Actual Evapotranspiration Framework for Arid Regions

This repository contains code for a comprehensive spatiotemporal framework to predict Actual Evapotranspiration (ETa) in Arid Regions, like the Central Valley of California using satellite data, machine learning techniques, and the Budyko Framework, as described in the study titled "A Novel Method to Estimate Actual Evapotranspiration using Modified Budyko Model under Data Scarcity Conditions in Arid Regions."

Overview

The project leverages:

• Google Earth Engine (GEE) for data collection
• Fitting different Budyko prameter (Fu, Turc-Pike, Zhang, and Wang-Tang)
• Machine learning models (Random Forest Regressor, AdaBoost Regressor, Bagging Regressor, Extra Trees Regressor, Gradient Boosting Regressor, Hist Gradient Boosting Regressor, Extreme Gradient Boosting, Ridge, Least Angle Regression (Lars), RidgeCV, Linear Regression, Multi-Layer Perceptron Regressor - MLPR, K-Nearest Neighbors - KNN, and Decision Tree Regressor - DTR) for ETa prediction
• SHAP analysis for feature importance
• Hyperparameter tuning
• Input preprocessing
• Model training and validation
• Feature selection
• Taylor diagram
• Hybrid Budyko-Deep learning models

The study covers data from 2001 to 2022 and includes climatic, landuse, and vegitation variables.

Key Results

1. Model Performance

• Best Models:

  • Wang-Tang-XGB7: NRMSE=0.06, R²=0.90, NSE=0.90, RMSE=3.21 mm

  • Turk-Pike-XGB7: NRMSE=0.06, R²=0.90, NSE=0.90, RMSE=3.23 mm

  • Zhang-XGB7: NRMSE=0.06, R²=0.90, NSE=0.90, RMSE=3.27 mm

  • Fu-XGB7: NRMSE=0.06, R²=0.89, NSE=0.88, RMSE=3.59 mm

  • Budyko model (Zhang): NRMSE=0.10, R²=0.78, NSE=0.77

  • Budyko model (Turc-Pike): NRMSE=0.10, R²=0.77, NSE=0.76

  • Budyko model (Fu): NRMSE=0.11, R²=0.77, NSE=0.76

  • Budyko model (Wang-Tang): NRMSE=0.11, R²=0.77, NSE=0.75

  • ML (XGB7): NRMSE=0.11, R²=0.73, NSE=0.73

  • ML (RF3): NRMSE=0.12, R²=0.70, NSE=0.70

  • ML (GB5): NRMSE=0.13, R²=0.68, NSE=0.68

  • ML (Ada6): NRMSE=0.14, R²=0.62, NSE=0.60

2. Key Predictors (SHAP) - Feature Selection

  - XGB7: ['ppt', 'lat', 'soil_water_1', 'r_max', 'Srad', 'lon', 'soilT_4']
  - RF3: ['ppt', 'r_max', 'lat']
  - GB5: ['ppt', 'lat', 'r_max', 'ss_flux_max', 'st_r_d_max']
  - Ada6: ['ppt', 'r_max', 'ss_flux_max', 'lai_high', 'pr', 'st_r_d_max']

3. Budyko Parameter (ω)

• Monthly Averaged Time scale:
  • Fu : 9.3
  • Turc-Pike : 8.6
  • Zhang : 5
  • Wang-Tang : 0.91
• long-term monthly average Time scale(LT monthly):
  • Fu : 4.1
  • Turc-Pike : 3.3
  • Zhang : 4.5
  • Wang-Tang : 0.8
• Annual Time scale:
  • Fu : 2.22
  • Turc-Pike : 1.46
  • Zhang : 0.39
  • Wang-Tang : 0.45

Evaluation ML Models

*Figure 2: Left - Taylor diagram of ETa prediction Machine Learning models during the calibration; Right - validation magnitudes*

Hyperparameter tuning

Model	Hyperparameters
Random Forest Regressor	Max. Depth: None Number of Trees: 50 Min. Samples to Split: 5 Best Nodes: 20 Bootstrap Samples: None
Gradient Boosting Regressor	Learning Rate: 0.1 Sample Fraction: 0.75 Boost Stages: 200 Max. Depth: 6
AdaBoost Regressor	Max. Estimators: 50 Regressor Weight: 0.25 Loss Function: Linear
Extreme Gradient Boosting	Number of boosting rounds: 100 Maximum depth of a tree: 3 Step size shrinkage: 0.1 Minimum loss reduction: 0.2 Subsample ratio (rows): 1 Subsample ratio (columns): 1

Repository Structure

• 1_1_era5_data.ipynb and 1_2_terraclimate_data.ipynb: Jupyter notebook for collecting satellite and meteorological data using Google Earth Engine (e.g., TerraClimate, ERA5).

• 1_3_eddy_data.ipynb: Input preprocessing.

• 2_1_budyko_models.R: Core modeling notebook implementing all 4 Budyko models.

• 2_2_ml_models.ipynb: Core modeling notebook implementing all 14 ML models.

• 2_3_hybrid_models.ipynb: Hybrid Budyko-Deep learning models using neural network

• 3_1_scatter_plots.ipynb: Scatter plot for ML and Budyko Models.

• 3_2_taylor_diagram.R: Taylor Diagram.

• 3_3_sensitivity_analysis.ipynb: SHAP analysis.

• 3_4_time_series_change.R: Time series scatter.

• requirements.txt: List of Python dependencies required to run the code.

Prerequisites

• Python: Version 3.8 or higher.
• Google Earth Engine Account: Required for data collection. Sign up here.
• Dependencies: Install required libraries using:

  pip install -r requirements.txt
  ## Key Dependencies
  

The main Python dependencies include:

• earthengine-api
• pandas
• numpy
• scikit-learn
• xgboost
• shap
• matplotlib (for visualizations)
• pyarrow
• xarray
• plotly
• scipy
• tqdm
• seaborn
• pyet
• datetime
• geopandas
• shapely
• wxee
• keras
• tensorflow

Setup Instructions

Clone the Repository

git clone git@github.com:mahdimohammadnezhad/Estimation_ETa.git
cd Estimation_ETa

Setup Instructions

Install Dependencies

pip install -r requirements.txt

Authenticate Google Earth Engine

Run the following command and follow the prompts to authenticate:

earthengine authenticate

Contact

For support or inquiries:

• 🐛 Report Issues:
  Open a new issue on GitHub
• 📧 Email Contact:
  mohammadnezhad@alumni.um.ac.ir
• 💬 Discussion Forum:
  Consider adding a GitHub Discussions tab to your repository for community questions

Citation

@article{mohammadnezhad2025novel,
  title={A novel hybrid model for actual evapotranspiration estimation in data-scarce arid regions: Integrating modified Budyko and machine learning models using deep learning},
  author={Mohammadnezhad, Mahdi and Davary, Kamran and Shirazi, Pooya and Rezvanpour, Mohammad Javad and Hasheminia, Seyed Majid},
  journal={Science of The Total Environment},
  volume={1001},
  pages={180438},
  year={2025},
  publisher={Elsevier}
}