5 Channels

1
Capabilities and Limitations of Neural Networks
in Snow Cover Mapping from Passive Microwave Data
Juan C. Arevalo ( NOAA/CREST Graduate student)
jc_at_ce.ccny.cuny.edu, Hosni Ghedira (Assistant
Professor) ghedira_at_ce.ccny.cuny.edu, Reza
Khanbilvardi (Professor) rk_at_ccny.cuny.edu.
The City College of the City University of New
York. Convent Avenue at 140th St, Steinman Hall,
New York, NY. 10031.
1. Introduction
2. Study Area, Data Acquisition and Remote
Sensing Data
SSM/I Images
Snow-cover parameters are being increasingly used
as input to hydrological models. Having an
accurate estimation of the snow cover
characteristics during the snowmelt season is
indispensable for an efficient hydrological
modeling and for an improved snowmelt runoff
forecasts. Passive microwave remote sensing
techniques have been investigated by numerous
researchers using various sensors and have been
demonstrated to be effective for monitoring snow
pack parameters such as spatial and temporal
distribution, snow water equivalent (SWE), depth,
and snow condition (wet/dry state). However, the
snow products derived from passive microwave
sensors are usually limited by the relatively low
resolution, especially when the purpose is to use
this product as input for hydrological models.
Moreover, the accuracy of passive-microwave-based
maps is usually affected by the presence of
vegetation. In this project, we used an adaptive
neural network system to generate the spatial
distribution of snow accumulation from
multi-channel SSM/I data in the Northern Midwest
of the United States. Five SSM/I channels were
used in this experiment (19H, 19V, 22V, 37V, and
85V). The normalized difference vegetation index
(NDVI) is used from the AVHRR to quantify the
vegetation dynamic in the snow mapping process.
Six snow days with high snow accumulation have
been selected during the 2001/2002 winter season
to train and test the neural network system. The
snow depths and NDVI values have been compiled
and gridded into 25 km x 25 km grid to match the
final SSM/I resolution. To ensure an accurate
selection of training pixels, different
approaches have been tested by varying the
selection criteria of snow pixels. The final
results have shown the importance of these
selection criterions on the neural network
performance.
Six days have been selected during the 2001/2002
winter season (01/16, 01/17, 01/18, 01/23, 01/24,
and 01/25). The following illustrations show the
channel 19H of SSM/I satellite with 25 km
resolution for the six selected days.
Source of Data Ground data from National Oceanic
and Atmospheric Administration, NOAA SSM/I data
from DMSP SSM/I Pathfinder daily EASE-Grid
brightness temperatures, January 2002. Boulder,
CO National Snow and Ice Data Center, NSIDC
(Armstrong, R.L., K.W. Knowles, M.J. Brodzik and
M.A. Hardman. 1994,updated current year).
The study area is located in the Northern Midwest
of the United States within 1103748W -
1020224W and 484236N - 404348N.. The
passive microwave data from the NOAA/NASA
Pathfinder Program Special Sensor
Microwave/Imager (SSM/I) Level 3 Equal Area
Scalable Earth-Grid (EASE-Grid) Brightness
Temperatures F13 satellite is used in both
ascending and descending orbits. These images
provide measurements of the brightness
temperature in seven channels with different
frequencies and polarizations (19 V, 19 H, 22 V,
37V, 37 H, 85 V, and 85 H).
Original gridded data in Northern Hemisphere
projection with coordinates 11944W - 9957W
and 4936W - 3434W
8. Snow Cover Maps
6. Neural Network Results
3. Data Acquisition Vegetation
4. Artificial Neural Network
The available training data, has been divided
into three subsets
AVHRR Image
The following color images represent two
snow-cover maps for each selected day generated
from the artificial neural network output. Each
map contains 34 X 30 pixels with spatial
resolution of 25 km.
The graph below shows the accuracy variation of
100 neural network trained with different initial
configurations. Threshold 0.6 for the approach 4,
which yields the net with the highest accuracy.
That net was used to simulate the corresponding
snow maps.

• The first one is the learning set, whish is used
  for computing and updating the network weights.
• The second subset is the validation set, which
  is used for stopping the training by monitoring
  the validation error during the training process.
• The third subset is the test set that is not
  used during the training process, and it is only
  used to assess the classification accuracy and to
  compare between different classifiers and
  different network configurations.

The Normalized Difference Vegetation Index (NDVI)
has been derived from the visible and the
near-infrared channels of NOAA-AVHRR Sensor over
our study area.
Approach 4. Five channels,Tb standard
deviation of NDVI. Threshold 0.6
5 Channels
5 Channels St Dev NDVI
8-Km, 10-days composite NDVI image, January 21-31
1994.
The 5 channels maps (first column) represent the
simulation generated by simulating the trained
neural network in the approach 1, (threshold
0.4).
Jan 16
The 190 training pixels have been set as
follow Learning set 90 pixels Validation
set 45 pixels Test set 50
pixels
The AVHRR data was acquired from the Distributed
Active Archive Center (DAAC) located at Goddard
Space Flight Center, NASA.
The overall accuracy varies between 60 and 80 ,
having a fairly stable pattern. However, some
high and low peaks have been observed 86 and
52 being the highest and lowest accuracy. The
Kappa coefficient, which assess the agreement in
the classification between the snow and non-snow
pixels, has a very irregular pattern.
Jan 17
The original 8 Km NDVI values have been gridded
into a regular grid of 25 Km over the study area.
Effect of the decision threshold on classification
For each vector of five brightness temperatures
presented to the input layer, a value equal to
one will be assigned in the output layer if the
presented vector correspond to a snow pixel.
Otherwise, a value equal to zero will be assigned
to the corresponding vector.
The standard deviation values have been measured
during the gridding process to quantify the
vegetation homogeneity for each pixel.
During the simulation process, a continuous range
from zero to one will be produced by the output
neuron. We have introduced a threshold value
(between 0 and 1) to decide if the pixel will be
classified as snow or no-snow pixel. The
optimal threshold value cannot be identified with
certainty without measuring its effect on the
overall accuracy of the neural network
classification. In this project, the threshold
value has been varied from 0.2 to 0.8. The effect
of the decision threshold on classification
accuracy of each class is illustrated in the
following figure
The other maps (2nd column) represent the
simulation results of neural network trained
with the approach 4 by using 5 SSM/I channels
plus the standard deviation of the NDVI as input.
For this configuration, the best performance was
obtained with a threshold equal to 0.6.
Ground data distribution
Jan 18
A total of 195 ground stations covering the study
area have been identified for this experiment.
The figure below shows the distribution of the
ground stations over the study area (red
rectangle).
Graph shows the pattern and difference from the
truth data snow/no-snow and the ANN result
The input layer size may vary depending on the
approach used from 5 to 7 neurons.
Actual snow depth from the truth data and the ANN
result
Jan 23
5. Neural Network Approaches
Results obtained for the four approaches (number
of input channels) 1. the 5 brightness
temperature (Tb) channels 2. the 5 Tb plus
NDVI 3. the 5 Tb plus NDVI and standard
deviation, and 4. the 5 Tb plus the standard
deviation of NDVI.
Jan 24
Ground snow map
7. Confusion matrices
A total of 100 runs of the neural network for
each approach have been performed. The following
graphs show for each approach the average
accuracy and the Kappa coefficient for the
100-runs and their corresponding standard
deviation. These results have shown that the
addition of the NDVI standard deviation
(homogeneity factor) improves the snow
identification accuracy.
As a part of the assessment of the capabilities
and limitations of neural networks for snow
mapping, the neural network output has been
evaluated with a confusion matrix that was
computed for each approach. The overall accuracy
and Kappa coefficient were measured. The
following matrices correspond to the net giving
the highest accuracy out of 100 runs.
Only pixels with ground stations inside their
boundaries are considered for the training and
validation of Neural Network. A total of 165
pixels satisfy this criterion.
Jan 25