Forecasting daily evapotranspiration using artificial neural networks for sustainable irrigation scheduling

Abstract

Evapotranspiration (ETo) forecasts can play an important role in irrigation scheduling and water resource management. Three state-of-the-art deep learning models, including long short-term memory (LSTM), one-dimensional convolutional neural networks (1D-CNN), and convolutional LSTM (ConvLSTM), were tested to forecast ETo on a daily time scale for seven days period. Daily air temperature (maximum, minimum and mean values), solar radiation, relative humidity, and wind speed data were collected for the 2011–2017 period from three weather stations (Harrington, North Cape, and Saint Peters) across Prince Edward Island (PEI), Canada. The relative importance method was used to determine the best-suited input variables for the models. The deep learning models were evaluated with the walk-forward validation technique using statistical measures of root mean square error (RMSE) and coefficient of determination (R²). The FAO Penman–Monteith modified equation (FAO-56) was used as the reference method for comparison purposes. For calibration and validation evaluation in annually daily ETo forecasts, the hybrid ConvLSTM model recorded lower errors than CNN and LSTM with the lowest calibration and validation daily RMSE of 0.64 and 0.62, 0.81 and 0.81, and 0.81 and 0.70 mm/day for Harrington, North Cape, and Saint Peters weather stations, respectively. The robustness and accuracy of these forecasted models may help farmers, water resource managers, and irrigation planners with improved and sustainable water management at the basin level, and for irrigation scheduling at farm/field level.