A python package aims to exploit state-of-the-art hydrological timeseries prediction and forcasting.

The package should supports the surrogate training, parameter learning and inference application components of the drought forecasting InterTwin's use case.

This package is currently under development.

git clone https://github.com/interTwin-eu/hython.git

cd ./hython

pip install .

Please review the workflow notebook for a demonstration of the expected inputs, outputs, and how to use the package.

Please open an issue if you have a bug, feature request or have an idea to improve the package.

• Domain sampling

Training the model on large domains is time and energy consuming. This functionality samples the full domain producing a smaller subsample, with different degree of representativeness based on the sampling strategy, enabling decisions about the trade-off between model performance and computation time. It is likely that good enough performance can be achieved with representative sampling scheme.

Planned strategies: - no sampling (implemented) - regular grid sampling (implemented) - stratified sampling (coming soon) - spatial correlation sampling (coming soon)

• Spatio-temporal validation consisting in (at least) three options: space, time and spacetime.

This feature generates (spatially, temporally or spatiotemporally) disjointed training and validation subsets for testing how well the model is performing in extrapolation tasks.

• Simulation of river discharge

Surrogate's simulation of river discharge in addition to soil moisture and evapotranspiration

• Parameter learning Calibratig the surrogate

• Add metrics with hydrological meaning

• Parallel and Distributed ML tasks.

• Uncertainty & Explainable AI

• Model evaluation

Assessing different model architectures and structures

For further information please contact:

[email protected]

[email protected]