Interactions Between the Cryosphere, Climate and Greenhouse

Interactions Between the Cryosphere, Climate and Greenhouse Gases (Proceedings of IUGG 99
35
Symposium HS2, Birmingham, July 1999). IAHS Publ. no. 256, 1999.
A m u l t i c h a n n e l t h r e s h o l d t e c h n i q u e for N O A A
A V H R R d a t a to m o n i t o r t h e e x t e n t of s n o w c o v e r
in t h e Swiss A l p s
STEFAN VOIGT, MICHAEL KOCH &
MICHAEL F. BAUMGARTNER
Department of Geography, University ofBern, Hallerstrasse 12, CH-3012 Bern, Switzer
e-mail: [email protected]
Abstract A multichannel threshold technique is presented that allows
monitoring of snow cover in mountainous terrain such as the Swiss Alps using
NOAA-AVHRR data. The data preparation and calibration steps necessary to
apply the method on time series are described. While some of the classifica­
tion thresholds are stable with time others vary over the season. The classi­
fication skill proves to be better than that of supervised and unsupervised
classification approaches. Comparison with ground measurements displays
accordance between 60% and 90%. The method shows to be reliable and easy
to apply and is therefore well suited for operational snow cover monitoring.
INTRODUCTION
Frequent observation of the snow cover extent in the Alps is needed for various
purposes such as snow mapping, meteorological modelling, estimation of stored water
equivalent or snowmelt runoff prediction. Some of these applications even require
reliable information on the snow covered area (SCA) in near real-time. Within the
frame of the HYDALP project methods are developed to apply spaceborne remote
sensing observations of SCA for snowmelt runoff predictions. This paper describes the
methods developed by the Swiss contributors to HYDALP to process NOAA AVHRR
data efficiently in near real time.
Several methods for classification of AVHRR data with respect to snow have been
suggested in the literature. These include supervised (Baumgartner, 1994),
unsupervised (Hu et ai, 1993), linear mixture modelling (Simpson et ai, 1998) and
threshold techniques (see below). The present study shows that various physicallybased thresholds, which define whether a pixel is snow influenced or not, can be
combined into one threshold scheme to allow operational SCA mapping. As opposed
to other methods described in the literature this method is designed to work in
mountainous terrain. The method is applied to a time series of 15 AVHRR scenes
dating from the snowmelt periods of 1996 and 1997. Results of these applications are
displayed and compared with other conventional methods and ground measurements.
THE DATA
The Advanced Very High Resolution Radiometer (AVHRR) is a multispectral scanner
operationally employed on the polar orbiting weather satellites of the US National
36
Stefan Voigt et al.
Oceanic and Atmospheric Administration (NOAA) for the last 15 years. The scanner
has five spectral bands in the visible as well as in the near, mid and thermal infrared
part of the electromagnetic spectrum. The method described here is designed to work
on noontime AVHRR images, since they show the smallest topography-related
shading effects. Daily coverage, the spatial resolution of 1.1 km at nadir and its
specific channel setting make the AVHRR sensor very well suited for SCA monitoring
in areas larger than 200 km (Rango et al, 1985). In order to enable the application of
objective and standardized classification methods the data need to be calibrated and
normalized for various influences.
2
CALIBRATION
The purely reflective channels W\ and R2 (at 0.7 pm and 0.9 urn) are calibrated to top
of atmosphere reflectance by applying calibration equations and coefficients published
by Rao & Chen (1995). In order to minimize illumination effects in rough terrain the
albedo values are normalized to the actual local incidence angle using a COSapproach. The seasonal changes in illumination due to changes in earth-sun distance
(max. 8%) are compensated by normalizing the reflective channels to mean earth-sun
distance.
The thermal channels are calibrated to brightness temperature (BT) using the in
flight calibration devices of the AVHRR. This includes the nonlinearity corrections
necessary for channel 4 and 5 at 11 pm and 12 |xm respectively (Lauritson et al, 1979;
and Cracknell, 1997). The probably most important calibration step is to divide the
radiation measured in channel 3 (at 3.6 pm) into its emissive and reflective fractions.
The reflective part in this channel (R ) carries essential information for the
differentiation of snow and cloud. This decomposition can be achieved by estimating
the emissive part and subtracting it from the measured radiance. This is done by using
the purely thermal channels 4 and 5 and the assumption that the earth's emission can
be approximated through the Planck law (Gesell, 1989). As a final preparation step the
Normalized Differential Vegetation Index (NDVI) is computed since it is an effective
measure to support the separation of snow and vegetation. By applying the calibration
and normalization steps described above, a fair level of standardization is achieved
which allows intercomparison of data sets that were recorded at different dates during
the year.
3
THRESHOLD APPROACH
The concept of the threshold approach is to define boundary conditions for
determining whether a pixel in a satellite image is significantly influenced by snow or
not. Various such boundary criteria for delimiting that part of AVHRRs feature space
that corresponds to snow exist. None of the criteria alone is sufficient to separate snow
from other surface types or features. However, a combination of these criteria is well
suited to extract the snow-dominated pixels from an AVHRR data set. Saunders &
Kriebel (1988), Allen et al. (1990) Gesell (1989) and Derrien et al. (1993) have
Multichannel threshold technique for NOAA-A VHRR data to monitor the extent of snow cover 37
suggested different thresholds to classify snow or to separate snow from cloud. The
presented threshold scheme is partially based on these approaches. While some of the
authors had to account for special problems such as the recognition of sea-ice or the
differentiation of different cloud types, the approach presented here could be
simplified and is designed to reliably classify the following three classes only:
"no-snow", "snow" and "cloud". The method has been shown to work well even in
alpine terrain with its strong topography.
A combination of six thresholds applied in a sequential and hierarchical way was
found to give good results. The tests are arranged such that the less critical and
therefore less constrained thresholds are applied first and the stricter and therefore
No snow
Cloud
Cloud
No snow
Cloud
No snow
Snow
Fig. 1 The hierarchical threshold test sequence. Three tests are based on brightness
temperature (BT), one test on the Normalized Differential Vegetation Index (NDVI)
and two on albedo (R).
38
Stefan Voigt et al.
tighter ones are applied later within the test sequence. Whenever a pixel fails to pass
one of the snow tests it is assigned as "no-snow" or "cloud". Pixels that pass all the
tests are assigned as "snow" assuming that these pixels are dominantly influenced by
snow characteristics.
The six sequential tests implemented in the threshold scheme are displayed in
Fig. 1. In the following section the tests are explained in detail. Standard values for
the different thresholds are given below. However it might be necessary to adapt
single values for individual scenes interactively to achieve optimum results. This
should be done with great care and within physically reasonable range only. If the
user chooses to adapt the thresholds it is necessary to display and visually interpret
the image for interactive threshold measurement.
"Warm brightness temperature test": BT < BT max(snow) This test defines
the maximum brightness temperature that the class "snow" can have. Objects warmer
than this temperature are classified as "no-snow". Since it is the very first test, it has
to be set on the safe side, which means rather too warm than too cold with respect to
snow. In early winter the value can be set close to 2°C or 3°C. In spring or summer
this value has to be set to brightness temperatures between 10°C and 15°C.
4
4
"Cold brightness temperature test": BT > min BT (snow) This threshold
tests the lowest expected BT for snow. It is used to eliminate clouds that are colder
than the snow cover. This value should also be set "loose"—meaning rather too cold
than too warm. For the Alps this value can be set to -25°C to -30°C. During
springtime values can be raised to about - 1 5 ° C t o - 1 0 ° C .
4
4
"Cirrus test": B T - BT < ABT (cirrus) This test assesses whether the density
of cirrus clouds is sufficiently low for proper snow classification. It defines the optical
thickness of cirrus clouds from which a pixel is declared as "cloudy" and therefore is
rejected from further snow tests. It is based on the fact that thick cirrus clouds cause
a large temperature difference between channels 4 and 5. However, a thin cirrus
cover may still allow snow to be properly detected. A typical value for this threshold
would be 2 K. The larger the temperature differences, the thicker cirrus clouds have
be in order to be rejected. If the threshold is set to 1.5 K almost all cirrus clouds are
identified as "cloud", but at 3 K hardly any are identified as cloud. It is recommended
to keep this threshold still rather loose—meaning at larger temperature differences.
4
5
45
"Vegetation test": NDVI < min(vegetation) The NDVI is a very sensitive
indicator for the transition between snow and vegetated areas. Pure snow pixels have a
negative NDVI. However, it is recommended to set this value closer to the edge of
vegetation rather than to snow since clouds can have NDVI values of about 0 to 0.1. A
reasonable threshold for the NDVI is at about 0.1. However, this test only makes sense
if the region in which the method is applied is at least partially covered by vegetation.
"Water cloud test": R < max R (snow) This is the most important test to
distinguish between snow and low altitude water clouds. It is based on the fact that the
reflectance of snow in the mid-infrared is much lower than that of clouds. It is very
3
3
Multichannel threshold technique for NOAA-A VHRR data to monitor the extent of snow cover 39
effective and stable. The value for this threshold is at about 8% albedo. If the R
albedo is larger than this the pixel is declared as "cloud".
3
"Albedo test": Ri > min Ri(snow) This is the key snow test. It makes use of the
large albedo differences between snow and other land-use forms or water in the visible
channel of the AVHRR. It is set to values between 20% and 25% albedo. Experience
with this test showed that it should not be set below 18% and not higher than about
30% without very good reason. This test is not capable of separating snow from
clouds. If large parts of a scene are covered with thin cirrus it is recommended to
increase this threshold by 5% to 8% in order to account for the increase of albedo in
the visible part of the spectrum due to cirrus. Only if all tests are positive is the pixel
declared as dominated by snow.
RESULTS
The threshold scheme described above is applied for SCA monitoring within the
HYDALP project. So far it has been used on a time series of 15 AVHRR scenes
covering Switzerland and dating from the years 1996 and 1997. Fig. 2 displays the
development of the thresholds during the season. All thresholds were derived
independently through visual interpretation of the snow cover and measurements in the
calibrated data sets. The time series of thresholds show that some of them vary with
Q.
E
10 •
8
6•
4•
2
50
40
•a 30
<u 20
10
-
—•— ABT„,
-a-^D-
- -<i, p. -o- -c- - <> • -o-
0.45
0.30
0.15
0.00
—v— NDVI
30
o 15
0Q.
-15 E
-30
- BT
» BT
1
ir
Fig. 2 Development of thresholds over the season. Where the line connecting the
points is missing individual thresholds were not applied for classification.
40
Stefan Voigt et al.
time (BTmin, B T , and NDVI) while others remain almost constant (ABT , R|, and
R ). However, even the stable ones exhibit single deviations as for the Ri or R on
12 March for example. On this day a hazy fog layer covered parts of the scene.
However, it was so homogeneous and translucent that the snow cover could still be
mapped by manually adjusting the Ri and R thresholds. The same can be done in the
case of thin cirrus clouds through adapting Ri and ABT45 correspondingly. The 15
dates evaluated so far are not enough to establish fixed relations for the seasonal
dependency of the thresholds but the results imply that this is possible once a larger
time series has been derived. Such relations between threshold values and time would
allow classifying the scenes automatically and with only minimum interaction during
extreme weather situations.
max
45
3
3
3
COMPARISON WITH OTHER REMOTE SENSING TECHNIQUES
As a first quality assessment the derived snow maps are checked against the satellite
images on a visual basis. For this purpose standardized visualization settings were
defined, which enable reproducible results of the comparison. This is realized by
standard display lookup tables and fixed channel combinations, which allow optimum
"no-snow", "snow" and "cloud" differentiation. In order to achieve a more objective
assessment of the classification skill the results were tested against other snow
mapping techniques, which were applied on the same data. These include
supervisedAtnsupervised classification and spectral unmixing.
For the noontime AVHRR data set dating from 19 April 1996 all the above
methods where applied and the results compared with each other.
The supervised classification was generated using a maximum likelihood classifier
on five classes taken from training samples in the image. The training samples were
selected for different types of the following features: vegetation, urban area, water,
cloud and snow. An ISODATA clustering method was used for the unsupervised
classification. This generated 16 spectral clusters, which were then grouped
interactively into the classes: no-snow, snow and clouds.
Table 1 shows how the supervised classification method reproduces the
classification results of the threshold approach. The results show that the method
overestimates the snow cover when compared with the threshold results. While the nosnow classes agree up to 93% the supervised method classifies 2 1 % of the cloud pixels
as snow. This implies that the supervised method could have difficulties in
distinguishing between snow and cloud, which in fact is a well-known problem with
statistical classificators on AVHRR data. Another indicator of this fact is that the snow
class is reproduced by 100%, which would be very unlikely if not any of the methods
was overestimating the results of the other. Table 2 displays the same comparison with
the unsupervised classification results. In this case the snow and no-snow classes show
very good congruence, while the cloud class can only be reproduced by about 67%.
This again implies that the unsupervised method also overestimates snow and
underestimates the cloud cover. Table 3 displays how the unsupervised method
reproduces the results of the supervised method. The no-snow class is reproduced up to
99%, while snow and cloud obviously have poorer accordance. This gives even more
Multichannel threshold technique for NOAA-A VHRR data to monitor the extent of snow cover 41
Table 1 Confusion-matrix of the threshold classification vs the supervised classification for 19 April
1996.
Threshold-classification
Supervised classification:
No-snow
Snow
93%
6%
0%
100%
4%
21%
No-snow
Snow
Cloud
Cloud
1%
0%
75%
Table 2 Confusion-matrix of the threshold classification v.s the unsupervised classification for 19 April
1996.
Threshold-classification
Unsupervised classification:
No-snow
Snow
95%
4%
2%
96%
15%
18%
No-snow
Snow
Cloud
Cloud
1%
2%
67%
Table 3 Confusion-matrix of the supervised v.s the unsupervised classification for 19 April 1996.
Supervised-classification Unsupervised classification:
No-snow
Snow
No-snow
99%
1%
Snow
11%
85%
Cloud
13%
8%
Cloud
0%
4%
79%
evidence to the assumption that the supervised and unsupervised methods have
difficulties in distinguishing between snow and cloud.
For the spectral unmixing approach channels Rj and R2 as well as the reflective
part of channel 3 (R ) were used to decompose the spectrum of each pixel into the end
members snow, water, vegetation and cloud applying a linear mixture modelling
algorithm. The results were compared visually with the other classifications, since the
products are of different types (object classes vs fractional components). The method
showed good accordance to the threshold results. However the spectral unmixing of
AVHRR data is not yet well established and needs further investigation.
3
COMPARISON WITH GROUND MEASUREMENTS
Using 180 snow gauge stations distributed over the whole of Switzerland as ground
truth the classification results achieved with the hierarchical thresholding technique
were compared with supervised and unsupervised classification methods. If a station
measured a snow height of more then 10 cm the AVHRR pixel containing that station
was assumed to be snow covered. The results are summarized in Table 4.
A comparison was carried out for those pixels which were not cloudy and within
which a ground station was located. Table 4 shows that all three methods performed
similarly. However, the threshold method shows the highest absolute count of correct
observations: 144 (86.2% of the cloud-free pixels). All three of the methods produced
the same classification over 151 (83.9%) of the 180 pixels, of these pixels seven were
42
Stefan Voigt et al.
Table 4 Result of the classification of 19 April 1996 AVHRR using the hierarchical thresholding,
supervised and unsupervised classification (180 ground measurements/pixels).
Method
Threshold:
Supervised:
Unsupervised
Number of pixels classified as
cloud
13
10
16
Correct classifications
(% of cloud-free pixels)
144(86.2%)
134(78.8%)
136(82.9%)
cloud covered while 127 of the remaining 144 cloud-free pixels were correctly
classified (88.2%).
CONCLUSION
It could be shown that the presented multichannel threshold technique serves well for
SCA monitoring in the Swiss Alps. The method is simple and robust and is therefore
well suited for operational near real-time application, as within the HYDALP project for
example. Although the method requires some interactive steps carried out by an
operator, the procedure is well defined and objective.
Compared with the highly subjective methods of supervised and unsupervised
classification the presented threshold approach proves to be more accurate. Especially
at the transition between classes the thresholds exactly define boundary conditions
while the supervised or unsupervised approaches are difficult to control. Since the
latter are based on mean class spectra the boundaries are not sharply defined.
A serious problem remaining with SCA monitoring using optical sensors such as
the AVHRR is the possible cloud cover. Whenever there are only sparse clouds in the
scene a statistical three-dimensional kriging post-classification procedure is used to
interpolate SCA information for the cloud-covered areas. However, often enough the
whole processing is hindered by cloudy overcast. Other limiting factors for the method
presented above are large densely forested areas, since they cause underestimation of
the SCA due to masking, as well as extreme terrain in combination with low sun
elevations as in wintertime at high latitudes.
Acknowledgements This work was carried out within the framework of the European
Union research project HYDALP (ENV4-CT96-0364). We like to thank the Swiss
Avalanche Service (SLF) and the Swiss Meteorological Service (SMA) for the supply
of snow depth measurements.
REFERENCES
Allen, R. C, Durkee, P. A. & Wash, C. H. (1990) Snow/cloud discrimination with multispectral satellite measurements. J.
Appt. Met. 29.
Baumgartner, M. F., Apfl, G. & Holzer, T. (1994) Monitoring alpine snow cover variations using NOAA-AVHRR data.
In: Proc. 14th IEEE International Geoscience and Remote Sensing Symp. (Pasadena, California), 2087-2089.
Cracknel!, A. P. (1997) The Advanced Very High Resolution Radiometer (AVHRR). Taylor & Francis, London.
Derrien, M., Farki, B., Harang, L., LeGléau, H., Noyalet, A., Pochic, D. & Sairouni, A. (1993) Automated cloud detection
applied to NOAA-11/AVHRR imagery. Remote Sens. Environ. 46,246-267.
Multichannel threshold technique for NOAA-A VHRR data to monitor the extent ofsnow cover 43
Gesell, G. (1989) An algorithm for snow and ice detection using AVHRR data—an extension to the APOLLO software
package. Int. J. Remote Sens. 10, 897-905.
Hu Xu, Bailey, J. O., Barret, E. C. & Kelly, E. J. (1993) Monitoring snow area and depth with integration of remote
sensing and GIS. Int. J. Remote Sens. 14, 3259-3268.
Lauritson, L., Nelson, G. J. & Porto, F. W. (1979) Data extraction and calibration of TIROS-N/NOAA radiometers. NOAA
Tech. Mem. NESS 107, Department of Commerce, Washington, DC, USA.
Rango, A., Martinec, J., Foster, J. & Marks, D. (1985) Resolution in operational remote sensing of snow cover. In:
Hydrological Applications of Remote Sensing and Remote Data Transmission (ed. by B. E. Goodiso
Hamburg Symp., August 1983), 371-381. IAHS Publ. no. 145.
Rao, C. R. N. & Chen, J. (1995) Inter-satellite calibration linkages for the visible and near-infrared channels of the
Advanced Very High Resolution Radiometer on the NOAA-7, -9 and -11 spacecraft. Int. J. Remote Sens. 16, 19311942.
Saunders, R. W. & Kriebel, K. T. (1988) An improved method for detecting clear sky and cloudy radiance from AVHRR
data. Int. J. Remote Sens. 9, 123-150.
Simpson, J. J., Stitt, J. M. & Sienko, M. (1998) Improved estimated of the areal extent if snow cover from AVHRR data.
J. Hydrol. 204, 1-23.