theweathermanda/gewittergefahr

End-to-end machine-learning library for predicting thunderstorm hazards.

GewitterGefahr

GewitterGefahr is an end-to-end machine-learning library for predicting thunderstorm hazards, primarily tornadoes and damaging straight-line wind. The machine-learning methods are storm-centered, which means that each case is one storm object (one storm cell at one time step). "End-to-end" means that this library includes code for data acquisition and pre-processing; training, validation, and testing of machine-learning models; and post-processing of machine-learning output.

External documentation is still a work in progress (this README is currently the only external documentation). I plan to add external documentation in the coming weeks, and I hope to have it finished by early November 2018. However, keep in mind that this library changes a lot (I use it for most of my Ph.D. work), so at any given time some portion of it will be probably be undocumented. I will try to be clear about what that portion is.

Despite the lack of external documentation, there are three types of internal documentation. First, there is a docstring at the top of each method, explaining the inputs and outputs along with their formats (e.g., number, string, list, numpy array, etc.). Second, the variable and method names are verbose and include units where applicable (e.g., DRY_AIR_GAS_CONSTANT_J_KG01_K01, specific_humidities_kg_kg01), so the code is self-documenting to some extent. Third, most modules (Python files) are accompanied by unit tests. For example, the unit tests for moisture_conversions.py are in moisture_conversions_test.py. With that said, the unit tests are not exhaustive and there are no integration tests, so I make no guarantee that the code is bug-free.

Requirements

• numpy
• scipy
• tensorflow
• keras
• scikit-image
• netCDF4
• pyproj
• scikit-learn
• opencv
• matplotlib
• basemap
• pandas
• shapely
• ambhas
• descartes
• geopy
• metpy