The task of SWE retrieval from passive
microwave sensor measurements for large land
areas is a difficult task because of high spatiotemporal variability of snowpack cover and low
quality of ground observations, leading to poor
territorial zoning by snow cover constitution
parameterization. As described in Nosenko et al.
(2005), these factors determine the futility of
obtaining sufficient accuracy of snow cover
depth and SWE retrieval for hydrological applications on the whole territory of the Russian
Federation. At the same time, the regression
formula (2) gives poor quality of SWE retrieval
in the deforested central and polar regions of the
European part of the Russia, with the situation
becoming even worse at a thaw period, especially
when it is followed by rains, which leads to
formation of impervious for microwave radiation
ice layer above the show. The more favorable
conditions for SWE retrieval arise when the
percentage of the forest cover comes to 40%
(Nosenko and Nosenko 2007). To overcome the
described problems, it is proposed to find new
regression relations for extended range of
microwave radiation frequencies, which are more
reliable for the territories of sufficient climatic
and landscape conditions uniformity (Nosenko
et al. 2005; Nosenko and Nosenko 2007).
The Special Sensor Microwave Imager
(SSM/I), as described by National Snow & Ice
Data Center (https://nsidc.org/data/docs/daac/
ssmi_instrument.gd.html), was carried aboard
Defence Meteorological Satellite Program
(DMSP) satellites F8 (launched on June 19,
1987), F10, F11, F12, F13 and F15 (launched in
December 1999). The SSM/I is a seven-channel,
four-frequency, orthogonally polarized, passive
microwave radiometric system that measures
atmospheric, ocean and terrain microwave
brightness temperatures at 19.35, 37.0, 85.5 GHz
(vertical/horizontal polarization with spatial resolution of 25 km) and 22.2 GHz (vertical polarization only with spatial resolution of 12.5 km).
The data for all frequencies, organized in the
Level-3 Equal-Area Scalable Earth-Grid (EASEGrid) for both Northern and Southern Earth
Hemispheres, with spatial resolution of 25 km
and collected on daily basis by SSM/I radiometer
is freely available from the National Snow & Ice
Data Center (ftp://sidads.colorado.edu/pub/
DATASETS/nsidc0032_ease_grid_tbs/).
For the purpose of research, the daily
observed snow cover characteristic data at 117
in situ meteorological observation stations that
widely cover the territory of the Russian Federation (see Fig. 18.1) were selected from the
dataset, freely available from All-Russian
Research Institute of Hydrometeorological
Information–World Data Center (RIHMI-WDC).
All the selected meteorological stations are
located within 12.5 km radius relative to the
centers EASE-Grid cells.
The artificial neural network (ANN) is used as
a mathematical model. In fact, neural approach
supposes settling the mathematical relationship
between a set of input parameters X i and a set of
output parameters Y i which is established by
training on a certain group of samples. One of the
most popular ANN architectures is a multilayer
feed forward artificial neural network, which
consists of a set of interconnected processing unit
groups called layers (Fig. 18.2).
In the first experiment, the following six artificial neural networks were constructed and
trained for retrieval of snow water equivalent from
SSM/I measurements of brightness temperatures
for channels 19.35, 37.0 and 85.5 GHz of different polarizations over subset of EASE-Grid cells,
containing snow observation courses located in
terrain, which share common characteristics:
(a) Artificial neural networks for horizontal,
vertical and both orthogonal polarizations on
the all forested snow courses;
(b) Artificial neural networks for horizontal,
vertical and both orthogonal polarizations on
all open terrain snow courses.
The multilayer perceptron (MLP) architecture
was chosen for every constructed artificial neural
network. Depending on the actual combination
of selected channel polarizations, each MLP had
the number of neurons in the input layer ranging
from 3 (horizontal or vertical polarization) to 6
(polarization in both directions). The hidden
layer had 10 neurons with sigmoid activation
238
A. A. Volchak et al.
microwave sensor measurements for large land
areas is a difficult task because of high spatiotemporal variability of snowpack cover and low
quality of ground observations, leading to poor
territorial zoning by snow cover constitution
parameterization. As described in Nosenko et al.
(2005), these factors determine the futility of
obtaining sufficient accuracy of snow cover
depth and SWE retrieval for hydrological applications on the whole territory of the Russian
Federation. At the same time, the regression
formula (2) gives poor quality of SWE retrieval
in the deforested central and polar regions of the
European part of the Russia, with the situation
becoming even worse at a thaw period, especially
when it is followed by rains, which leads to
formation of impervious for microwave radiation
ice layer above the show. The more favorable
conditions for SWE retrieval arise when the
percentage of the forest cover comes to 40%
(Nosenko and Nosenko 2007). To overcome the
described problems, it is proposed to find new
regression relations for extended range of
microwave radiation frequencies, which are more
reliable for the territories of sufficient climatic
and landscape conditions uniformity (Nosenko
et al. 2005; Nosenko and Nosenko 2007).
The Special Sensor Microwave Imager
(SSM/I), as described by National Snow & Ice
Data Center (https://nsidc.org/data/docs/daac/
ssmi_instrument.gd.html), was carried aboard
Defence Meteorological Satellite Program
(DMSP) satellites F8 (launched on June 19,
1987), F10, F11, F12, F13 and F15 (launched in
December 1999). The SSM/I is a seven-channel,
four-frequency, orthogonally polarized, passive
microwave radiometric system that measures
atmospheric, ocean and terrain microwave
brightness temperatures at 19.35, 37.0, 85.5 GHz
(vertical/horizontal polarization with spatial resolution of 25 km) and 22.2 GHz (vertical polarization only with spatial resolution of 12.5 km).
The data for all frequencies, organized in the
Level-3 Equal-Area Scalable Earth-Grid (EASEGrid) for both Northern and Southern Earth
Hemispheres, with spatial resolution of 25 km
and collected on daily basis by SSM/I radiometer
is freely available from the National Snow & Ice
Data Center (ftp://sidads.colorado.edu/pub/
DATASETS/nsidc0032_ease_grid_tbs/).
For the purpose of research, the daily
observed snow cover characteristic data at 117
in situ meteorological observation stations that
widely cover the territory of the Russian Federation (see Fig. 18.1) were selected from the
dataset, freely available from All-Russian
Research Institute of Hydrometeorological
Information–World Data Center (RIHMI-WDC).
All the selected meteorological stations are
located within 12.5 km radius relative to the
centers EASE-Grid cells.
The artificial neural network (ANN) is used as
a mathematical model. In fact, neural approach
supposes settling the mathematical relationship
between a set of input parameters X i and a set of
output parameters Y i which is established by
training on a certain group of samples. One of the
most popular ANN architectures is a multilayer
feed forward artificial neural network, which
consists of a set of interconnected processing unit
groups called layers (Fig. 18.2).
In the first experiment, the following six artificial neural networks were constructed and
trained for retrieval of snow water equivalent from
SSM/I measurements of brightness temperatures
for channels 19.35, 37.0 and 85.5 GHz of different polarizations over subset of EASE-Grid cells,
containing snow observation courses located in
terrain, which share common characteristics:
(a) Artificial neural networks for horizontal,
vertical and both orthogonal polarizations on
the all forested snow courses;
(b) Artificial neural networks for horizontal,
vertical and both orthogonal polarizations on
all open terrain snow courses.
The multilayer perceptron (MLP) architecture
was chosen for every constructed artificial neural
network. Depending on the actual combination
of selected channel polarizations, each MLP had
the number of neurons in the input layer ranging
from 3 (horizontal or vertical polarization) to 6
(polarization in both directions). The hidden
layer had 10 neurons with sigmoid activation
238
A. A. Volchak et al.