Predicting EvapoTranspiration Index using time-series ML algorithms

Droughts have become a major problem in every part of the world, including in my home state of California. This is reducing crop yields, raising prices for critical food supplies, and impacting global food security.

Evapotranspiration is the combination of evaporation from the soil and transpiration from plants. When farmers can assess the evapotranspiration, they can estimate how much water is being lost to the atmosphere, and plan irrigation accordingly. Reference evapotranspiration, or ET, is combined with crop specific coefficients to assess the ET for each plant. I use historical ET values to forecast future ET values, which can help farmers and farming authorities better manage our water resources.

Repository Structure

├── src/                         # Empty folder.
├── scripts/
│   ├── download_data.py
│   ├── install_dependencies.py  # This script installs all required Python packages.
│   ├── train_model.py           # Empty script, as we train our models on-the-fly.
│   └── launch_app.py            # This script launches the predictET App.
├── app/
│   └── app.py                   # The code for the Streamlit-based predictET App.
├── static/                      # Directory for assets for this documentation.
├── .project-metadata.yaml       # Declarative specification of this project.
├── README.md                    # This file!
└── requirements.txt             # Python 3 package requirements.

Creating the project on CML

I recommend manual setup:

1. In a CML workspace, click "New Project" and add a Project Name. Select "Git" as the Initial Setup option, copy in the repo URL https://github.com/danikagupta/predictET, and click "Create Project".
2. Launch a Python 3 Workbench Session with 4GiB RAM and run !pip3 install -r requirements.txt to install requirements. You may need to run this step twice to complete the installation after initial failure.
3. Then create a CML Application as described in the CML documentation, using scripts/launch_app.py as the script.

Launching the app

Once the CML Application has been created, you can launch it from the Applications pane. This should open a browser window, with a Streamlit application running at a URL similar to http://predict-et.cdsw.3.13.98.205.nip.io/.

If everything worked, you should see an application like this:

Using the app

Please use dark mode for the best viewing experience.

All controls are in the sidebar on the left side:

• Three different viewing modes: Two-column, Tabbed, and Full-page.
• Choose from several different choices of backgrounds.
• Choose from ten different cities in the Bay Area.
• How many months of original data to use for training.
• How many additional months to forecast.

The sidebar also shows:

• Map with the city location.
• Input data visualized as a graph.
• Raw input data.

Predictions are based on four models:

• Prophet
• SARIMA
• Theta
• Ensemble, with median of these three models

The main screen shows the following for each model:

• A graph showing the input data along with the model predictions for visual analysis.
• Six quality metrics - in general, the lower the better.

Here are six screenshots:

References

OpenET: Filling a Critical Data Gap in Water Management for the Western United States. Forrest S. Melton, Justin Huntington, Robyn Grimm, Jamie Herring, Maurice Hall, Dana Rollison, Tyler Erickson, Richard Allen, Martha Anderson, Joshua B. Fisher, Ayse Kilic, Gabriel B. Senay, John Volk, Christopher Hain, Lee Johnson, Anderson Ruhoff, Philip Blankenau, Matt Bromley, Will Carrara, Britta Daudert, Conor Doherty, Christian Dunkerly, MacKenzie Friedrichs, Alberto Guzman, Gregory Halverson, Jody Hansen, Jordan Harding, Yanghui Kang, David Ketchum, Blake Minor, Charles Morton, Samuel Ortega-Salazar, Thomas Ott, Mutlu Ozdogan, Peter M. ReVelle, Mitch Schull, Carlos Wang, Yun Yang, Ray G. Anderson