Comparison of Neural network based and Kernel based Machine Learning approaches for daily forecasting of reference evapotranspiration in data scarce regions

Efficient estimation and forecast of reference evapotranspiration (ET_O) is crucial for water resources management and for developing an efficient irrigation practice that will help better utilization of scanty water resources. This is a more challenging task in data scarce regions. This study aimed at multi-step ahead prediction of ET_O across different cropping seasons and agro-climatic regions in India with six Machine learning (ML) based techniques using globally available fifth-generation European Centre for Medium-Range Weather Forecasts (ECMWF) gridded reanalysis products (ERA5). For real-time, one-day, two-day, and seven-day ahead prediction of ET_O, this study evaluates and compares the capability and prediction accuracy of deep learning algorithms, i.e., Support Vector Regression (SVR), Multivariate Adaptive Regression Splines (MARS), Random Forest (RF), Multi-Layer Perceptron (MLP), one-dimensional Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM). ML based models were developed using meteorological observations along with ERA5 inputs at three meteorological stations: Nagpur, Hyderabad and Bhubaneswar. The results indicate that MLP, SVR and CNN outperform other ML algorithms. High performing models, one each (MLP and SVR) from neural network-based models and kernel-based models respectively, are further utilized to scale up the analysis for gridwise ET_O forecasting across the whole of India. The Global Land Evaporation Amsterdam Model (GLEAM) dataset has been used as reference to evaluate gridwise ET_O forecasts. ET_O predicted by MLP model shows better agreement with GLEAM ET_O values during the Rabi cropping season (October-March) (MAE = 0.103 mm/day and NRMSE = 3.9 %) than during the Kharif season (June-September) (MAE = 0.151 mm/day and NRMSE = 4.5 %). As expected, the accuracy of the models drops with increase in the prediction horizon from real-time to seven-day; for instance, with MLP, MAE = 0.146 mm/day, R² = 0.955 for real-time and MAE = 0.173 mm/day, R² = 0.939 for seven-day ahead prediction over arid agro-climatic zone during Rabi season. However, even the minimum forecast performance observed in the semi-arid tropics region during Rabi season is reasonably good (MAE = 0.396 mm/day, R² = 0.704 for real-time evaluation and MAE = 0.445 mm/day, R² = 0.56 for seven-day ahead). This strengthens the potential of the proposed models for multi-step ahead ET_O forecasting across varied cropping seasons and agro-climatic regions without depending on meteorological station data.