Transfer learning augmented enhanced memory network models for reference evapotranspiration estimation

Abstract

Accurate estimation of reference evapotranspiration is a prerequisite for achieving sustainable development, designing irrigation plans, and efficient utilization of water resources. Traditionally, several conventional and statistical methods have been developed for this time-series prediction task. However, these methods are not well capable of handling the non-linear complexities, historical dependencies of the data and are majorly external features driven. In recent years, machine learning and deep learning models have gained significant attention for achieving state-of-the-art results in nearly all engineering domains. In this context, the present research study proposes three deep learning-based models to estimate future evapotranspiration values. Initially, a moving window-based baseline model is proposed for prediction by utilizing the minimum available historic evapotranspiration data only. The proposed model provides support for handling historical data dependencies by employing the Long Short Term Memory Network (LSTMN) model. Then, the concept of transfer learning is introduced/augmented for improving the prediction performance of the baseline model. Two transfer learning augmented deep learning prediction models (TL−LSTM1 and TL−LSTM2) are developed to forecast the future reference evapotranspiration values. The application of the proposed three models (Baseline, TL−LSTM1 and TL−LSTM2) is demonstrated on the dataset of nine different locations of State Punjab, India. The performance of the proposed models is compared with the prediction results of three supervised benchmark models in terms of well-known performance measures. The overall prediction results indicate that the transfer learning-based models achieve promising results.