11/09/2017
Deep Learning Methods Extend Soil Moisture Observations Seamlessly in Space and Time
New advanced analytic methods applied to soil data for enhancing Earth System Modeling.
The Science
A deep neural network method can learn soil moisture responses to atmospheric forcings and the patterns and conditions of land surface model errors, and by utilizing them, can extend soil moisture data decades back without auxiliary observations.
The Impact
The application of this new reflection method to Earth system codes has the potential to greatly enhance model productivity and code-reusability.
Summary
Soil moisture is a crucial variable affecting runoff, evapotranspiration, energy balance, and ecosystem functioning. It is simulated by most land surface models. Soil moisture observations can be used to evaluate model simulations and reduce uncertainties. The Soil Moisture Active Passive (SMAP) mission has delivered valuable sensing of surface soil moisture since 2015. However, it has a short time span with irregular revisit schedules. Deep learning techniques recently gained rapid adoption and are promoting transformational changes across many disciplines. However, it has not been widely employed in hydrology or Geosciences. Utilizing a state-of-the-art time series deep learning neural network, Long Short-Term Memory (LSTM), DOE-funded researchers created a system that predicts SMAP level-3 moisture product with atmospheric forcings, model-simulated moisture, and static physiographic attributes as inputs. The system removes most of the bias with model simulations and improves predicted moisture climatology, achieving small test root-mean-square errors (<0.035) and high-correlation coefficients >0.87 for over 75% of Continental United States, including the forested southeast. The study showed the proposed network avoids overfitting and is robust for both temporal and spatial extrapolation tests. It also showed that the inclusion of numerical models in a big data machine learning setting helps reduce prediction bias. LSTM generalizes well across regions with distinct climates and environmental settings. With high fidelity to SMAP, LSTM shows great potential for hindcasting, model evaluation, data assimilation, and weather forecasting. This study is the first use of state-of-the-art time series deep learning for hydrology or hydrologic variables.
Principal Investigator(s)
Chaopeng Shen
Pennsylvania State University
[email protected]
Funding
Partial support from the U.S. Department of Energy Office of Science, Biological and Environmental Research, Earth System Modeling Program.
References
Fang, K., C. Shen, D. Kifer, and X. Yang. “Prolongation of SMAP to Spatio-temporally Seamless Coverage of Continental US Using a Deep Learning Neural Network.” Geophysical Research Letters (2017). DOI: 10.1002/2017GL075619