bask0 / mach-flow

Rainfall-runoff modeling (Mach-flow project)

---

CH-RUN: A data-driven spatially contiguous
runoff monitoring product for Switzerland

Fig. 8 from Kraft et al. (in prep.): Spatially contiguous reconstruction of runoff from 1962 to 2023. The maps represent the yearly catchment-level ruoff quantiles relative to the reference period (1971 to 2000) empirical distribution. The bottom bars show the decadal deviation in mm y−1 of the national- level runoff relative to the reference period.

Description

This repository contains the code for performing deep learning based reconstruction of runoff for Switzerland (Fig. 1). More details can be found in this publication.

Fig. 1 from Kraft et al. (in prep.): From sparse observations with low human impact to contiguous coverage. The 98 observational catchments highlighted in magenta were selected by domain experts and served as a base for training and evaluating the data-driven models. Those catchments are only marginally affected by anthropogenic factors and are of similar size as the target catchments for reconstruction (grey).

How to run

First, make sure that you have conda installed. Then, install dependencies:

# Clone the project.
git clone https://github.com/bask0/mach-flow.git

# Install the project.
# You may need to change this line `source $CONDA_PREFIX/etc/profile.d/mamba.sh` in `create_env.sh`).
cd mach-flow
bash create_env.sh

Next, run the model tuning and cross validation. GPU should automatically be detected if available. This takes a while.

# Run LSTM tuning and cross validation.
bash basin_level/run_experiment.sh -m LSTM

# Run TCN tuning and cross validation.
bash basin_level/run_experiment.sh -m TCN

Create figures:

# Create figures.  
bash create_plots.sh

Citation

@article{kraft_chrun_2024,
  title={CH-RUN: A data-driven spatially contiguous runoff monitoring product for Switzerland},
  author={B. Kraft, W. Aeberhard, M. Schirmer, M. Zappa, S. I. Seneviratne and L. Gudmundsson},
  journal={HESS},
  year={in prep.}
}