Passive microwave satellite remote sensing systems (e.g., SMAP and SMOS) can provide reliable near-real-time observations of surficial soil moisture at a coarse resolution of 9-to-30 km. Recent efforts to downscale these observations have focused on the fusion of Visible-Infrared or Synthetic Aperture Radar imagery. However, these methods are limited by the availability of the ancillary satellite imagery datasets, and their resolution is limited to 1-to-2 km. The state-of-the-art in neural network downscaling methods is also limited by the extreme sparseness of soil moisture probe datasets. We present an alternative downscaling procedure that leverages fully-integrated groundwater-surface water models for their insights into the spatial distributions of soil moisture.
We constructed a feedforward neural network with 30-m aggregated input parameters including topographic wetness index, hydraulic conductivity, land cover class, and soil moisture observations interpolated from daily passive microwave soil moisture products. The neural network contained one input layer, two hidden layers, and one output layer. The novel aspect is that the neural network was trained on nodal soil moisture values predicted at daily intervals by a HydroGeoSphere model spanning Southern Ontario. With ~900,000 total nodes at surface, this vast training dataset spans ~75,000 km2 at a resolution of 10-to-500 m.
The trained neural network was able to delineate sharp features such as wetland boundaries, and produces plausible-looking soil moisture patterns over various geophysical features such as ravines and moraines. Neural network predictions were comparable to in-situ soil moisture probe data.
We demonstrate that a neural network can be trained using the outputs from a fully-integrated hydrological model and applied in practice for the downscaling of passive microwave soil moisture retrievals.