Short- and mid-term forecasts of actual evapotranspiration with deep learning

Abstract

• Deep learning is applied for nowcasting and forecasting of evapotranspiration. • Weather variables and soil moisture are used as input for the deep learning models. • The models were applied to seven distinct climate zones within the continental U.S. • Results reveal great potential for evapotranspiration forecasting. Evapotranspiration is a key component of the hydrologic cycle. Accurate short-, medium-, and long-term forecasts of actual evapotranspiration (ET a ) are crucial not only for quantifying the impacts of climate change on the water and energy balance, but also for real-time estimation of crop water demand and irrigation water allocation in agriculture. Despite considerable advances in satellite remote sensing technology and the availability of long ground-measured and remotely sensed ET a timeseries, real-time ET a forecasts are deficient. Applying a state-of-the-art deep learning (DL) approach, Long Short-Term Memory (LSTM) models were employed to nowcast (real-time) and forecast (ahead of time) ET a based on (1) major meteorological and ground-measured (i.e., soil moisture) input variables and (2) long ET a timeseries from the Moderate Resolution Imaging Spectroradiometer (MODIS) onboard of the NASA Aqua satellite. The conventional LSTM and convolutional LSTM (ConvLSTM) DL models were evaluated for seven distinct climatic zones across the contiguous United States. The employed LSTM and ConvLSTM models were trained and evaluated with data from the National Climate Assessment-Land Data Assimilation System (NCA-LDAS) and with MODIS/Aqua Net Evapotranspiration MYD16A2 product data. The obtained results indicate that when major atmospheric and soil moisture input variables are used for the conventional LSTM models, they yield accurate daily ET a forecasts for short (1, 3, and 7 days) and medium (30 days) time scales, with normalized root mean squared errors (NRMSE) and Nash-Sutcliffe efficiencies (NSE) of less than 10% and greater than 0.77, respectively. At the watershed scale, the univariate ConvLSTM models yielded accurate weekly spatiotemporal ET a forecasts (mean NRMSE less than 6.4% and NSE greater than 0.66) with higher computational efficiency for various climatic conditions. The employed models enable precise forecasts of both the current and future states of ET a , which is crucial for understanding the impact of climate change on rapidly depleting water resources.