Improving GRACE spatial resolution to small river basin scale in Ireland using a deep learning model - a twenty-year study.

Since the deployment of the GRACE mission in 2002, data products from the mission have found multiple applications in studies of the physical earth. Some of the greatest successes amongst these applications have been achieved in the tracking of groundwater dynamics over large terrestrial regions, and to some extent over smaller regions where the underlying geomorphology is reasonably homogenous. Studies over smaller regions with mixed bedrock, soil and aquifer types have encountered challenges, mostly due to the limitations of the spatial resolution of the GRACE data sets. The current literature describes several different schemes to overcome this spatial resolution challenge, many of which employ machine learning models.

 

This study aims to improve the spatial resolution of GRACE measurements over Ireland by incorporating a number of additional data sets into a deep learning model. In the context of groundwater dynamics monitoring from orbit, the Irish land mass presents a number of unique challenges. The island of Ireland is entirely covered by about eleven GRACE one-degree pixels, most of which straddle a land-sea interface. In addition, the underlying bedrock and aquifer types are highly inhomogeneous with large regions, particularly in the west of the country, consisting of porous limestone aquifers that  have characteristically different groundwater flow dynamics from the gravel aquifers that underly much of the rest of the country.

 

The geological features of Ireland are well mapped by the Geological Survey of Ireland (GSI) so that the locations of the different aquifer types is reasonably well known. The Irish Environmental Protection Agency (EPA) operates a network of automated wells that record groundwater levels at fifteen-minute intervals.  Both of these primary data sets formed valuable inputs and validation for our models. Other input data sets included meteorological data, MODIS NDVI and LST, GLDAS climate variables, GPM precipitation and SRTM DEM.

 

The groundwater depth time series from the EPA wells were clustered using a k-means clustering algorithm to classify wells that behaved similarly in response to the natural precipitation and runoff cycles. These classifications aligned well spatially with the aquifer type maps and helped to identify the regions consisting of primarily gravel aquifer types. All data sets were resampled to 0.25-degree resolution using a ConvLSTM deep learning model.  The model was trained on GLDAS climate variables, MODIS NDVI and LST, GPM precipitation, SRTM DEM, EPA well maps and GSI aquifer type maps. The resulting predictions were validated with the EPA well data and as expected, the results exhibited heavily location-dependent correlations with the best agreement occurring in areas with predominantly gravel aquifer types. In these locations, the results suggest that this model can improve the resolution of GRACE measurements to 0.25-degree resolution quite well by capturing the spatiotemporal dynamics of this unusually varied hydrological system. At this resolution, GRACE measurements could form a key component of a more complete groundwater model for Ireland.