Deep Multi-Stage Reference Evapotranspiration Forecasting Model: Multivariate Empirical Mode Decomposition Integrated With the Boruta-Random Forest Algorithm

Abstract

Evapotranspiration, as a combination of evaporation and transpiration of water vapour, is a primary component of the global hydrological cycle. It accounts for significant losses of soil moisture from the earth to the atmosphere. Thus, reliable methods to monitor and forecast evapotranspiration are required for decision-making in many sectors. Reference evapotranspiration, denoted as ET, is a major parameter that is useful in quantifying soil moisture in a cropping system. This article aims to design a multi-stage deep learning hybrid Long Short-Term Memory (LSTM) predictive model that is coupled with Multivariate Empirical Mode Decomposition (MEMD) and Boruta-Random Forest (Boruta) algorithms to forecast ET in the drought-prone regions (i.e., Gatton, Fordsdale, Cairns) of Queensland, Australia. Daily data extracted from NASA’s Goddard Online Interactive Visualization and Analysis Infrastructure (GIOVANNI) and Scientific Information for Land Owners (SILO) repositories over 2003–2011 are used to build the proposed multi-stage deep learning hybrid model, i.e., MEMD-Boruta-LSTM, and the model’s performance is compared against competitive benchmark models such as hybrid MEMD-Boruta-DNN, MEMD-Boruta-DT, and a standalone LSTM, DNN and DT model. The test MEMD-Boruta-LSTM hybrid model attained the lowest Relative Root Mean Square Error (≤ 17%), Absolute Percentage Bias (≤ 12.5%) and the highest Kling-Gupta Efficiency (≥ 0.89) relative to benchmark models for all study sites. The proposed multi-stage deep hybrid MEMD-Boruta-LSTM model also outperformed all other benchmark models in terms of predictive efficacy, demonstrating its usefulness in the forecasting of the daily ET dataset. This MEMD-Boruta-LSTM hybrid model could therefore be employed in practical environments such as irrigation management systems to estimate evapotranspiration or to forecast ET.