Comparison of neuron-based, kernel-based, tree-based and curve-based machine learning models for predicting daily...

Area of Science:

• Agricultural Science
• Environmental Science
• Data Science

Background:

• Accurate reference evapotranspiration (ET0) prediction is essential for efficient irrigation scheduling and agricultural water management.
• Limited climatic data presents a challenge for reliable ET0 estimation.

Purpose of the Study:

• To evaluate the performance of eight machine learning models for daily ET0 prediction using temperature and precipitation data.
• To compare model accuracy across diverse climatic regions in China.

Main Methods:

• Utilized neuron-based (MLP, GRNN, ANFIS), kernel-based (SVM, KNEA), tree-based (M5Tree, XGBoost), and curve-based (MARS) models.
• Trained and tested models on daily data from 14 stations across China (2001-2015).
• Assessed prediction accuracy using R², RMSE, NRMSE, and MAE metrics.

Main Results:

• Machine learning models using only temperature data achieved satisfactory ET0 estimates (R² ≈ 0.829).
• Incorporating precipitation data improved prediction accuracy by 7.6% nationwide, especially in humid regions.
• Kernel-based (SVM, KNEA) and curve-based (MARS) models generally outperformed others, with SVM showing the best overall accuracy and stability.

Conclusions:

• Support Vector Machine (SVM) is highly recommended for daily ET0 prediction across China due to its accuracy and stability.
• Kernel-based KNEA and curve-based MARS models are also powerful and promising alternatives.
• Machine learning offers a robust approach for ET0 prediction with limited climatic data, enhancing water resource management.