3

A reproducible benchmarking framework for Indian monsoon onset prediction

MonsoonBench

A unified, reproducible benchmarking framework for Indian monsoon onset prediction.

MonsoonBench provides a standardized workflow for loading rainfall and forecast datasets, computing monsoon onset, and evaluating forecasting skill across space and time.
It is designed for climate researchers, forecasters, and data scientists aiming to compare deterministic, probabilistic, and climatology-based onset models using consistent methods.

The framework follows WeatherBench-style principles: clean APIs, reproducible configuration, modular components, and shareable outputs.

---

Documentation Overview

MonsoonBench includes detailed module-specific guides. Use the links below to navigate the documentation.

Core Package Overview & Pipeline

High-level explanation of the evaluation pipeline, CLI interface, onset metrics, and NetCDF outputs.
Path: monsoonbench/README.md
Open Metrics & Pipeline README

---

Data Loading Guide

How to load IMD rainfall, deterministic/probabilistic forecasts, and threshold datasets using the unified API.
Path: monsoonbench/data/dataloader_quickstart.md
Open DataLoader QuickStart

---

Visualization & Metric Export Tools

How to generate spatial scorecards and export skill metrics in NetCDF, CSV, Parquet, or JSON formats.
Path: monsoonbench/visualization/README.md
Open Visualization README

---

Examples (Configs, Scripts, Notebooks)

Example YAML configs, runnable scripts, and tutorial notebooks demonstrating end-to-end usage.
Path: examples/README.md
Open Examples README

---

Installation

MonsoonBench is available on PyPI:

pip install monsoonbench

Verify installation:

monsoonbench --help

Python API Example

from monsoonbench.metrics import DeterministicOnsetMetrics
from monsoonbench.visualization import create_model_comparison_table

# Initialize metrics calculator
metrics = DeterministicOnsetMetrics()

# Compute metrics for multiple years
df, onset_data = metrics.compute_metrics_multiple_years(
    years=[2019, 2020, 2021, 2022],
    model_forecast_dir="data/model_forecast_data/fuxi/...",
    imd_folder="data/imd_rainfall_data/4p0",
    thres_file="data/imd_onset_threshold/mwset4x4.nc4",
    tolerance_days=3,
    verification_window=1,
    forecast_days=15,
)

# Create spatial metrics
spatial = metrics.create_spatial_far_mr_mae(df, onset_data)

# Generate comparison table
comparison = create_model_comparison_table({"FuXi": spatial})
print(comparison)

Repository Structure

monsoon-bench/
│
├── monsoonbench/ # Core package
│ ├── data/ # Dataloaders
│ │ └── dataloader_quickstart.md
│ ├── metrics/ # Onset detection + evaluation pipeline
│ ├── utils/ # Notebook-facing reusable helpers (e.g., onset_timeseries)
│ │ └── README.md
│ ├── visualization/ # Scorecards + metric downloaders
│ │ └── README.md
│ ├── README.md # Module-level pipeline documentation
│ └── ...
│
├── examples/ # Configs, scripts, tutorial notebooks
│ └── README.md
│
├── tests/ # Unit tests
├── Dockerfile
├── Makefile
└── pyproject.toml

Development Process with branches

Each team member created their own branch to implement specific fixes or features, such as the data loader, data downloader, and visualizations. We regularly merged these branches during TA meetings to ensure that the codebase stayed consistent and that everyone remained aligned on progress and design decisions.