Reconstructing a 1980s DEM of the Greenland ice sheet for model initialization

Abstract One of the current limitations of reproducing the past evolution of the Greenland ice sheet with numerical models is the lack of a digital elevation model (DEM) prior to the year 2000 that can provide an initial state based on historical observations. Here we combine historical DEMs from aerial photography of the margins with laser altimetry data from the interior to reconstruct a complete DEM at 600 m grid resolution that represents the state of the Greenland ice sheet in the 1980s. To this end, we apply two methods: First, we use Gaussian process regression constrained by empirically derived spatial covariance functions. We find that after standardization, the signal is stationary and correlated over more than 100 km, allowing for interpolation over large data gaps. In a second, more novel approach, we use the singular value decomposition to extract spatial patterns from a series of model-generated DEMs. The Gaussian process regression shows smaller cross-validation errors and yields robust uncertainties; for the second method, we expect it could reduce nonphysical transient behavior in the historical simulations. With these reconstructions, we provide initial states that predate the earliest high-resolution ( less than 1 km $ \lt 1~\mathrm{km}$ 1   k m ) DEM of the Greenland ice sheet by around 20 years.