Application of Deep Learning for Regional Evapotranspiration Projections Based on Satellite and Weather Observations

Abstract:

Spatiotemporal data about actual evapotranspiration (ETa) is of crucial importance for numerous Earth and environmental science applications that directly impact society and the environment. In‑depth knowledge of ETa dynamics is essential for weather forecasting, monitoring and projecting of drought conditions, conserving declining water resources, and for crop production, to name just a few. Over the past decade, satellite remote sensing techniques have been applied to quantify land surface variables and fluxes at large scales. Despite the impressive progress in improving the temporal resolution of satellite data, real-time estimates and projections are still deficient. Furthermore, currently applied ETa estimation methods mainly assimilate remote sensing (RS) observations and climate variables into surface energy balance models and unrealistically assume homogeneity of land surface properties. They also do not incorporate essential soil parameters that can significantly affect regional ET_a dynamics. Machine learning techniques such as deep learning (DL) have shown promise for efficient and automatic extraction of temporal features from time-series information. Deep learning not only can benefit satellite remote sensing applications for real-time estimation and projection of ET_a, but also enhance the accuracy of ET_a retrievals, which has not been attempted to date. Here we integrate long-term ET_a estimates obtained from MODIS satellite observations and weather forecasts into a DL model to provide daily estimates as well as projections of ET_a at regional scales. A time-series ET_a forecast model based on a long short-term memory (LSTM) based deep recurrent neural networks that is capable of learning multi-range and multi-level features from ground, satellite, and weather time series was developed. Weather parameter forecasts (i.e., precipitation, air temperature, wind speed, etc.) were acquired from the NOAA Global Forecast System. The LSTM-ET_a model estimates and projections were evaluated via comparison with eddy covariance ET_a estimates at U.S. AmeriFlux sites with diverse climates, soil types, and land covers.