The Impact of Land Cover Data on Rainfall-Runoff Prediction Using an Entity-Aware-LSTM

The code in this repository relates to the abstract The Impact of Land Cover Data on Rainfall-Runoff Prediction Using an Entity-Aware-LSTM presented as a vPICO presentation at the EGU General Assembly 2021 in the session HS3.4 - 'Deep learning in hydrological science'. All results and graphics can be reproduced following the README instructions and running the Jupyer Notebook in this repository.

Drost, S., Netzel, F., Wytzisk-Ahrens, A., and Mudersbach, C.: The Impact of Land Cover Data on Rainfall-Runoff Prediction Using an Entity-Aware-LSTM, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-1136, https://doi.org/10.5194/egusphere-egu21-1136, 2021.

Get started

To get prepared for running the code and reproducing the results you first have to set up your local environment by installing all required Python dependencies and downloading those datasets that are not included in this repository. Especially, we make extensive use of the great neuralHydrology Python package, which implements several Deep Learning models for use cases from the hydrology.

1. Make sure, you have locally installed a conda distribution (e.g., Miniconda or Anaconda)

2. Recreate the conda environment by running conda env create -f environment.yml from the conda command line and activate it with conda activate egu21-rrp-landcover-lstm

3. Install the neuralHydrology package by running pip install git+https://github.com/neuralhydrology/neuralhydrology.git. If you'd rather wish to install a specific version, you can clone the repository to your local machine by running git clone https://github.com/neuralhydrology/neuralhydrology.git, checkout a certain branch and install it with pip install -e .. For our studies, we used the 1.0.0-alpha release.

4. Download the CAMELS-GB dataset which contains hydro-meteorological timeseries for 671 catchments across Great Britain. This dataset is required for training the LSTM and EA-LSTM models for certain experiments. You can access the dataset via https://catalogue.ceh.ac.uk/documents/8344e4f3-d2ea-44f5-8afa-86d2987543a9 by creating an account for the Environmental Information Data Centre.

5. Extract the CAMELS-GB archive and copy the whole timeseries folder to ./data/camels-gb.

Project structure

• ./config: The ./config folder contains config files for training and evaluating the models via neuralHydrology used in 4 different experiments.

• ./data: The ./data folder holds all datasets, required for model training and evaluation. In particular, these are:

  • ./wv: Timeseries data and land cover frequencies as static attributes for 12 subbasins from the Wupper region as well as some geospatial datasets related to these.

  • ./camels-gb: After downloading the CAMELS-GB dataset and copying the timeseries folder, ./camels-gb contains timeseries data for the 671 CAMELS-GB catchments as well as land cover frequencies as static attributes.

• ./runs: After model training and evaluation, this folder contains the results created by the neuralHydrology package. These results will be used in the Jupyter Notebook for producing the figures.

• ./scripts: For convenience, this folder contains some scripts which will be utilized by the Jupyter Notebook.

Model training and evaluation

We trained and evaluated LSTM and EA-LSTM models to predict in 4 different experimental setups to predict rainfall runoff of 12 subbasins located in the Wupper region. You can train the models from scratch by using the configurations files provided within the ./config folder. For a detailed documentation of the neuralHydrology package have a look in its documentation.

Model training

To start training one of the models, run nh-run --config-file /path/to/config.yml. For instance, to train the model for the ealstm-camelsgb-wv experiment, run nh-run --config-file ./ealstm-camelsgb-wv.yml from the root of this repository. To train multiple models the neuralHydrology package supports certain execution parameters you can read about in the package documentation.

Model evaluation

After training, you can evaluate your models by running nh-run evaluate --run-dir /path/to/run_dir/. Just specify one of the generated run directories, which contains the model states.
Note: For ealstm-camels-wv and lstm-camels-wv you have to replace the config file within the created run directories by the ealstm-camels-wv-test.yml and lstm-camels-wv-test.yml configuration files, since we use CAMEL-GB for training and the Wupper region dataset for testing.

Run the code

For reproducing the figures you can run the ./analysis.ipynb via Jupyter Notebook. Start the server by running jupyter notebook from conda prompt and have fun.

License

Apache License, Version 2.0