PE&RS February 2015 - page 119

An Improved Liberal Cloud-Mask for
Addressing Snow/Cloud Confusion with MODIS
Jeffery A. Thompson, David. J. Paull, and Brian G. Lees
The utility of the daily
snow products depends on the
ability of the
cloud-masking algorithm to differentiate
between snow- and cloud-cover. Although few studies have
explored the issue, snow/cloud confusion is a key issue limit-
ing the accuracy of the
snow products. Recent studies
from the Southern Hemisphere suggested that snow/cloud
confusion limited the utility of the
snow products there.
In this study,
snow/cloud confusion over Australia was
investigated using an improved liberal cloud-mask in con-
junction with a snow-detection algorithm. The performance of
the proposed cloud-mask was assessed using high-resolution
imagery and in situ observations. Results indicated that
the improved liberal cloud-masking algorithm reduced snow/
cloud confusion, successfully identifying snow-covered pixels
that were previously identified as cloudy. The analysis further
suggested that scale-related differences in imagery used in
the standard
cloud-masking workflow might be the
source of some snow/cloud confusion previously reported.
In recent years, remotely sensed time-series have been
increasingly used for monitoring climate related changes
in snow-covered areas (Vaughan
et al.
, 2013). Most studies
exploring these changes have focused on the Northern Hemi-
sphere, which is not surprising considering that snow-cover
is spatially extensive and plays an important role in both
climatological and hydrological processes there (Barry, 2002).
In contrast, outside of Antarctica, snow-cover in the Southern
Hemisphere is largely confined to alpine environments with
limited spatial distribution (Barry, 2008). Early attempts to
operationally monitor snow-cover across the entire Southern
Hemisphere were abandoned, as the spatial resolution of early
satellite platforms was too coarse to depict the finer scale het-
erogeneity of snow affected land surfaces (Dewey and Heim,
1983). In theory, the advent of the MODerate Resolution Im-
aging Spectrometer (
) with its high temporal resolution,
wide view angles and its standardized workflow processes,
resulted in the availability of daily snow-cover observations
at 500 × 500 m for the entire surface of the Earth, including
the Southern Hemisphere (Hall
et al.
, 2002).
In practice, the operational utility of the
products depends on an ability to differentiate between
snow-cover and clouds, as snow/cloud discrimination affects
the accuracy of these snow products (Hall and Riggs, 2007).
Differentiating between snow and cloud remains a challeng-
ing problem, because the spectral reflectance properties of
clouds and snow are governed by the same physical phenom-
ena. Cloud spectral reflectance is principally a function of the
diffraction and absorption of electro-magnetic radiation by
constituent water molecules, with the number of water drop-
lets, their cross-sectional area (equivalent spherical radius),
droplet density and depth (optical thickness) within the local-
ized areas of cloud-cover being the governing factors (Chahine
et al.
, 1983). Similarly, the optical properties of snow-cover
are also governed by diffraction and absorption characteristics
of snow crystals within the snowpack, with the equivalent
spherical radius, density and optical thickness of the snow-
pack being of primary importance (Wiscombe and Warren
1980). Clouds are often characterized as having smaller parti-
cle sizes relative to snow (Dozier, 1989), though the spherical
radii of actively precipitating clouds are similar to those of
fine snow particles (Chahine
et al.,
1983). That snow crystals
form within clouds and that optically thin cloud cover can
overlay snow-cover present additional challenges for snow/
cloud discrimination in remotely sensed imagery.
Considering the potential for confusing snow with cloud
and cloud with snow, it is surprising that
confusion has received relatively little attention within the
snow literature. In their synopsis of
snow product
accuracy, Hall and Riggs (2007) collated results from earlier
studies (e.g., Ault
et al.,
2006; Bitner
et al.,
2002; Klein and
Barnett, 2003; Maurer
et al,.
2003), indicating that the
snow products were ~93 percent accurate under cloud-
free conditions. Rittger
et al.
(2013) recently compared 172
snow-covered Landsat scenes with daily (MOD10A1) images
from four different study areas across the Northern Hemi-
sphere. Using similar thresholds to those employed in the
snow products, Rittger
et al.
(2013) found that their
precision descriptor for the daily
snow products (MO-
D10A1) ranged from 68 to 94 percent across four sites. Rittger
et al.
(2013) also noted that the accuracy of the MOD10A1
products was influenced by vegetation, viewable fraction of
snow-cover, and snow-cover heterogeneity.
A similar accuracy assessment of the MOD10A1 product
over Australia was conducted by Bormann
et al.
(2012), who
compared snow-cover mapped in 358 Landsat images with
MOD10A1 results. Bormann
et al.
(2012) reported that their
precision descriptor for the
products was 30 percent
across the Australian Alps. These results from Bormann
(2012) sharply contrast with those of Rittger
et al.
suggesting that there are further limitations with the MO-
D10A1 products that have not been reported in the literature.
Although they did not provide quantitative details, Sirguey
(2009) noted that the accuracy of the
snow product
School of Physical, Environmental and Mathematical
Sciences, University of New South Wales, Australia, P.O. Box
7916, Canberra, BC 2610, ACT 2600, Australia, (jeffery.thomp-
Photogrammetric Engineering & Remote Sensing
Vol. 81, No. 2, February 2015, pp. 119–129.
© 2014 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.81.2.119
February 2015
83...,109,110,111,112,113,114,115,116,117,118 120,121,122,123,124,125,126,127,128,129,...170
Powered by FlippingBook