Innovative approach for predicting daily reference evapotranspiration using improved shallow and deep learning models...

Area of Science:

• Hydrology
• Agricultural Science
• Data Science

Background:

• Accurate Reference Evapotranspiration (ETo) estimation is critical for water management, hydrology, and agriculture.
• The FAO-56 Penman-Monteith (FAO-56PM) is the standard but requires extensive data, often unavailable globally.
• Shallow and deep learning models offer alternatives for ETo estimation with limited data.

Purpose of the Study:

• To compare shallow learning (SL) and deep learning (DL) models for daily ETo estimation against the FAO-56PM standard.
• To evaluate the effectiveness of a semi-supervised pseudo-labeling (PL) technique in enhancing model performance.
• To identify the optimal meteorological variable combinations for ETo prediction in data-scarce coastal regions.

Main Methods:

• Employed Catboost Regressor (CBR) as a novel SL model and 1D-CNN, LSTM, GRU as DL models.
• Integrated a semi-supervised pseudo-labeling (PL) technique with all tested models.
• Developed six scenarios using various combinations of meteorological variables (temperature, humidity, wind speed, etc.).

Main Results:

• The PL technique effectively reduced systematic errors during model training across all scenarios.
• Input combinations including Tmin, Tmax, Tmean, and RH yielded superior performance for all models.
• The CBR-PL model demonstrated the best generalization, accuracy, and stability, with lower computational cost than DL models.

Conclusions:

• The CBR-PL model is a highly effective and recommended tool for predicting daily ETo, particularly in data-limited coastal environments.
• Shallow learning models, enhanced with PL, can be a viable and efficient alternative to complex deep learning approaches for ETo estimation.
• Meteorological data including temperature and relative humidity are key predictors for accurate ETo modeling.