Probabilistic estimation of glacier surface elevation changes from DEM
differentiation: a Bayesian method for outlier filtering, gap filling
and uncertainty estimation with examples from High Mountain Asia
Abstract
Various interdisciplinary studies have shown substantial discrepancies
between modeled and remotely sensed glacier surface elevation change.
It is therefore crucial to better understand and quantify uncertainties
associated to both methods.
We design a probabilistic framework with the aim to filter outliers, fill
data voids and estimate uncertainties in glacier surface elevation
changes computed from Digital Elevation Model (DEM) differentiation.
The technique is based on a Bayesian formulation of the DEM difference
problem and specifically targets surging and debris-covered glaciers,
both at glacier and regional scales.
We first define a set of physically admissible surface elevation changes
as an elevation-dependent probability density function. In a second
step, terrain roughness is defined as the main descriptor for DEM
uncertainty. Each surface elevation change pixel is a probability
distribution. We present validation experiments in High Mountain Asia
and show that the model produces results consistent with conventional
DEM differencing, while avoiding the caveats of already existing
methods.
We further demonstrate that accounting for unstable glacier dynamics is
crucial for accurate outlier filtering and robust uncertainty estimation.
The technique can be applied to other types of remotely sensed glacier
quantities (surface velocity
etc.) and so would help to improve the characterization of uncertainty
associated with changes in glacier mass and dynamics.