Performance of machine learning algorithms for multi-step ahead prediction of reference evapotranspiration across various agro-climatic zones and cropping seasons

Abstract

This study forecasts multi-step-ahead Potential Evapotranspiration (ETO) in India using globally available fifth-generation European Centre for Medium-Range Weather Forecasts (ECMWF) gridded climate reanalysis products (ERA5). For this purpose, the potential of six machine learning approaches are examined across different agro climatic zones and cropping seasons. 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) models are first evaluated at the station level for up to 28-day-ahead prediction of daily ETO using meteorological station data as well as gridded ERA5 products. Using meteorological observations, at all three stations, SVR performs the best for a prediction horizon (PH) up to 2-days whereas LSTM performs the best for PH of 7-days. Using ERA5 datasets as input, among the three stations, the best performance is observed in Nagpur, where the best models for real-time and 28-day ahead prediction are LSTM (R2 = 0.847 and MAE = 0.474 mm/day) and RF (R2 = 0.722 and MAE = 0.635 mm/day) respectively. As the PH increases from 2-day to 28-day-ahead, models using ERA5 datasets performs better than those using station observations for most of the cases. Although the prediction performance drops initially with the increase in lead time, the drop in performance between 7-day and 28-day-ahead prediction is negligible. Evaluation of gridwise ETO prediction across entire India, using Global Land Evaporation Amsterdam Model (GLEAM) dataset as a reference, indicates MLP and CNN as the top performing models. Considering the crop seasons, model performance during Rabi season (October-March) ranged from 0.103 to 0.145 mm/day (MAE) and 0.977 to 0.988 (R2), which is better than the Kharif season (June-September) where MAE ranged from 0.140 to 0.234 mm/day and R2 ranged from 0.906 to 0.962. During the Rabi season, the ETO prediction performance of the arid agro-climatic zone is found to be superior to the other three agro-climatic zones, with the highest range of R2 (0.939 to 0.955) and lowest range of MAE (0.146 to 0.182 mm/day). Even the worst prediction performance, which is observed in the Humid region during the Rabi season, is also reasonably good (R2 = 0.656 to 0.79); thereby establishing the potential of the proposed models in multi-step ahead ETO prediction across various agro-climatic zones and cropping seasons.