Quick Start Browse Models Browse Benchmarks

Overview

PyHazards brings together public dataset catalogs, registry-based models, benchmark families, experiment configs, and shared training or reporting workflows across wildfire, earthquake, flood, and tropical cyclone tasks.

It is designed for researchers and practitioners who need one coherent library for reproducing baselines, comparing methods, and extending hazard-ML workflows without rebuilding the software stack for each hazard family.

Hazard Families

4

Wildfire, earthquake, flood, and tropical cyclone workflows under one library.

Public Datasets

20

Curated dataset pages covering forcing sources and hazard-specific benchmark adapters.

Implemented Models

24

Public implemented baselines and variants surfaced through the model catalog.

Benchmark Families

4

Shared evaluator families with linked ecosystems, smoke configs, and reports.

Start Here

Use one of these four paths to move from overview to action quickly.

Quick Start

Run the first benchmark-aware workflow and verify the package.

Open: Quick Start

Browse Datasets

Explore forcing sources, benchmark adapters, and inspection entrypoints.

Open: Datasets

Browse Models

Compare implemented baselines, variants, and benchmark-linked model detail pages.

Open: Models

Browse Benchmarks

Compare hazard benchmark families, ecosystem mappings, and smoke coverage.

Open: Benchmarks

Why PyHazards

Unified Datasets

Public datasets, forcing sources, and inspection surfaces are documented through one hazard-first catalog.

Benchmark-aligned Evaluation

Shared benchmark families, smoke configs, and report exports make model comparisons more reproducible.

Registry-based Models

Baselines and adapters are exposed through a consistent build surface instead of one-off scripts.

Shared Training and Inference

One engine layer supports training, evaluation, prediction, and benchmark execution across hazard tasks.

Hazard Coverage

PyHazards spans four hazard families with public datasets, models, and benchmark pages designed to work together.

Wildfire

Danger forecasting, weekly forecasting, spread baselines, fuels, burn products, and active-fire sources.

Explore: Datasets | Models

Earthquake

Waveform picking, dense-grid forecasting adapters, and linked benchmark ecosystems for phase-picking workflows.

Explore: Models | Benchmarks

Flood

Streamflow and inundation baselines with benchmark-backed datasets, configs, and evaluation coverage.

Explore: Datasets | Benchmarks

Tropical Cyclone

Track-and-intensity forecasting baselines plus shared benchmark ecosystems and experimental weather-model adapters.

Explore: Models | Benchmarks

Explore the Docs

Installation

Set up PyHazards from PyPI or source and verify the environment.

Open: Installation

Quick Start

Run the shortest end-to-end workflow in the library.

Open: Quick Start

Datasets

Browse hazard-grouped dataset cards, detail pages, and inspection entrypoints.

Open: Datasets

Models

Compare implemented models, variants, and benchmark-linked detail pages.

Open: Models

Benchmarks

Review benchmark families, ecosystem mappings, and smoke-config coverage.

Open: Benchmarks

Reports and Configs

Load reproducible experiment YAML files and export benchmark summaries.

Open: Configs | Reports

For Contributors

PyHazards is registry-driven and uses dataset cards, model cards, and benchmark cards to generate the public catalogs. If you plan to extend the library, use Implementation Guide for the contributor workflow and Coverage Audit for the audited gap list behind the current roadmap work.

Citation

If you use PyHazards in your research, please cite:

@misc{pyhazards2025,
  title        = {PyHazards: An Open-Source Library for AI-Powered Hazard Prediction},
  author       = {Cheng et al.},
  year         = {2025},
  howpublished = {\url{https://github.com/LabRAI/PyHazards}},
  note         = {GitHub repository}
}

Community

Use the RAI Lab Slack channel for project discussion and coordination.