Wringing out moisture from deep learning: high-res (250m) soil moisture predictions across the Western USA to estimate tree seedling survival and forest regeneration.

Abstract:

Regeneration of tree seedlings is increasingly considered a demographic bottleneck in the resilience of forests to climate change and disturbance. Maybe more true for seedlings than adult trees, soil moisture is a critical determinant of survival, but the current suite of data sources on soil moisture lack the spatial resolution from coarse satellite data or coverage from field sensors for landscape-scale studies of seedling survival. As a result, the scientific community relies heavily on physical models to simulate soil moisture dynamics at spatial and temporal scales relevant to plant response. The high spatial and temporal resolution needed for such simulations limits the spatial and temporal extent of applications. To address this scaling issue, we developed a Long Short-Term Memory (LSTM) deep neural network model to emulate a high-resolution (250 m) physically-based distributed eco-hydrology model. Training the LSTM to emulate the physical model, and specifying physical constraints on the learning and prediction process, ensured that LSTM predictions were plausible even under conditions outside of the training range. To improve the generalization of the LSTM, we developed a large source of simulated and validated data representative of a broad range of climatic, topographic, and edaphic conditions.We optimized the physical model using daily mean-field soil moisture from satellite retrievals (Soil Moisture Active Passive, SMAP satellite) and gauged streamflow, ensuring that the physical model simulated realistic hydrological dynamics across 175 catchments distributed across the Western U.S. We verified that the physical model was able to reproduce temporal dynamics of SMAP retrievals (2015-2017), and independently, the land surface temperatures as compared from MODIS AQUA (2003-2017). We then trained the LSTM using these calibrated simulation data. We evaluate LSTM predictions against spatial and temporal out-of-sample subsets of model simulation data, SMAP, and LST retrievals. Finally, we integrated daily LSTM soil moisture predictions into a Soil-Plant-Atmosphere-Continuum model of plant hydraulic function to estimate the frequency, duration, and intensity of drought events that affect plant loss of hydraulic conductivity (PLC) in Ponderosa pine. The LSTM demonstrated high predictive power for simulating the spatial and temporal distribution of soil moisture evaluated on out-of-sample catchments. Simulations of PLC showed both broad scale sensitivity to climatic gradients and local sensitivity patterns related to the topographic redistribution of water, highlighting areas where forest regeneration might no longer be possible. We show that regional simulations of potential plant hydraulic stress are consistent with the low elevation distribution of forests. The use of the LSTM for simulating soil moisture dynamics results in an order of magnitude improvement in computational savings and makes mechanistically based studies of seedling mortality feasible on landscape to continental scales. We present insights gained during the development and integration of remote sensing data into a deep learning model, and present first results from LSTM soil moisture predictions applied to a case of tree seedling survival in the Western U.S.