Estimation of Daily Reference Evapotranspiration using Machine Learning and Deep Learning Techniques with Sparse Meteorological Data

Abstract

Accurate estimation of evapotranspiration is very crucial for enhancing the real time irrigation scheduling and decision making in water resources planning. Traditionally, empirical methods are used to calculate the reference evapotranspiration using available meteorological data. However, in many areas, such data is limited or unavailable for ETo estimation. Hence, this study aims to explore data-driven models like machine learning (ML) and deep learning (DL) for estimating ETo with minimal meteorological data. In this study, five ML models, including linear regression (LR), random forest (RF), support vector regression (SVR), XGBoost, KNN regression, and two deep learning methods such as feedforward neural networks and long-term short-term memory (LSTM), were used to estimate the reference evapotranspiration (ETo) over the Phulnahara canal command area using various combinations of meteorological parameters. The results of these models were compared with the Penman-Monteith-based ETo. The Penman-Monteith-based ETo exhibits a significant positive correlation with sunshine hour and maximum temperature, displaying correlation coefficients of 0.8 and 0.6, respectively, while RHmin and RHmax demonstrate a negative correlation. The findings revealed that when all climate data is available, the coefficient of determination (R2) rises to 0.98. However, when data is limited, it drops to 0.78. The SVR model outperformed other ML models with all input combinations. However, KNN emerged as the most reliable model for estimating ETo with input data of maximum and minimum temperature. Interestingly, we found that even using just three parameters (temperature, wind speed, and relative humidity) or two-parameter combinations (temperature and relative humidity or temperature and wind speed) can yield promising results in ETo estimation. The findings of this study offer valuable insights for estimating ETo in regions with limited climate data, which is crucial for effective agricultural water management.