To allow a comparison between GPR snow depth measurements
and the ADS measurements, we assigned all individ
ual 18 136 GPR point measurements to the 22m ADS
raster and calculated the mean of all GPR values within
each cell, resulting in 1522 cells with GPR-based comparison
data. The variability of the GPR snow depth within these
cells amounted to between 0.1 and 0.3 m. Parts of the GPR
data have been obtained close to taller vegetation such as
trees and bushes. However, heavily affected measurements
have been masked out before comparison, as ADS data cannot
represent snow depth under forest canopy.
Comparing GPR to ADS data results in an overall RMSE
of 0.43m and an NMAD of 0.36 m. This is approx. 0.1m
worse compared to the reference data sets acquired at the
Wannengrat area. The overall correlation coefficient between
both data sets is 0.45 (Fig. 7a) only; note, however, that the
GPR data set features a significantly lower range in snow
depth when compared to the TLS data set (Fig. 6), mainly
because it was acquired at the valley bottom. When analyzing
different segments of the GPR data set we find considerable
differences. While the correlation is acceptable for individual
GPR segments that feature large snow depth variability
(Fig. 7b), it appears less favorable for GPR segments
with a small variability in snow depth (Fig. 7c). By comparing
the profiles of the snow depth values along the two
segments numbers 1 and 5 (Fig. 7d, e), we find the ADS values
to be too low over large parts of the transects. The agricultural
zones at the Dischma Valley bottom are covered by
grass with a length of 0.1 to 0.5m during summertime, when
the ADS data was acquired. This explains partially why the