De6.2 report

Preliminary Algorithm Theoretical BasisDocuments for Snow/Ice/Frozen soil Properties

Fraction Cover, Water Equivalent, and Frozen/Thaw

Status

Deliverable De6.2

The WorkPackage 6 group1,2,3

1Cold and Arid Regions Envrironmental and Engineering Research Institute,

CAS, P.R. China

2 Institute Tibetan Plateau Research, Chinese Academy of Science, P.R.China

3 Beijing Normal University, Chinese Academy of Science, P.R.China

Dissemintation level: Programme ParticipantsLead beneficiary ID: CAREERI

ISSN/ISBN:c!2010

Edited by the CEOP-AEGIS Project O!ceLSIIT/TRIO, University of Strasbourg

BP10413, F-67412 ILLKIRCH Cedex, FrancePhone: +33 368 854 528; Fax: +33 368 854 531

e-mail: [email protected]

No part of this publication may be reproduced or published in any formor by any means, or stored in a database or retrieval system, without thewritten permission of the CEOP-AEGIS Project O!ce.

CEOP-AEGIS Report De 6.2

II

MODIS SNOW PRODUCTS ALGORITHM ABSTRACT......................................................................................................................................................... 1

1. INTRODUCTION ............................................................................................................................................. 2 1.1 Identification ......................................................................................................................................... 2

1.2 Overview ................................................................................................................................................ 3 2. ALGORITHM DESCRIPTION OF SNOW COVER.............................................................................................. 5

2.1 Introduction........................................................................................................................................... 5 2.2 Background and Data........................................................................................................................... 6

2.3 Calculation of ground reflectance ....................................................................................................... 8 2.4 Adjust of NDSI .................................................................................................................................... 10

2.5 Additional Algorithms ........................................................................................................................ 11 2.6 Image fusion ........................................................................................................................................ 12

2.7 Backup Algorithm............................................................................................................................... 13 3. ALGORITHM DESCRIPTION OF FRACTIONAL SNOW COVER ..................................................................... 13 4. VALIDATION PLAN ..................................................................................................................................... 13

4.1 Introduction......................................................................................................................................... 13 4.2 Approach ............................................................................................................................................. 14

4.3 Validation Sites.................................................................................................................................... 14 4.4 Auxiliary Measurements .................................................................................................................... 14

4.5 Scaling .................................................................................................................................................. 14 5. ANCILLARY DATA ...................................................................................................................................... 15

6. PROGRAMMING AND PROCEDURAL CONSIDERATIONS ............................................................................ 15 6.1 Programming Issues ........................................................................................................................... 15

6.2 Processing Issues ................................................................................................................................. 15 6.3 Quality Assurance ............................................................................................................................... 15

REFERENCES .................................................................................................................................................. 15

PART II

SNOW WATER EQUIVALENT RETRIEVAL ALGORITHM

ABSTRACT....................................................................................................................................................... 19 1. INTRODUCTION ........................................................................................................................................... 19

1.1 Identification ....................................................................................................................................... 19

1.2 Overview .............................................................................................................................................. 20 2. ALGORITHM DESCRIPTION ........................................................................................................................ 21

2.1 Introduction......................................................................................................................................... 21 2.2 Theoretical Basis of the Algorithm.................................................................................................... 24

2.3 Description of Retrieval Concept ...................................................................................................... 25 2.4 Description of Retrieval Algorithm ................................................................................................... 25

2.5 Backup Algorithm............................................................................................................................... 26 3. ALGORITHM PROTOTYPING ...................................................................................................................... 26

3.1 Data Analysis ....................................................................................................................................... 26 3.2 Prototyping of the Algorithm............................................................................................................. 28


III

4. VALIDATION PLAN ..................................................................................................................................... 29 4.1 Introduction......................................................................................................................................... 29

4.2 Approach ............................................................................................................................................. 29 4.3 Validation Sites.................................................................................................................................... 32

4.4 Auxiliary Measurements .................................................................................................................... 32 4.5 Scaling .................................................................................................................................................. 32

4.6 Data Protocols and Dissemination..................................................................................................... 32 4.7 Proposed Validation Tests.................................................................................................................. 32

5. ANCILLARY DATA ...................................................................................................................................... 32

6. PROGRAMMING AND PROCEDURAL CONSIDERATIONS ............................................................................ 33 6.1 Programming Issues ........................................................................................................................... 33

6.2 Processing Issues ................................................................................................................................. 33 6.3 Quality Assurance ............................................................................................................................... 33

References .................................................................................................................................................. 33

PART III

SURFACE SOIL FREEZE/THAW STATE DATASET USING THE DECISION TREE CLASSIFICATION ALGORITHM

ABSTRACT....................................................................................................................................................... 37

1. INTRODUCTION......................................................................................................................................... 37

1.1 IDENTIFICATION ....................................................................................................................................... 38

1.2 OVERVIEW ................................................................................................................................................ 38

2. ALGORITHM DESCRIPTION .................................................................................................................. 39

2.1 INTRODUCTION ......................................................................................................................................... 39 2.2 TARGETS TO BE OBSERVED ...................................................................................................................... 39

2.3 RADIATIVE TRANSFER PROBLEM ............................................................................................................ 39 2.4 MATHEMATICAL BASIS OF THE ALGORITHM ......................................................................................... 40

2.5 DESCRIPTION OF RETRIEVAL CONCEPT ................................................................................................. 41 2.6 DESCRIPTION OF RETRIEVAL ALGORITHM ............................................................................................ 41

2.7 BACKUP ALGORITHM............................................................................................................................... 41

3. ALGORITHM PROTOTYPING ................................................................................................................ 41

3.1 DATA ANALYSIS........................................................................................................................................ 41

3.1.1 Analysis of the brightness temperature characteristics of each land surface type .................... 41 3.1.2 Cluster analysis and decision tree for freeze/thaw status classification...................................... 44

4. VALIDATION PLAN................................................................................................................................... 45

4.1 INTRODUCTION ......................................................................................................................................... 45

4.2 APPROACH ................................................................................................................................................ 46 4.3 VALIDATION SITES ................................................................................................................................... 49

4.4 AUXILIARY MEASUREMENTS................................................................................................................... 49


IV

4.5 SCALING.................................................................................................................................................... 49 4.6 DATA PROTOCOLS AND DISSEMINATION ................................................................................................ 49

4.7 PROPOSED VALIDATION TESTS ............................................................................................................... 49

5. ANCILLARY DATA .................................................................................................................................... 49

6. PROGRAMMING AND PROCEDURAL CONSIDERATIONS............................................................ 50

6.1 PROGRAMMING ISSUES ............................................................................................................................ 50

6.2 PROCESSING ISSUES ................................................................................................................................. 50 6.3 QUALITY ASSURANCE .............................................................................................................................. 50

REFERENCES.................................................................................................................................................. 50

PART I

MODIS Snow Products Algorithm

Authors: Xiaohua Hao, Jian Wang, Hongyi Li, Zhe Li

Affiliations: Cold and Arid Regions Environment and Engineering Research Institute, Chinese Academy of Sciences.


1

MODIS Snow Products Algorithm

Abstract

The algorithms of MODIS Terra and MODIS Aqua versions of the snow products have

been developed by the NASA National Snow and Ice Data Center (NSIDC). The MODIS

global snow-cover products have been available through the NSIDC Distributed Active

Archive Center (DAAC) since February 24, 2000 to Terra and July 4, 2002 to Aqua. The

MODIS snow-cover maps represent a potential improvement relative to hemispheric-scale

snow maps that are available today mainly because of the improved spatial resolution and

snow/cloud discrimination capabilities of MODIS, and the frequent global coverage. In

China, the snow distribution is different to other regions. Their accuracy on Qinghai-Tibet

Plateau (QTP), however, has not yet been established. There are some drawbacks about

NSIDC global snow cover products on QTP:

1. The characteristics of snow depth distribution on QTP: Thin, discontinuous. Our research

indicated the MODIS snow-cover products underestimated the snow cover area in QTP

(Hao xiaohua, 2008).

2. The snow on QTP belongs to alpine snow. Errors due to the effects of topography can be

large. Without the terrain correction of a digital elevation model, the NSIDC global snow

products can underestimate the snow cover in QTP.

3. The snow products can separate snow from most obscuring clouds. However, there are

still many cloud pixels in daily snow cover product.

The study developed a new daily snow cover algorithm through improving the NSIDC

snow algorithms and combining MODIS-Terra and MODIS-Aqua data in QTP. The study

also developed a method of mapping fractional snow cover from MODIS in QTP. The new

snow cover products will provide daily snow cover at 500-m resolution in QTP. The new

snow cover algorithm employs the CIVCO topographic correction, a grouped-criteria

technique using the Normalized Difference Snow Index (NDSI) and other spectral threshold

tests and image fusion technology to identify and classify snow on a pixel-by-pixel basis.

The usefulness of the NDSI is based on the fact that snow and ice are considerably more

reflective in the visible than in the shortwave IR part of the spectrum, and the reflectance of


2

most clouds remains high in the short-wave IR, while the reflectance of snow is low. In

order to reduce the effect on cloud, snow cover over MODIS-Terra and MODIS-Aqua is

composed as maximum snow extent. At last, a MODIS-Terra fractional snow cover data

were added to the product base on linear relationship between NDSI and fractional snow

cover.

Validation of the MODIS snow cover and fractional snow cover products is an on-going

process. Two types of validation are addressed in the study-absolute and relative. To derive

absolute validation, the MODIS maps are compared with field measurements. Relative

validation refers to comparisons with other high resolution image snow cover maps, which

are considered to be the ‘truth’ snow maps. We have validated the daily snow cover product

MOD10A1 and 8-day snow cover product MOD10A2 using snow depth from 47 climate

stations in North Xinjiang, China. The accuracy of MODIS snow cover mapping algorithm

under varied topography, snow depth and land cover types was analyzed. Analysis showed

that the MODIS snow cover underestimated the snow cover area in alpine regions.

Vegetation cover has an important influence in the accuracy of MODIS snow cover maps.

We also validated the MOD10A1 by Landsat-ETM+ images with 30-m resolution in QTP.

Results suggest that the snow mapping algorithm of MODIS also underestimates the snow

cover. We intend to design a field experiment focused on validating our snow cover

products in QTP this winter. Recent advances in the area of snow remote sensing have lead

to further algorithm development to more accurately measure snow cover from different

sensors. In future, a blended snow product to map snow cover area utilizing MODIS,

AMSR-E passive microwave data, QuikSCAT scatterometer data and ICESTA laser radar

data will be developed.

1. Introduction 1.1 Identification

Snow is an important, though highly variable, earth surface cover (Klein et al., 1998).

Because of its high albedo, snow is an important factor in determining the radiation balance,

with implications for global climate studies (Foster and Chang, 1993). Midlatitude alpine

snow cover and its subsequent melt can dominate local to regional climate and hydrology,

and more and more notice in the world’s mountains regions snow cover. Because of its

importance, accurate monitoring of snow cover extent is an important research goal in the

science of Earth systems. Satellites are well suited to measurement of snow cover because

the high albedo of snow presents a good contrast with most other natural surfaces except


3

cloud. Fortunately, the physical properties of snow make it highly amenable to monitoring

via remote sensing. The objective of the MODIS snow mapping is to generate snow cover

area and fractional snow cover products on Qinghai-Tibet Plateau.

1.2 Overview

Remote sensing of snow cover is more than 40 years old. Snow was observed in the

first image obtained from the TIROS-1 weather satellite following its April 1960 launch

(Singer and Popham, 1963). However, it was in the mid-1960s that snow was successfully

mapped from space on a weekly basis following the launch of the ESSA-3 satellite. ESSA-3

carried the Advanced Vidicon Camera System (AVCS) that operated in the spectral range of

0.5 - 0.75 mm with a spatial resolution at nadir of 3.7 km. Using a variety of sensors,

including the Scanning Radiometer (SR), Very High Resolution Radiometer (VHRR) and

AVHRR sensors, snow cover has been mapped in the Northern Hemisphere on a weekly

basis since 1966 by NOAA (Matson et al., 1986; Matson, 1991). Initially, the weekly

NOAA National Environmental Satellite Data and Information System (NESDIS)

operational product was determined from visible satellite imagery from polar-orbiting and

geostationary satellites and surface observations. Where cloud cover precluded the analyst’s

view of the surface for an entire week, the analysis from the previous week was carried

forward (Ramsay, 1998). The maps were hand drawn, and then digitized using an 89×89

line grid overlaid on a stereographic map of the Northern Hemisphere. In 1997, the older,

weekly maps were replaced in 1997, by the IMS product. The IMS product provides a daily

snow map that is constructed through the use of a combination of techniques including

visible, near-infrared and passive-microwave imagery and meteorological-station data at a

spatial resolution of about 25 km (Ramsay, 1998 and 2000). Regional snow products, with

1-km resolution, are produced operationally in 3000 - 4000 drainage basins in North

America by the National Weather Service using NOAA National Operational Hydrologic

Remote Sensing Center (NOHRSC) data (Carroll, 1990 and Rango, 1993). Passive-

microwave sensors on-board the Nimbus 5, 6, and 7 satellites and the Defense

Meteorological Satellite Program (DMSP) have been used successfully for measuring snow

extent at a 25 to 30 km resolution through cloud-cover and darkness since 1978 (Chang et

al., 1987). Passive-microwave sensors also provide information on global snow depth

(Foster et al., 1984). The NOAA/AVHRR and the DMSP Special Sensor Microwave Imager

(SSM/I) are currently in operation. The Landsat Multispectral Scanner (MSS) and TM

sensors, with 80-m and 30-m resolution, respectively, are useful for measurement of snow


4

covered area over drainage basins (Rango and Martinec, 1982). Additionally, Landsat TM

data are useful for the quantitative measurement of snow reflectance (Dozier et al., 1981;

Dozier, 1984 and 1989; Hall et al., 1995; Winther, 1992).

The Moderate Resolution Imaging Spectroradiometer (MODIS), a major NASA EOS

instrument, was launched aboard the Terra satellite on December 18, 1999 (10:30 AM

equator crossing time, descending) for global monitoring of the atmosphere, terrestrial

ecosystems, and oceans. On May 4, 2002, a similar instrument was launched on the EOS-

Aqua satellite (1:30 PM equator crossing time, descending) (Salomonson et al., 2001).

MODIS data are now being used to produce snow-cover products from automated

algorithms at Goddard Space Flight Center in Greenbelt, MD. The products are transferred

to the National Snow and Ice Data Center (NSIDC) in Boulder, CO, where they are archived

and distributed via the Warehouse Inventory Search Tool (WIST). The MODIS snow

products are produced as a series of six products, including MOD10_L2, MOD10L2G,

MOD10A1, MOD10A2, MOD10C1 and MOD10C2. MOD10_L2 is swath product that is

generated using the MODIS calibrated radiance data products (MOD02HKM and

MOD021KM), the geolocation product (MOD03), and the cloud mask product (MOD35_L2)

as inputs. The MODL2G product is the result of mapping all the MOD10_L2 swaths

acquired during a day to grid cells of the Sinusoidal map projection. The Earth is divided

into an array of 36 x 18, longitude by latitude, tiles, about 10°x10° in size in the Sinusoidal

projection. The daily snow product MOD10A1 is a tile of data gridded in the sinusoidal

projection. Tiles are approximately 1200 x 1200 km (10°x10°) in area. Snow data arrays are

produced by selecting the most favorable observation (pixel) from the multiple observations

mapped to a cell of the MOD10_L2G gridded product from the MOD10_L2 swath product.

In addition to the snow data arrays mapped in from the MOD10_L2G, snow albedo is

calculated. There are four SDSs (or data fields) of snow data; snow cover map, fractional

snow cover, snow albedo and QA in the data product file. The MOD10A2 is eight-day

composited of MOD10A1. The MOD10A2 is generated by merging all the MOD10A1

products (tiles) for an eight-day period. MOD10C1 and MOD10C2 snow product gives a

global view of snow cover at 0.05° resolution global climate modeling grid (CMG) by a

geographic projection. The detail of MODIS products can be found from MODIS Snow

Products User Guide (Riggs et al. 2003). MODIS snow-cover products represent potential

improvement to or enhancement of the currently available operational products mainly

because the MODIS products are global and 500-m resolution, and have the capability to

separate most snow and clouds. The MODIS snow-mapping algorithms are automated,


5

which means that a consistent data set may be generated for longterm climate studies that

require snow-cover information. MODIS Terra and MODIS Aqua versions of the snow

products are generated. Bias to Terra is because the snow detection algorithm is based on

use of near infrared data at 1.6 µm. A primary key to snow detection is the characteristic of

snow to have high visible reflectance and low reflectance in the near infrared, MODIS band

6. MODIS band 6 (1.6 µm) on Terra is fully functional however, MODIS band 6 on Aqua is

only about 30% functional; 70% of the band 6 detectors non-functional. That situation on

Aqua caused a switch to band 7 (2.1 µm) for snow mapping in the swath level algorithm. In

addition, a fractional snow cover data array has been added to the product from collection 5.

In our study, mapping snow cover in mountainous regions remains an omission

limitation to the MODIS snow products from NSIDC (Hao Xiaohua et al. 2008). The

MODIS snow cover products rely on analysts to fine-tune the maps. So we describe and

validate a method that retrieves snow-covered area in Xinjiang and Qinghai-Tibet Plateau

regions, China by Terra MOD09 surface reflectance data. Develop an improved algorithm

suited for mapping MODIS snow cover and fraction snow cover on Qinghai-Tibet Plateau.

2. Algorithm Description of snow cover 2.1 Introduction The new snow cover algorithm employs the CIVCO topographic correction, a grouped-

criteria technique using the Normalized Difference Snow Index (NDSI) and other spectral

threshold tests and image fusion technology to identify and classify snow on a pixel-by-

pixel basis. The new algorithm was selected for the following reasons:

(1) The new snow cover algorithm is more accurate than algorithm of NSIDC on Qinghai-

Tibet Plateau.

(2) It corrects the effect of atmospheric and topographic conditions.

(3) It can minimize the limitation of the cloud.

(4) It runs automatically and fast. It is straightforward, computationally frugal, and thus

easy for the user to understand exactly how the product is generated.

Snow has strong visible reflectance and strong short-wave IR absorbing characteristics.

The Normalized Difference Snow Index (NDSI) is an effective way to distinguish snow

from many other surface features. Both sunlit and some shadowed snow is mapped

effectively. A similar index for vegetation, the Normalized Difference Vegetation Index

(NDVI) has been proven to be effective for monitoring global vegetation conditions

throughout the year (Tucker, 1979 and 1986). Additionally, some snow/cloud discrimination


6

is accomplished using the NDSI. Other promising techniques, such as traditional supervised

multispectral classifications, spectral-mixture modeling, or neural-network analyses have

not yet been shown to be usable for automatic application at the global scale. However,

these techniques may progress to regional applications.

2.2 Background and Data

2.2.1 Area of interest

The Qinghai-Tibet Plateau is the highest plateau over the world. It not only had an

important influence on the atmospheric circulation of the northern hemisphere, but also

directly affected the climatic and eco-environmental evolution of China in the Quaternary

period (Huairen and Xin, 1985).The Qinghai-Tibet Plateau is the largest, nonpolar cold

desert in the world, with an average elevation above 4000 m. The presence of snow cover

plays a key role in the cold desert ecosystem by affecting the hydrology, ecology and

climate. Snow cover in Qinghai-Tibet Plateau is highly variable both spatially and

temporally. Thin, discontinuous sheets of snow can occur year round (Zheng et al. 2000). In

the absence of snow, soils are more vulnerable to freezing and potentially decreased rates of

microbial transpiration, which can alter the soil’s ability to sequester carbon. Due to the

remoteness and topographic complexity of the Qinghai-Tibet Plateau, remote sensing offers

the most practical tool for monitoring its snow cover area.

2.2.2 Elevation data

The Digital Elevation Model (DEM) of the area at 500 m spatial resolution was created

from SRTM (Shuttle Radar Topography Mission) data at 3 arc-seconds, which is 1/1200th

of a degree of latitude and longitude, or about 90 meters as a source of topography

correction. From the DEM dataset, information about the slope, aspect and illumination

according to the sun angle and elevation were generated for input to the topographic

corrections algorithms for MODIS image.

2.2.3 MODIS data

In the new algorithm, we rely on MOD09 surface reflectance products (MOD09GA,

MYD09GHK) to get the MODIS snow cover. MOD09 (MODIS Surface Reflectance) is a

seven-band product computed from the MODIS Level 1B land bands 1 (620-670 nm), 2

(841-876 nm), 3 (459-479), 4 (545-565 nm), 5 (1230-1250 nm), 6 (1628-1652 nm), and 7

(2105-2155 nm). MOD is the MODIS/Terra data and MYD is the MODIS/Aqua data. The

product is an estimate of the surface spectral reflectance for each band as it would have been


7

measured at ground level as if there were no atmospheric scattering or absorption. It corrects

for the effects of atmospheric gases, aerosols, and thin cirrus clouds. The data can be

obtained from the National Snow and Ice Data Center Distributed Data Archive. Six

MOD09 tiles (h23v05, h24v05, h25v05, h26v05, h2506, h26v06) were used in the study

region.

Other MODIS product suite that include cloud mask data (MOD35 and MYD35) and

temperature data (MOD11A1 and MYD11A1) were regard as auxiliary inputs. The MODIS

daily snow cover product (MOD10A1 and MYD10A1) is regard as the reference data of the

snow cover from the new algorithms.

2.2.4 Landsat-ETM+ data and analysis

The ETM+ was launched on April 15, 1999, on the Landsat-7 satellite

(http://www.landsat.gsfc.nasa.gov/project/satellite.html). The ETM+ has eight discrete

bands ranging from 0.45 to 12.5 Am, and the spatial resolution ranges from 15 m in the

panchromatic band, to 60 m in the thermal-infrared band. All of the other bands have 30-m

resolution. Landsat-ETM+ data provide a high-resolution view of snow cover that can be

compared with the MODIS and operational snow-cover products. In the study, Landsat-

ETM+ path 143 row 30, path 136 row 38, path134 row 38, path 136 row 39, path134 row 40

path were used to produce a validation dataset for the MODIS snow cover products. The

figure1 shows the detail of study region.


8

Figure 1. The study region and the Landsat-ETM+ location. A, B ,C, D and E are respectively path 143 row 30, path 136

row 38, path134 row 38, path 136 row 39, path134 row 40.

2.3 Calculation of ground reflectance

The objective of any radiometric correction of airborne and spaceborne imagery of

optical sensors is the extraction of physical earth surface parameters such as reflectance,

emissivity, and temperature. The imagery available in the MOD09 (MODIS surface

reflectance product) provides measurements of surface reflectance with the atmosphere

correction by ‘6S’ model. However, in rugged terrain and in the case of multi-temporal

dataset these measurements are affected strongly by changes of topographic conditions. Our

research indicates that such variability reduces the identification of snow in shadow. To

getting the true ground reflectance the topography correction of the MOD09 is necessary in

QTP.

The problem of differential terrain illumination on satellite imagery has been

investigated for at least 20 years. At present, there are many methods in terrain correction,

such as physical models, Semi-empirical and empirical models. Although physical models can

be quite successful to eliminate atmospheric and topographic effects they inherently rely on an


9

accurate spectral and radiometric sensor calibration and on the accuracy and appropriate spatial

resolution of a digital elevation model (DEM) in rugged terrain and the computer is complex. The

MODIS data is large quantity. The empirical based approach offers the fast and accurate

correction. Law (2004) tested and compared three topographic correction methods, which

are the Cosine Correction, Minnaert Correction and a CIVCO model. By comparing, he

offered an improved CIVCO model. In our study, we used the improved CIVCO model.

The CIVCO method used here is modified from the two stage normalization proposed

by Civco, 1989, and consists of two stages. In the first stage, shaded relief models,

corresponding to the solar illumination conditions at the time of the satellite image are

computed using the DEM data. This requires the input of the solar azimuth and altitude

provided by the metadata of the satellite image. The resulting shaded relief model would

have values between 0 and 1. After the model is created, a transformation of each of the

original bands of the satellite image is performed to derive topographically normalized

images using equation (1) and (2).

(1)

( 2)

where !Ref"ij= the normalized radiance data for pixel(i, j) in band(!)

Ref"ij= the raw radiance data for pixel(i, j) in band(!)

µk= the mean value for the entire scaled shaded relief model (0,1)

µij= the scaled (0,1) illumination value for pixel(i, j)

C" = the correction coefficient for band(!)

N! = the mean on the slope facing away the sun in the uncalibrated data for the forest

category

S! = the mean on the slope facing to the sun in the uncalibrated data for the forest category

µk = the mean value for the entire scaled shaded relief model

µN = the mean of the illumination of forest on the slope facing away from the sun.


10

µS = the mean of the illumination of forest on the slope facing to the sun.

By the topography correction, we can get the MODIS surface reflectance. It will improve

the accuracy of snow cover mapping in mountainous regions.

2.4 Adjust of NDSI

The MODIS snow cover products algorithm is essentially designed for the evaluation of

the threshold value of the NDSI (Normalize Difference Snow Index) threshold value. For

MODIS data the NDSI is calculated as:

Erreur ! Des objets ne peuvent pas être créés à partir des codes de champs

de mise en forme. (3)

The NDSI threshold of the MODIS snow cover products distributed by the NSIDC is 0.40.

The NDSI values of the MODIS scenes greater than or equal to 0.40 represent snow cover

pixels. In addition, since water may also have an NDSI 0.4, an additional test is necessary to

separate snow and water. Snow and water may be discriminated because the reflectance of

water is <11% in MODIS band 2. Hence, if the reflectance of MODIS band 4 >11%, and the

NDSI 0.40, the pixel is initially considered snow covered. However, validation of the

current NDSI threshold has being accomplished only by the measurements in the United

States and Europe. In China, therefore, there is not reliable NDSI threshold value for the

MODIS snow mapping and a credible threshold can be established.

In the study, the snow cover area of A, B and C were selected for this study. First, the

Landsat-ETM+ snow cover maps were produced by the method of the SNOMAP. Then, the

snow cover maps, produced obtained from the way mentioned above, were compared with

the ones derived by the manual photo interpretation classification technique. Overall

agreement which is simply a comparison of the number of snow pixels, is high at 96%. Thus,

the Landsat-ETM+ snow cover maps can be reliable served as the “groudtruth”， with

which then the snow cover maps of the study area extracted from the MOD09

measurements by NDSI method were compared. For the MODSI snow cover maps of the

study areas, the NDSI threshold value for snow was increased gradually for 0.30 to 0.40 in

steps of 0.01. At Last, the comparisons focused on comparing the MODIS snow cover maps

following with NDSI threshold value and the Landsat-ETM+ snow cover maps serving as

absolute standard. The result suggests that the MODIS snow cover products distributed by

the NSIDC using NDSI threshold of 0.40 underestimated the SCA (snow-covered area) of


11

the study areas. In the study areas, the credible NDSI threshold value is respectively 0.34,

0.36and0.38 in A, B and C regions. As computer the average value, it is approximately

0.36,which is less than the one from the 0.40 of NSIDC.

Table 1. MODIS snow cover accuracy of different NDSI threshold in A, B and C region.

NDSI

threshold

value

The overall accuracy, Kappa

coefficient and fractional snow

cover area of A region.



cover area of B region.



cover area of C region.

0.39 93.00%、0.669、11.37% 86.82%、0.676、27.73% 94.73%、0.708、10.17%

0.38 93.02%、 0.672、11.53% 86.81%、0.678、28.36% 94.74%、0.711、10.48%

0.37 93.07%、0.675、11..66% 86.76%、0.679、29.02% 94.62%、0.709、10.79%

0.36 93.11%、0.679、11.83% 86.73%、0.680、29.63% 94.51%、0.707、11.08%

0.35 93.16%、0.683、11.97% 86.63%、0.679、30.25% 94.39%、0.706、11.48%

0.34 93.17%、0.685、12.13% 86.54%、0.679、30.87% 94.26%、0.703、11.82%

0.33 92.89%、0.678、12.66% 86.45%、0.679、31.51% 94.16%、0.702、12.16%

0.32 92.91%、0.681、12.80% 86.28%、0.677、32.13% 94.04%、0.700、12.53%

0.31 92.91%、0.683、12.98% 86.13%、0.676、32.66% 93.88%、0.697、12.89%

0.30 92.90%、0.684、13.18% 86.05%、0.676、33.23% 93.69%、0.692、13.28%

2.5 Additional Algorithms

In forested locations, many snow covered pixels have an NDSI lower than 0.4. To

correctly classify these forests as snow covered, a lower NDSI threshold is employed. The

normalized difference vegetation index (NDVI) and the NDSI are used together in order to

discriminate between snow-free and snow covered forests. The NDSI-NDVI field is

designed to capture as much of the variation in NDSI-NDVI values observed in the snow

covered forests as possible while minimizing inclusion of non-forested pixels. It was

designed to include forestcovered pixels that have NDSI values lower than 0.4, yet have

NDVI values lower than would be expected for snow-free conditions (Klein et al., 1998).

For MODIS data the NDVI is calculated as:

Erreur ! Des objets ne peuvent pas être créés à partir des codes de champs de

mise en forme. ( 4)

Last, a threshold of 10% in MODIS band 4 was used to prevent pixels with very low visible

reflectances, for example black spruce stands, from being classified as snow as has

previously been suggested (Dozier, 1989).

The NDSI can separate snow from most obscuring clouds, it does not always identify or


12

discriminate optically-thin cirrus clouds from snow. Clouds are masked by using the

MODIS cloud masking data product (MOD35).

One of the problems facing the MODIS snow-mapping algorithm is the mapping of

snow in regions where it is known not to exist. One of the more common locations for this

problem is in dark, dense forests, particularly in the tropics. The nature of the snow-

mapping algorithm is such that it is particularly sensitive to small changes in the NDSI or

NDVI over dark, dense vegetation. To correct false-snow mappings in tropical forests, the

MODIS temperature mask product (MOD11) was used to improve the accuracy of snow

cover map. A tentative threshold of 277 K has been set. When this threshold is applied in

tropical regions, e.g., the Congo, it eliminates from 93% to 98% of the false snow (Barton,

et al. 2001).

2.6 Image fusion

MODIS cloud masking data product was used to map MODIS snow cover product.

Nevertheless, inaccurate detection of clouds in the MOD35 cloud mask product revealed to

be problematic in high-elevation regions such as the QTP, China (Hall et al. 2002). The

Collection 5 of the MODIS snow products has been infused and expanded with information

regarding characteristics and quality of snow products at each level. It improves the cloud

mask product, thus permitting more snow covet to be mapped. However, the accurate

monitoring of SCA using optical imagery of high spatial resolution is seriously reduced by

cloud cover due to the similar reflective nature of snow and clouds. The ground object under

cloud remains unknown. Whether in MODIS terra or MODIS aqua daily snow cover

product, either way, it's always was effected by large cloud.

In the context of remote sensing, image fusion consists of merging images from

different sources, which hold information of a different nature, to create a synthesized image

that retains the most desirable characteristics of each source (Pohl & Genderen, 1998). In

my study, the method was applied to composite the MODIS/Terra and MODIS/Aqua snow

cover product to minimize the effect of cloud. In selecting the image fusion technique for

the daily composites, we decided that it would be most useful to use maximum snow cover.

In other words, if snow were present on any image in any location on the Terra or Aqua. tile

product, it will show up as snow-covered on the daily composite product. Maximum snow

cover is a more useful parameter than minimum or average snow cover. Using either

minimum or average snow cover would result in failure to map some snow cover. The


13

compositing technique also minimizes cloud cover. The figure 2 shows the flow process of our

new MODIS snow cover map algorithm.

Figure 2. The flow process chart of the new snow cover algorithms.

2.7 Backup Algorithm Future enhancements to MODIS snow cover maps include improving snow cover

resolution, fusing the polygenetic remote sensing data and producing more abundant applied

snow products.

3. Algorithm Description of fractional snow cover The work are doing. 4. Validation Plan 4.1 Introduction

The accuracy of snow cover products from optical remote sensing is of particular

importance in hydrological applications and climate models. In the study, using in situ

observation data during the five snow seasons at 47 climatic stations from January 1 to

MOD09GA MYD09GA

CIVCO Terrain correction

NDSI≥0.36, B2＞0.11

Snow，Cloud, Other

Klein MODEL，b4>0.1

LST mask:MOD11A1 Threshold value≤283

Cloud mask: MOD35， Land/water mask: MOD03

other

Snow in forest，Cloud，Other

MYDSNOW MODSNOW

Snow Cover Map

Maximum Composition


14

March 31of year 2001 and from November 1 to March 31 of year 2001 to 2005 in northern

Xinjiang area, China, the accuracy of MODIS snow cover products (MOD10A1 and

MOD102) and VEGETATION snow cover products (VGT-S10 snow cover products)

algorithm under varied terrain and land cover types were analyzed. The study shows the

overall accuracy of MOD10A1、MOD10A2 and VGT-S10 snow cover products is high at

91.3%, 90.6%, 87.9% respectively in all climatic stations. However, the overall accuracy of

the snow cover products in mountain regions is low. In mountain climatic stations the snow

omission of the three products is 32.4、21.7%、36.3% respectively. The cloud limitation

ratio of MOD10A1 reaches to 61.8%.;but the MOD10A2 and VGT-S10 are only 7.6%,

1.8%. The comparison result of user-defined 10-day MODIS snow products and VGT-S10

snow products shows that the snow identification ability of MODIS are more accuracy than

VGT-S10 snow cover products. However, the VGT-S10 snow cover products are little

affected by cloud than MODIS snow cover products.

4.2 Approach Two types of validation are addressed in our study-absolute and relative. To derive

absolute validation, the MODIS maps are compared with ground measurements or

measurements of snow cover from Landsat data, which are considered to be the ‘truth’ for

this work. Relative validation refers to comparisons with other snow maps, most of which

have unknown accuracy. Thus for the studies of relative validation, it is not generally

known which snow map has a higher accuracy.

4.3 Validation Sites QTP-Naqu. Lake Namtso.

4.4 Auxiliary Measurements Snow density, snow water liquid, snow grain size, snow temperature and snow pit works.

4.5 Scaling


15

5. Ancillary Data

DEM data , snow depth from climate stations.

6. Programming and Procedural Considerations 6.1 Programming Issues

The difficulty in establishing the accuracy of any of these maps is that it is not known

which map is the ‘‘truth’’ (if any) and the techniques used to map snow cover in the various

maps are different, resulting in different products.

6.2 Processing Issues 6.3 Quality Assurance References Barton, J.S., D.K. Hall and G.A. Riggs, unpublished document, 2001: Thermal and geometric thresholds in the

mapping of snow with MODIS, July 11, 2001.

Carroll T R. Operational airborne and satellite snow cover products of the National Operational Hydrologic

Remote Sensing Center[C]. Proceedings of the forty-seventh annual Eastern Snow Conference, Bangor,

Maine, CRREL Special Report. June 7-8, 1990: 90-44.

Chang, A.T.C., J.L. Foster and D.K. Hall. Microwave snow signatures (1.5 mm to 3 cm) over Alaska, Cold

Regions Science and Technology. 1987, 13:153-160.

Civco D L. Topographic Normalization of Landsat Thematic Mapper Digital Imagery[J]. Photogrammetric

Engineering and Remote Sensing. 1989, 55(9): 1303-1309.

Dozier J, Schneider S R, McGinnis J D F. Effect of grain size and snowpack water equivalence on visible and

near-infrared satellite observations of snow[J]. Water Resources Research.1981,17(4): 1213-1221.

Dozier, J. Snow reflectance from Landsat-4 thematic mapper. I.E.E.E. Transactions on Geoscience and

Remote Sensing, 1984,22: 323-328.

Dozier, J. Spectral signature of alpine snow cover from the Landsat Thematic Mapper, Remote Sensing of

Environment. 1989, 28: 9-22.


16

Foster, J.L., D.K. Hall, A.T.C. Chang and A. Rango. An overview of passive microwave snow research and

results. Reviews of Geophysics. 1984, 22: 195-208.

Foster, J.L., A.T.C. Chang. Snow cover. In Atlas of Satellite Observations Related to Global Change R.J.

Gurney, C.L. Parkinson, and J.L. Foster (eds.), Cambridge University Press, Cambridge. 1993: 361-370.

Hao Xiaohua, Wang Jian, Li Hongyi. Evaluation of the NDSI threshold value in mapping snow cover of

MODIS—A case study of snow in the middle Qilian Mountains. Journal of Glaciology and Geogryology.

2008,30 (1): 132-138.

Hall D K, Riggs G A, Salomonson V V. Development of methods for mapping global snow cover using

moderate resolution imaging spectroradiometer data. Remote Sensing of Environment. 1995, 54: 127–140.

Hall D K, Riggs G A, Salomonson V V, et al. MODIS snow-cover products[J]. Remote Sensing of

Environment. 2002, 83: 181-194.

Law K H, Nichol J. Topographic correction for differential illumination effects on IKONOS satellite

imagery[C]. ISPRS Congress, Istanbul, Turkey Commission 3. 12-23 July 2004.

Huairen Y. Climatic change in Quaternary. In: Tingdong L. Contribution to the Quaternary glaciology and

Quaternary geology, Geological Publishing House, P.R. China,1985,2:135–144.

Klein A, Hall D K, Riggs G A. Global snow cover monitoring using MODIS. In 27th International

Symposium on Remote Sensing of Environment. June 8-12, 1998: 363-366.

Matson, M., C.F. Ropelewski and M.S. Varnadore. An atlas of satellitederived northern hemisphere snow

cover frequency, National Weather Service, Washington, D.C. 1986, 75 pp.

Matson, M.. NOAA satellite snow cover data, Palaeogeography and Palaeoecology. 1991, 90: 213-218.

Pohl, C., & Genderen, J. L. V. (1998). Multisensor image fusion in remote sensing: Concepts, methods and

applications. International Journal of Remote Sensing, 19(5), 823#854.

Ramsay, B. The interactive multisensor snow and ice mapping system. Hydrological Processes. 1998,

12:1537-1546.

Ramsay B. Prospects for the interactive multisensor snow and Ice Mapping System (IMS) [C]. Proceedings of

the 57th Eastern Snow Conference, Syracuse, NY, East Snow Conference. 2000: 161-170.

Rango, A. Snow hydrology processes and remote sensing. Hydrological Processes. 1993, 7:121-138.

Rango, A. and J. Martinec. Snow accumulation derived from modified depletion curves of snow coverage,

Symposium on Hydrological Aspects of Alpine and High Mountain Areas, IAHS Publication.

1982,138:83-90.

Salomonson V V, Guenther B, Masuoka, E A. A summary of the status of the EOS Terra Misson MODIS and

attendant data product development after one year of on-orbit performance. In: Proceedings of the


17

International Geoscience and Remote Sensing Symposium/IGARSS’2001, Sydney, Australia, 9-13 July, 2001.

Singer, F.S. and R.W. Popham. Non-meteorological observations from weather satellites, Astronautics and

Aerospace Engineering. 1963, 1(3): 89-92.

Tucker, C.J. Red and photographic infrared linear combinations for monitoring vegetation, Remote Sensing of

Environment. 1979, 8: 127-150.

Tucker, C.J. Maximum normalized difference vegetation index images for sub-Saharan Africa for 1983-1985,

International Journal of Remote Sensing, 1986,7: 1383-1384.

Winther, J.G. Landsat thematic mapper (TM) derived reflectance from a mountainous

watershed during the snow melt season, Nordic Hydrology. 1992, 23: 273-290.

PART II

Snow Water Equivalent Retrieval Algorithm

Authors: Tao Che Affiliations: Cold and Arid Regions Environment and Engineering

Research Institute, Chinese Academy of Sciences.


19

Snow Water Equivalent Retrieval Algorithm

Abstract

We report spatial and temporal distribution of seasonal snow depth derived from

passive microwave satellite remote sensing data (e.g. SMMR from 1978 to 1987 and SMM/I

from 1987-2006) in China. We first modified the Chang algorithm and then validated it

using meteorological observations data, considering the influences from vegetation, wet

snow, precipitation, cold desert and frozen ground. Furthermore, the modified algorithm is

dynamically adjusted based on the seasonal variation of grain size and snow density. The

snow depth distribution is indirectly validated by MODIS snow cover products by

comparing the snow extent area from this work. The final snow depth datasets from 1978 to

2006 show that the inter-annual snow depth variation is very significant. The spatial and

temporal distribution of snow depth is illustrated and discussed, including the steady snow

cover regions in China and snow mass trend in these regions. Though the area extent of

seasonal snow cover in the Northern Hemisphere indicates a weak decrease in a long time

period, there is no clear trend in change of snow cover area extent in China. However, snow

mass over the Qinghai-Tibet Plateau and Northwestern China has increased, while it has

weakly decreased in Northeastern China. Overall, snow depth in China during the past three

decades shows significant inter-annual variations with a weak increasing trend.

1. Introduction 1.1 Identification Snow plays an important role at the climatic system due to its high surface albedo and

heat insulation effect which influences energy exchange between the land surface and the

atmosphere. It also influences the hydrological processes though snow water storage and

release. To obtain the large scale and long time period snow depth datasets, the passive

microwave remote sensing data (e.g. SMMR and SSM/I) have shown their capability in the

past three decades (Armstrong and Brodzik, 2002). The deeper the snowpack, the more

snow crystals are available to scatter microwave energy away from the sensor. Hence,

microwave brightness temperatures are generally lower for deep snowpack while they are


20

higher for shallow snowpack (Chang and others, 1987). Based on this fact, both snow depth

and snow water equivalent retrieval algorithms were developed by using brightness

temperature difference between 18 and 37 GHz (spectral gradient, e.g. Chang and others,

1987). With the utility of the Chang algorithm in the global scale, it was shown that a single

algorithm cannot describe all kinds of snow conditions (Foster and others, 1997). Regional

algorithms to retrieve snow depth have been developed in the past decade for North

America and Eurasia snowpack (Foster and others, 1997; Tait, 1998; Kelly and others,

2003).

1.2 Overview

In fact, the global snow depth retrieval algorithms overestimate snow depth in China

according to the records of meteorological station observations (Chang and others, 1992).

Snow depth retrieved from passive microwave remote sensing data can be influenced by the

condition of snowpacks, such as snow crystal (England, 1975; Chang and others, 1976;

Foster and others, 1997), snow density (Wiesmann and Matzler, 1999; Foster and others,

2005), and vegetation (Foster and others, 1997). Tait (1998) reported the different

algorithms for different snow features. For this reason, it is necessary to develop an

algorithm favorable to snow depth study in China.

It is reported that snow grain size and density determine the coefficient of spectral

gradient for snow depth retrieval. For example, using the Chang algorithm with a grain size

of 0.3 mm, the coefficient is 1.59, and with a grain size of 0.40 mm, the coefficient becomes

0.78 (Foster and others, 1997). Josberger and Mognard (2002) reported that while the

snowpack was constant, the spectral gradient continued to increase with time due to the

metamorphism of snow. Larger snow grains cause increased microwave scattering with the

result that an algorithm based on a fixed value for grain size will tend to overestimate snow

depth. (Armstrong and others, 1993). So, the spectral gradient will increase with the time

lapses due to the grouping snow grain size and snow density.

Liquid water content in snow layer (Ulaby and others, 1986; Matzler, 1994) and large

water bodies (Dong, 2005) can also lead to large errors in retrieving snow water equivalent.

These two factors should be considered before the linear regression for the coefficient

modification as in the Chang algorithm. Microwave radiation will not determine snow depth

accurately when snow is wet (Matzler, 1994). The dry snow and wet snow criteria were


21

used to discriminate the wet snow brightness temperature data, while the lake and land-sea

boundary were collected for removing the meteorological stations that near to the large

water body. After the work of Neale and others (1990), the NOAA-NASA SSM/I Pathfinder

(NNSP) program also uses SSM/I data to derive land surface classifications and to establish

criteria of dry snow and wet snow (Singh and Gan, 2000).

Grody (1991) reported it was necessary to remove the rain signal to identify snow cover.

When it is raining, snow parameters may not be retrieved. For obtaining the long-time series

dataset of snow depth, the Grody’s decision tree method based on the passive microwave

remote sensing data can be adopted so that the snow depth retrieval algorithm only is

focused on the snow pixels.

In this study, we will modify the Chang snow algorithm to make it suitable for snow

depth retrieval in China using SMMR and SSM/I remote sensing data and snow depth data

recorded at the China national meteorological stations. We will further analyze the accuracy

and uncertainty of the new snow product produced from the modified Chang algorithm. The

daily snow depth datasets in China from 1978/1979 to 2005/2006 will be produced, and

their spatial and temporal characteristics will be analyzed.

2. Algorithm Description 2.1 Introduction The coefficient of spectral gradient algorithm

Based on theoretical calculations and empirical studies, Chang and others (1987)

developed an algorithm for passive remote sensing of snow depth over relative uniform

snowfields utilizing the difference between the passive microwave brightness temperature

of 18 and 37 GHz in horizontal polarization.

SD = 1.5*(TB(18H) – TB(37H)) （1）

SD is snow depth in cm, and TB(18H) and TB(37H) are brightness temperature at 18

and 37 GHz in horizontal polarization, respectively. Here, brightness temperature at 37GHz

is sensitive to snow volume scattering, while that at 18GHz includes the information from


22

the ground under the snow. Therefore, the basic theory of the spectral gradient algorithm is

the snow volume scattering, which can be used to estimate the snow depth after the

coefficient (slope) was modified by the snow depth observations in the field.

Based on Foster and others’s results (1997) of forest influence, the forest area fraction

was considered here:

SD = a*(TB (18H) – TB (37H))/ (1-f) （2）

where a is the coefficient, while f is the forest area fraction.

In this study, snow depth observations at the meteorological stations in 1980 and 1981

were regressed with the spectral gradient of SMMR at 18 and 37GHz in horizontal

polarization. Before regression, the adverse factors should be taken into account, such as

liquid water content within the snowpack, which lead to a large uncertainty due to the big

difference between dry snow and water dielectric characteristics. The brightness

temperature data influenced by liquid water content were eliminated based on the following

dry snow criteria: TB(22V)-TB(19V) ≤ 4, TB(19V)-TB(19H)+TB(37V)-TB(37H)>8,

225<TB(37V)<257, and TB(19V)≤266 (Neale and others 1990). Mixed pixels with large

water bodies were removed according to the Chinese lake distribution map and the Chinese

coastline maps.

According to the regression between the spectral gradient of TB(18H) and TB(37H) and

the snow depth measured at the meteorological stations, the coefficient (slope) is 0.78 and

the standard deviations from the regression line is 6.22cm for SMMR data. For the SSM/I

brightness temperature data, the 19GHz channel replaced the 18GHz of SMMR. Results

show that the coefficient is 0.66 and the standard deviations from the regression line are

5.99cm. So, the modified algorithm is:

SD = 0.78*(TB(18H) – TB(37H))/(1-f) (for SMMR data from 1978 to 1987) SD = 0.66*(TB(19H) – TB(37H))/(1-f) (for SSM/I data from 1987 to 2006) (3) There are 2217 snow depth observations available in 1980 and 1981, while 6799

observations in 2003 due to the SSM/I has an improved swath width and acquiring period

than the SMMR has (See Figure 1 and 2).


23

Figure 1. Snow depth estimated from passive microwave brightness temperature data and observed in

meteorological stations: (a) SMMR in 1980 and 1981 and (b) SSM/I in 2003.

Figure 2 Percentage of error frequency distribution of snow depth estimated from passive microwave brightness temperature data and observed in meteorological stations. (a) SMMR in 1980 and 1981 and (b)

SSM/I in 2003. A simple dynamically adjusted algorithm

Snow density and grain size are two sensitive factors affecting microwave emission

from snowpacks (Foster and others, 1997, 2005), because it can partly affect the volume

scattering coefficient of snow. Although Josberger and Mognard (2002) developed a

dynamic snow depth algorithm, it is difficult to use the algorithm to mapping snow depth

estimation in China because the lack of reliable ground and air temperature data for each

passive microwave remote sensing pixel. In this study, we adopted a statistical regression

method to adjust the coefficient dynamically based on the error increasing ratio within the

snow season from October to April. The original Chang algorithm underestimated the snow

depth in the beginning of snow season and overestimated snow depth in the end of snow

season (Figure 4). As the results of statistic, the average offsets can be obtained in every

month for SMMR and SSM/I, respectively (Table 1). Table 1 Average offsets to remove the influence from snow density and grain size variations for each

month within the snow season based on the linear regression method


24

Average offset (cm) Month SMMR SSM/I Oct -3.64 -4.18 Nov -3.08 -3.58 Dec -1.91 -1.93 Jan -0.19 0.29 Feb 1.51 2.15 Mar 2.65 3.31 Apr 3.32 3.80

Figure 3 Error increases from snow density and grain size variations within the snow season from October to next April based on the estimations of SMMR and SSM/I data and observations in meteorological stations.

Here (a): SMMR and (b): SSM/I

2.2 Theoretical Basis of the Algorithm

To obtain the large scale and longtime period snow depth datasets, the passive

microwave remote sensing data (e.g. SMMR and SSM/I) have shown their capability in the

past three decades (Armstrong and Brodzik, 2002). The deeper the snowpacks, the more

snow crystals are available to scatter microwave energy away from the sensor. Hence,

microwave brightness temperatures are generally lower for deep snowpacks while they are

higher for shallow snowpacks (Chang and others, 1987). Based on this fact, both snow

depth and snow water equivalent retrieval algorithms were developed by using brightness

temperature difference between 18 and 37 GHz (spectral gradient, e.g. Chang and others,

1987). With the utility of the Chang algorithm in the global scale, it was shown that a single

algorithm cannot describe all kinds of snow conditions (Foster and others, 1997). Regional

algorithms to retrieve snow depth have been developed in the past decade for North

America and Eurasia snowpacks (Foster and others, 1997; Tait, 1998; Kelly and others,

2003).


25

2.3 Description of Retrieval Concept 2.4 Description of Retrieval Algorithm

The spectral gradient algorithm for the snow depth retrieval is based on the volume

scattering of snowpacks, which means other scattering surfaces can influence the results as

well. However, it will overestimate the snow cover area if the spectral gradient algorithm is

directly used to retrieve snow depth (Grody and Basist,1996). This is because that the snow

cover produces a positive difference between low and high-frequency channels, but the

precipitation, cold desert, and frozen ground show a similar scattering signature. Grody and

Basist (1996) developed a decision tree method for the identification of snow. The

classification method can distinguish the snow from other scattering signatures (i.e.

precipitation, cold desert, frozen ground).

Within the decision tree flowchart, there are four criteria related to the 85GHz channel.

For its utility of SMMR brightness temperature data which do not have the 85GHz channel,

we only adopted other relationships, such as the TB(19V)-TB(37V) as the scattering

signature rather than the TB(22V)-TB(85V). For the SMMR measures, the simplified

decision tree can be described as following relationships:

1. TB(19V)-TB(37V)>0, for scattering signature;

2. TB(22V)>258 or 258$TB(22V)%254 and TB(19V)-TB(37V)$2, for precipitation;

3. TB(19V)-TB(19H)%18 and TB(19V)-TB(37V)$10, for cold desert;

4. TB(19V)-TB(19H)≥8K and TB(19V)-TB(37V)≤2K and TB(37V)-TB(85V)≤6K,

for frozen ground.

For the more detail description of the decision tree method, please see Grody and Basist

(1996).

In this study, we adopted the Grody’s decision tree method to obtain snow cover from

SMMR (1978-1987) and SSM/I (1987-2004). Then, the snow depth data were calculated

only on those pixels by the snow depth retrieval algorithm. The return periods of SMMR

and SSM/I measurements are about every 3-5 days depending on the latitude. To obtain the


26

daily snow depth dataset, the intervals between swaths were filled up by the most recent

data available.

2.5 Backup Algorithm 3. Algorithm Prototyping 3.1 Data Analysis Passive microwave remote sensing data

The Scanning Multichannel Microwave Radiometer (SMMR) is an imaging 5-frequency

radiometer (6, 10, 18, 21, and 37 GHz) flown on the Nimbus-7 earth satellites launched in

1978. The SSM/I sensors on the DMSP satellite collect data for 4 frequencies: 19, 22, 37,

and 85 GHz. Both vertical and horizontal polarizations are measured for all except 22 GHz,

for which only the vertical polarization is measured. At NSIDC (National Snow and Ice

Data Center), the SMMR and SSM/I brightness temperatures are gridded to the NSIDC

Equal-Area Scalable Earth grids (EASE-Grids). Because China is located in a mid-latitude

region, we used the brightness temperature data with the global cylindrical equal-area

projection (Armstrong and others, 1994; Knowles and others, 2002).

Meteorological station snow depth observations

Snow depth observations at national meteorological stations from the China

Meteorological Administration (CMA) were used to modify and validate the coefficient of

the Chang algorithm. We used 178 stations within the main snow cover regions in China,

covering the Northeastern China, Northwestern China, and the QTP (Qinghai-Tibet Plateau)

(Figure 4). For modification of the Chang algorithm, we collected snow depth data from the

daily observations in 1980 and 1981 for SMMR, and 2003 for SSM/I, respectively. Then,

snow depth data in 1983 and 1984 (for SMMR) and 1993 (for SSM/I) were used to validate

the modified algorithm.

MODIS snow cover area products

Hall and others (2002) described the Moderate Resolution Imaging Spectroradiometer

(MODIS) snow cover area algorithm for the EOS Terra satellite. At present, the MODIS


27

snow products are created as a sequence of products beginning with a swath (scene) and

progressing, through spatial and temporal transformations, to an eight-day global gridded

product. In the NASA Goddard Space Flight Center (GSFC), the daily Climate Modeling

Grid (CMG) snow product gives a global view of snow cover at 0.05 degree resolution.

Snow cover extent is expressed as a percentage of snow observed in the raw MODIS cells at

500 m when mapped into a grid cell of the CMG at 0.05 degree resolution. These MODIS

snow cover products can be downloaded from NASA Earth Observing System Data

Gateway. In this study, we projected the 0.05 degree daily CMG product to register with the

EASE-Grids projection for the accuracy assessment of snow area extent derived from

passive microwave satellite data.

Vegetation distribution map in China

Snow depth retrieval from passive microwave remote sensing data will be influenced by

vegetation, in particular, the dense forest. Hu (2001) published the vegetation atlas of China

(1:1,000,000), which is the most detailed and accurate vegetation map of the whole country

up to now. It was based on the result of the nationwide vegetation surveys and their

associated researches in 50 years since 1949 and the relevant data from the aerial remote

sensing and satellite images, as well as geology, pedology and climatology. In this study, we

digitized and vectorized the vegetation atlas of China, and projected it into cylindrical

equal-area projection to register the EASE-GRID data. The forest area fraction will be used

to reduce the forest influence for the snow depth retrieval from passive microwave

brightness temperature data.


28

Figure 4. Position of meteorological stations within main snow cover regions in China (NWC:

Northwestern China, QTP: Qinghai-Tibet Plateau, NEC: Northeastern China, and other region).

Lake distribution map/Land-sea boundary

Based on the results of Dong and others (2005), large water bodies will seriously

influence the brightness temperature. Before the modification of snow depth retrieval

algorithm, those brightness temperature data and meteorological station data nearby the

lakes or ocean were removed to eliminate the mixed pixel effect. We used the 1:1,000,000

lake distribution maps from the Lake Database in China, which was produced by the

Nanjing Institute of Geography and Limnology, Chinese Academy of Sciences (CAS) and

was shared for scientific and educational group at Data-Sharing Network of Earth System

Science, CAS (http://www.geodata.cn). The Data-Sharing Network also archived the

1:4,000,000 coastline maps. These spatial data also was projected to register the EASE-

GRID data.

3.2 Prototyping of the Algorithm

We adopted the Grody’s decision tree method to obtain snow cover from SMMR (1978-

1987) and SSM/I (1987-2004). Then, the snow depth data were calculated only on those

pixels by the snow depth retrieval algorithm. The return periods of SMMR and SSM/I

measurements are about every 3-5 days depending on the latitude. To obtain the daily snow

depth dataset, the intervals between swaths were filled up by the most recent data available.

The flow chart to obtain the snow depth data in China can be described by Figure 5.


29

Figure 5 Flow chart of snow depth data in China derived from passive microwave brightness temperature data. 4. Validation Plan 4.1 Introduction

The validation used meteorological observations data, considering the influences from

vegetation, wet snow, precipitation, cold desert and frozen ground. The snow depth

distribution is indirectly validated by MODIS snow cover products by comparing the snow

extent area from this work.

4.2 Approach Accuracy assessment (Snow depth)

To assess the accuracy of snow depth retrieved from the modified algorithm, we used

measured snow depth data at the meteorological stations in 1983 and 1984 to compare with

the SMMR results, and that in 1993 for the SSM/I results. Both of the absolute errors less

than 5cm hold about 65% of all data (Figure 6). The standard deviations are 6.03cm and

5.61cm for SMMR and SSM/I, respectively.

Accuracy assessment (Snow cover)


30

We collected MODIS snow cover products from December 3, 2000 to February 28,

2001 to compare with the results of this study. Though MODIS snow cover products can not

provide snow depth information, we can compare the agreement or disagreement of MODIS

and SSM/I snow extent in each of SSM/I pixels by resampling the MODIS snow cover

products into the EASE-Grids projection. For a SSM/I pixel, when the snow depth is larger

than 2cm, we consider the pixel to be snow covered. For the resampled MODIS pixel, the

snow cover area is a fraction of snow covered, and when the snow cover area is larger than

50% we consider it as a snow cover pixel. Congalton (1991) described several accuracy

assessment methods of remotely sensed data. First of all, we considered the MODIS snow

cover products as the truth because the optical remote sensing has higher spatial resolution

and better comprehensive algorithm than the passive microwave remote sensing. Then, we

established the error matrixes of the SSM/I results for each day according to MODIS snow

cover products. Finally, two methods (overall accuracy and kappa analysis) were used to

assess the accuracy.

The two data sets have a good agreement by the overall accuracy analysis. The overall

accuracy is about from 0.8 to 0.9 after using Grody’s decision tree method (Grody and

Basist, 1996), while the accuracy from 0.7 to 0.8 without using the method (Figure 7(a)).

The results show that the overall accuracy can be improved by Grody’s decision tree

method by 10%.

The Kappa analysis is a more strict method to assess the coincidence in two data sets.

The Khat statistic was defined as (Congalton, 1991):

(4)

Where r is the number of rows in the error matrix, xii is the number of MODIS

observations in row i and column i, xi+ and x+i are the marginal totals of row i and column

i, respectively. N is the total number of data. The results of Khat statistics show that the

accuracy can be improved by Grody’s decision tree method by 20% (Figure 7(b)).


31

Figure 7 Accuracy assessment of overall accuracy and Kappa analysis methods based on the MODIS daily

snow cover area products from December 1, 2000 to February 28, 2001. Solid line is the results with Grody’s decision tree method to identify the snow cover, and Dash line is the results without the decision tree method.

(a) Overall accuracy, and (b) Kappa coefficient.

Uncertainty Effect of Vegetation

Vegetation cover has a significant influence on snow depth estimation from remote

sensing data (Foster and others, 1997, 2005). In this study, we used the forest cover

parameter to remove this influence (Foster and others, 1997). In fact, this method is not

appropriate out for dense forest regions. We overlap the stable snow cover map with the

Chinese Vegetation Map and find dense forests with a large forest cover fraction (greater

0.5) mainly distribute in the Xing’aling regions (Heilongjiang Province and the Eastern Inter

Mongolia) with about 160 EASE-Grid pixels (100,000km2). Although snow depth derived

from the modified algorithm may be questionable, the total area of the dense forest regions

is very limited.

Effect of Snow Crystal


32

The snow grain size can influence the algorithm coefficient of snow depth retrieval (e.g.

formula (1) and (2)). With a snow grain size of 0.3mm the coefficient is 1.59, but with a

snow grain size of 0.4mm the coefficient becomes 0.78 (Foster and others, 1997). Snow

crystal size can depend on the snowfall condition, such as the wind and temperature. It also

varies with snow metamorphism after the snow is on the ground. In this study, we

characterized this influence using a statistical regression method and adjusted the seasonal

offsets. These offsets can not interpret the regional differences of snow conditions.

Effect of Liquid Water Content

The snow depth can not be retrieved when snow is wet because the liquid water within

snow layer will remove the volume scatter of microwave signals. Therefore, only morning

brightness temperature data were used to minimize the errors associated with melting snow

in the afternoon.

4.3 Validation Sites

The specific validation sites still under-investigation which will be presented in later

vrsion 4.4 Auxiliary Measurements

Still under-investigation which will be presented in later version 4.5 Scaling

Still under-investigation which will be presented in later version 4.6 Data Protocols and Dissemination 4.7 Proposed Validation Tests

Still under-investigation which will be presented in later version

5. Ancillary Data


33

The ancillary data need in this algorithm is: meteorological station snow depth

observations, MODIS snow cover area products, vegetation distribution map in China and

lake distribution map/Land-sea boundary. Detailed information for each dataset can be fund

in Section 3.1 Data Analysis

6. Programming and Procedural Considerations

The whole part still under-investigation which will be presented in later version

6.1 Programming Issues 6.2 Processing Issues 6.3 Quality Assurance

References

1. Armstrong, R. L., A. T. C. Chang, A. Rango, and E. Josberger. 1993. Snow depths and

grain-size relationships with relevance for passive microwave studies, Ann. Glaciol.,

17, 171–176.

2. Armstrong, R. L., K. W. Knowles, M. J. Brodzik and M. A. Hardman. 1994, updated

current year. DMSP SSM/I Pathfinder daily EASE-Grid brightness temperatures, [list

dates of data used]. Boulder, Colorado USA: National Snow and Ice Data Center.

Digital media

3. Armstrong, R.L., and M.J. Brodzik. 2002. Hemispheric-scale comparison and evaluation of passive-microwave snow algorithms. Ann. Glaciol,. 34, 38-44.

4. Chang, A. T. C., P.Gloersen, T. Schmugge, T. T. Wilheit, and H. J.Zwally. 1976.

Microwave emission from snow and glacier ice. J. Glaciol., 16, 23-39.

5. Chang, A. T. C., J. L. Foster, and D. K. Hall. 1987. Nibus-7 SMMR derived global snow

cover parameters. Ann. Glaciol,. 9, 39-44.

6. Chang, A. T. C., D. A. Robinson, L. Peiji, and C. Meisheng. 1992. The use of

microwave radiometer data for characterizing snow storage in western China. Ann.


34

Glaciol., 16, 215-219.

7. Congalton, R. 1991. A review of assessing the accuracy of classification of remotely

sensed data. Remote Sens. Environ,.37, 35-46,.

8. Dong, J. R., J. P.Walker, and P. R. Houser. 2005. Factors affecting remotely sensed

snow water equivalent uncertainty. Remote Sens. Environ, 97, 68-82.

9. England, A.W. 1975. Thermal microwave emission from a scattering layer. J. Geophys.

Res., 80 (32), 4484-4496.

10. Foster, J. L., A. T. C. Chang, and D. K. Hall, 1997. Comparison snow mass estimates

from a prototype passive microwave snow algorithm, a revised algorithm and snow

depth climatology. Remote Sens. Environ. 62, 132-142, 1997.

11. Foster, J.L., C.J. Sun, J.P. Walker, R. Kelly, A.C.T. Chang, J.R. Dong, H. Powell. 2005.

Quantifying the uncertainty in passive microwave snow water equivalent observations.

Remote Sens. Environ. 94, 187-203.

12. Grody, N C. 1991. Classification of snow cover and precipitation using the Special

Sensor Microwave Imager. J. Geophys. Res., 96, 7423-7435.

13. Grody, N. C., and A. N. Basist. 1996. Global identification of snowcover using SSM/I

measurements. IEEE Trans. Geosci. Remote Sensing.34, 237-249.

14. Hall, D. K., G. A. Riggs, V. V. Salomonson, N. E. DiGirolamo, and K. J. Bayr. 2002.

MODIS snow-cover products. Remote Sens. Environ.83, 181-194.

15. Hu, X. Y. 2001. The Vegetation Atlas of China (1:1,000,000). Beijing: Science press.

16. Josberger, E. G., and Mognard, N. M. 2002. A passive microwave snow depth

algorithm with a proxy for snow metamorphism. Hydrological Processes, 16(8), 1557-

1568.

17. Kelly, R.E., A.C.T. Chang, and T. Leung T. 2003. A prototype AMSR-E global snow

area and snow depth algorithm. IEEE Trans. Geosci. Remote Sens., 41(2), 230-242.

18. Knowles, K., E. Njoku, R. Armstrong, and M.J. Brodzik. 2002. Nimbus-7 SMMR

Pathfinder daily EASE-Grid brightness temperatures. Boulder, CO: National Snow and


35

Ice Data Center. Digital media and CD-ROM.

19. Li, P. J. and D. S. Mi. 1983. Distribution of snow cover in China. Journal of glaciology

and cryopedology, 5(4), 9-18. (In Chinese)

20. Matzler, C. 1994. Passive microwave signatures of landscapes in winter. Meteorol.

Atmos. Phys. 54, 241–260.

21. Neale, C. M. U., M. L. McFarland, and K. Chang. 1990. Land-surface-type classification using microwave brightness temperatures from the special sensor microwave/imager. IEEE Trans. Geosci. Remote Sens. 28(5), 829-837.

22. Qin, D., S. Liu, and P. Li. 2006. Snow cover distribution, variability, and response to

climate change in Western China. J. Climate, 19(9), 1820-1833.

23. Rikiishi, K. and N. Nakasato. 2006. Height dependence of the tendency for reduction in

seasonal snow cover in the Himalaya and the Tibetan Plateau region, 1966-2001. Ann.

Glaciol., 43, 369-377.

24. Singh, P. R., and T. Y. Gan. 2000. Retrieval of snow water equivalent using passive

microwave brightness temperature data. Remote Sens. Environ, 74, 275-286.

25. Tait, A.B. 1998. Estimation of snow water equivalent using passive microwave

radiation data. Remote Sens. Environ.64, 286-291.

26. Ulaby, F., R.Moore, , and A. Fung. 1986. Microwave Remote Sensing, Artech House,

Dedham, MA, Vol. III, 1602-1634.

27. Wiesmann, A, and C. Matzler. 1999. Microwave emission model of layered snowpacks.

Remote Sens. Environ, 70, 307-316.


PART III

Surface Soil Freeze/Thaw State Dataset Using The Decision Tree

Classification Algorithm

Authors: Rui Jin

Affiliations: Cold and Arid Regions Environment and Engineering

Research Institute, Chinese Academy of Sciences.


37

Surface Soil Freeze/Thaw State Dataset Using The Decision Tree

Classification Algorithm Abstract

A new decision tree algorithm to classify the surface soil freeze/thaw states has been

developed. The algorithm uses SSM/I brightness temperatures recorded in the early morning.

Three critical indices are introduced as classification criteria—the scattering index (SI), the

37 GHz vertical polarization brightness temperature (T37V), and the 19 GHz polarization

difference (PD19). And the discrimination of the desert and precipitation from frozen soil is

considered, which improve the classification accuracy. Long time series of surface soil

freeze/thaw statuses can be obtained using this decision tree, which potentially can provide

a basic dataset for research on climate and cryosphere interactions, carbon cycles,

hydrological processes, and general circulation models.

1. Introduction

Globally, about 50&106 km2 of surface soil undergoes freeze/thaw cycles annually

(Kimball et al., 2001; Zhang et al., 2003a). The soil freeze/thaw status has a profound

influence on the energy and water exchange between the land surface and the atmosphere,

the hydrological cycle, crop growth, and the carbon cycle (Cao & Chang, 1997; Goodison et

al., 1998; Judge et al., 1997; Zhang & Armstrong, 2001; Zuerndorfer et al., 1990;

Zuerndorfer & England, 1992). The timing, duration, and area of surface soil freeze/thaw

status can be taken as an indicator of climate change because of its sensitivity (Goodison et

al., 1998; Li et al., 2008; Zhang & Armstrong, 2001; Zhang et al., 2003b).

A new decision tree algorithm was developed to classify the soil freeze/thaw state with

SSM/I data. New indices are introduced, and the discrimination of the desert and

precipitation from frozen soil is considered. Long time series of surface soil freeze/thaw

statuses can be obtained using this decision tree, which potentially can provide a basic

dataset for research on climate and cryosphere interactions, carbon cycles, hydrological


38

processes, and general circulation models (Allison et al., 2001; Jin & Li, 2002; Judge et al.,

1997; Zhang & Armstrong, 2001; Zuerndorfer et al., 1990).

1.1 Identification 1.2 Overview

Many studies were published during the 1980s and 1990s on detecting the surface soil

freeze/thaw state using passive microwave radiometers such as SMMR and SSM/I. There

are two major types of near-surface soil freeze/thaw states classification algorithm

comprising the dual-indexes algorithm (Zuerndorfer et al., 1990; Zuerndorfer et al., 1992;

Judge et al., 1997; Zhang and Armstrong, 2001; Zhang et al., 2003), and change detection

algorithm (Smith et al., 2004). All above algorithms were based on the unique microwave

radiative characteristics associated with frozen soils, such as lower thermo-dynamical

temperature, higher emissivity and volume scattering darkening effect. (1) Dual-indexes Algorithm The dual-indexes algorithm using T37 brightness temperature and the spectral gradient

(SG) between T37 and T18/T19 was most widely used in 1990s. The dual-index algorithm

was easily for the operational application with the unified thresholds throughout the

research region, however the thresholds of both indices were determined through a

statistical analysis of training samples, which need to be recalibrated when applied in other

regions (Jin and Li, 2002).

(2) Change Detection Algorithm The change detection algorithm for surface soil freeze/thaw states classification was

originated from the active microwave remote sensing based on the time series of the

backscattering coefficient. Smith developed an algorithm applicable to passive microwave

remote sensing (Smith et al., 2004) by using the difference between the brightness

temperature at 37 and 19 (or 18) GHz to identify the transition from frozen to thawed soil.

However, the gradual process of soil temperature with freezing, the coarse spatial resolution

of the passive microwave radiometers, and the opposite effect of increased emissivity and

decreased thermal temperature of frozen soil on the brightness temperature may resulted in

no abrupt changes in brightness temperature or spectral signals at the daily scale.


39

Furthermore, both of above algorithms only separate frozen and thawed soil. The

desert in the winter season and snow were both easily misclassified as frozen soil because of

their similar volumetric scattering characteristics (Fiore Jr & Grody, 1992; Cao & Chang,

1997). In addition, precipitation may mask the radiation emitted from the land surface

(Grody & Basist, 1996). Therefore, it is necessary to distinguish these types to improve the

classification accuracy of frozen/thawed soil. 2. Algorithm Description 2.1 Introduction

A new decision tree algorithm was developed to classify the soil freeze/thaw state with

SSM/I data. New indices, i.e. scattering index, polarization difference, are introduced, and

the discrimination of the desert and precipitation from frozen soil is considered, which will

improve the classification accuracy of the surface soil freeze/thaw states.

2.2 Targets to be observed

Due to the coarse spatial resolution of passive microwave remote sensing, “pure”

training samples from SSM/I data need to be collected to analyze the brightness temperature

characteristics of different land surface types and to determine the threshold of each node in

the decision tree. We selected four types of samples, including frozen soil, thawed soil,

desert and snow. The latter two sample types have volume scattering characteristics similar

to those of frozen soil. Grody’s method was adequately validated by previous research

(Grody & Basist, 1996), so it was adopted directly to identify precipitation.

2.3 Radiative Transfer Problem

The soil brightness temperature Tb can be simply expressed as the product of the soil

effective temperature Teff and the emissivity e if we consider the soil as a semi-infinite

medium (Ulaby et al., 1986). When the soil freezes, its thermodynamic temperature

decreases, but the emissivity increases due to the decreased permittivity. Therefore, the

change in radiobrightness may be either positive or negative, mainly depending on the soil

moisture (Zuerndorfer et al., 1990; Zuerndorfer & England, 1992). For dry soil, the soil

emissivity changes little between the thawed and frozen states, so the brightness temperature

generally decreases with soil temperature. For moist soil, the emissivity increases

significantly when it changes from the thawed to the frozen state, but the Teff may only


40

drop a few Kelvin, so the Tb may increase (Dobson et al., 1985; Jin & Li, 2002; Zuerndorfer

et al., 1990). According to the above analysis, although the brightness temperature of frozen

soil is low, the brightness temperature cannot be taken as an unambiguous index to identify

the soil freeze/thaw status (Zuerndorfer et al., 1990). Moreover, the brightness temperature

of moist regions near rivers and lakes is also low because of abundant moisture and the

corresponding lower emissivity, which may cause confusion in distinguishing between

frozen soil and very moist soil when using the brightness temperature alone (England, 1990).

Both the permittivity and the dielectric loss factor decrease with soil freezing (Hoekstra

et al., 1974). The dielectric loss factor is reduced more than the permittivity, resulting in a

decrease of the loss tangent ( ), which means that the emission depth will be

greater and there will be volume scattering. The effective emission depth Ze (1-e-1 of the

total emission in the zenith direction originates above Ze) is about 10% of the free space

wavelength in moist soil, and increases to more than 30% of the free space wavelength

when the soil is frozen (Zuerndorfer et al., 1990). The higher the microwave frequency the

more heterogeneous the soil column is, and the stronger the scattering volume will be (Cao

& Chang, 1997; England et al., 1991; Zuerndorfer et al., 1990). The brightness temperature

of frozen soil at high frequencies is therefore generally lower than that at low frequencies.

In summary, the microwave emissions and scattering characteristics have several

differences between frozen and thawed soil, such as a lower thermodynamic temperature

and brightness temperature, a higher emissivity, and a stronger volume scatter darkening

effect that can be used to select proper indices to identify the soil freeze/thaw state.

2.4 Mathematical Basis of the Algorithm

There are three critical indices used in the decision tree:

(1) Scattering Index (SI): The SI was proposed based on a regression analysis of the

training data covering various land surface types and atmospheric conditions (Grody, 1991),

expressed as follows:

, (1)


41

where, T19V, T22V and T85V are vertical polarization brightness temperatures at 19,

22 and 85 GHz, respectively. F represents the simulated 85 GHz vertical polarization

brightness temperature under the ideal condition of no scattering effect. SI is the deviation

of the actual SSM/I T85V observation from F. Because the volume scattering darkening of

frozen soil at 85 GHz is stronger than that at lower frequencies, SI is a more reliable index

than SG for distinguishing between scatterering and non-scatterering samples.

(2) 37 GHz vertical polarization brightness temperature (T37V): A correlation analysis

was carried out between the SSM/I brightness temperature at each channel and the SMTMS

4 cm deep soil temperature, revealing that T37V has the highest correlation coefficient of

0.87 with the 4 cm deep soil temperature. T37V was therefore used as a criterion to indicate

the thermal regime of the surface soil.

(3) 19 GHz Polarization Difference (PD19 = T19V - T19H). The polarization

difference at 19 GHz reveals the surface roughness. A rougher surface decreases the

coherent reflection and increases incoherent scattering, resulting in the tendency of the

surface reflectivity to be independent of polarization, diminishing the polarization difference.

The PD19 was used to identify the desert, which has a relatively small roughness.

2.5 Description of Retrieval Concept 2.6 Description of Retrieval Algorithm 2.7 Backup Algorithm 3. Algorithm Prototyping 3.1 Data Analysis 3.1.1 Analysis of the brightness temperature characteristics of each land surface type

The variation of the time series of the above three indices was analyzed for each

sample type, providing a priori knowledge necessary to create a decision tree.

Frozen/thawed soil

Figure 1 shows the time series of T37V, SI and PD19 at the Tuotuohe and MS3608

stations from June 29, 1997 to August 31, 1998. The SMTMS 4 cm deep soil temperatures


42

and soil moistures are also shown as ancillary information to indicate the surface soil

freeze/thaw status. Both stations are located in the seasonally frozen ground region. The soil

moisture of MS3608 was higher than that of Tuotuohe.

Although the hydrothermal conditions are different between the two stations, the three

indices have many characteristics in common when the soil is frozen or thawed. In the

middle of October, the 4 cm deep soil temperature fell below the soil freezing point; the

liquid water in the soil changed its phase to ice and suddenly dropped. The 37 GHz

brightness temperature therefore decreased, and the SI increased due to volume scattering

darkening. When the reverse phase change process occurred during middle to late April of

the next year, the 4 cm deep soil temperature increased; the 37 GHz brightness temperature

accordingly increased and the SI decreased due to dominant surface scattering. The frozen

soil scatters with an SI between 10 and 3 because the volume fraction of soil matrix and ice

particles in the frozen soil is very large, about 0.5 to 0.8, which results in the attenuation of

the volume scattering effect. The high value of SI at the MS3608 station in December 1997

resulted from the snow cover. The PD19 of frozen soil fluctuated modestly with soil

temperature and soil moisture, and was commonly smaller than 25.

(1) Desert

Two years (1999-2000) of SSM/I brightness temperatures and daily mean air

temperatures were acquired for the Tazhong station, located in the hinterland of the

Taklimakan desert and operated by the CMA (China Meteorological Administration). There

were no soil temperature observations at the Tazhong station. The polarization difference of

the desert at each SSM/I channel was larger than that of other land types because it is

smoother (Neale et al., 1990). Fig. 2 shows that the PD19 of the desert was above 30 for

most of the year, the SI was mainly between 5 and 10, and the brightness temperature

variation of the desert agreed well with the air temperature variation due to the very low

moisture content in the desert. Compared to dry snow and frozen ground, the desert is a

weaker scatterer due to the large volume fraction, and the homogeneous particle size and

dielectric properties. The effective emissivity of the desert at 37 GHz vertical polarization

was about 0.95 on average, calculated by dividing the 37 GHz vertical polarization

brightness temperature by the daily mean air temperature.


43

(a) Tuotuohe

(b) MS3608

Fig. 1 Time series of T37V, SI and PD19 of frozen/thawed soil at Tuotuohe (a) and MS3608 (b) station


44

Fig. 2 Time series of T37V, SI and PD19 of the desert at Tazhong station, Taklamakan Desert.

(2) Snow cover

The microwave radiative characteristics of snow cover are very similar to those of

frozen soil, including a low temperature, a low complex dielectric constant, and strong

volume scattering (Edgeton et al., 1971). The shallow and dry snow is transparent to

microwaves, so most of the brightness contribution comes from the underlying soil, which

may cause confusion in separating shallow snow and frozen soil. The snow depth for each

snow sample was calculated using Equation 2 (Che et al., 2008). The SI of shallow snow

samples (<10 cm) are generally between 0 and 20, close to the SI of frozen soil. An increase

in the snow depth enhances the volume scattering effect in snow. Therefore, the SI of snow

deeper than 10 cm is above 30, and even reaches 80 for deep snow.

(2)

Furthermore, the patchily-distributed shallow snow cover over China cannot

effectively play a role in the heat preservation and insulation of the underlying soil. The soil

under the snow cover remained frozen most of the time (Cao et al., 1997). The snow cover

was therefore not targeted as a classification type in this decision tree.

3.1.2 Cluster analysis and decision tree for freeze/thaw status classification

The spatial distribution of the randomly selected training samples shows that each type

converges as a cluster in the 3-dimensional space composed by the three indices (Fig. 3a).

The decision rules in the decision tree (Fig. 4) were determined from the mean and standard

deviation of each index calculated for each type. These rules are:

(1) The PD19 of desert is 36.28±2!2.22 (mean±2!standard deviation), obviously larger

than that of other land surface types. A threshold of PD19>30 was used to identify most

desert (Fig. 3b), and the remaining desert can be further separated in the sub-branches of the

decision tree by using PD19>25. (2) Both frozen soil and snow are strong scatterers with

high SI values. The threshold of SI"5.0 was used to separate more than 95% (18.69±2&6.04)

(Fig. 3c) of frozen soil samples into the left branch of the decision tree (Fig. 4). (3) In terms

of brightness temperature, the T37V of frozen soil is 232.57±2&9.40, while that of thawed

soil is 259.1±2&5.33. The threshold of T37V=252 K can separate frozen and thawed soil

samples with the least misclassification (Fig. 3a and d). (4) Because of the strong scattering

from ice particles, some of the precipitation pixels would be divided into the left branch of


45

the decision tree after using SI"5.0. However, the precipitation is still warmer than frozen

ground. Grody’s index T22V"165+0.49&T85V was therefore directly adopted to identify

deep convective precipitation with ice particles. Furthermore, the discriminant

T85V/T19V<0.9 was used to identify hail clouds and rainstorms (He & Chen, 2006). For

precipitation with weak scattering, the discrimination of 254K$T22V$258K and SI$2 were

used in the right branch of the decision tree (Grody & Basist, 1996). The decision tree to

classify soil surface freeze/thaw status was finally set up in Fig. 4.

3.2 Prototyping of the Algorithm 4. Validation Plan 4.1 Introduction

In order to evaluate the accuracy of the decision tree algorithm, the daily classification

results were first validated by SMTMS 4 cm deep soil temperature observations at the local

time of 6:00 am for eight stations on the Qinghai-Tibetan Plateau measured during CEOP-

EOP3. We also conducted a grid-to-grid validation by the Kappa statistics using the map of

geocryological regionalization and classification in China (Zhou et al., 2000) as a reference

(Fig. 5b).


46

Fig. 3 Cluster analysis on the samples of frozen soil, thawed soil, desert and snow (a) and the statistical

characteristics of PD19 (b), SI (c) and T37V (d) for different land surface types.

4.2 Approach

In order to evaluate the accuracy of the decision tree algorithm, the daily classification

results were first validated by SMTMS 4 cm deep soil temperature observations at the local

time of 6:00 am for eight stations on the Qinghai-Tibetan Plateau measured during CEOP-

EOP3. Only the classification of frozen or thawed soil was validated. The number of

validated pixels was 1695, and the number of misclassifications was 219. The average

classification accuracy reached 87% (Table 1).


47

Fig. 4 Flow chart of the decision tree for the surface soil freeze/thaw status classification.

Table 1. Validation of the classification results by 4 cm deep soil temperature observations at selected

CEOP stations. Station Validation data Misclassified data Accuracy (%)

AMDO 219 25 88.58 MS3608 207 24 88.41 MS3637 209 27 87.08

D66 217 15 93.09 D105 209 39 91.34 D110 211 41 80.57

BJ 207 19 90.82 Tuotuohe 216 29 86.57

Total 1695 219 87.08

As for the misclassification, among 219 pixels, 18 cases of thawed soil were

misclassified into the desert type due to the high PD19 value of the flat and dry surfaces.

This kind of misclassification can be avoided using a reliable desert map. The freeze or

thaw statuses of the remaining 201 pixels were misclassified. We first analyzed this kind of

misclassification from the viewpoint of soil temperature; it was found that 40% and 73% of

the misclassification occurred when the 4 cm deep soil temperature was in the range of -0.5

°C-0.5 °C and -2.0 °C-2.0 °C, respectively, according to the frequency histogram of

misclassification pixels numbered against 4 cm deep soil temperatures (Fig. 6a). Then we

determined that from the viewpoint of timing, most misclassifications occurred during the

transition period between the cold and warm seasons. For instance, the proportions of error

in April-May and September-October to the total number of misclassifications were about

33% and 38%, respectively (Fig. 6b). It is understandable that most of the misclassifications

were in the transition periods because the heterogeneity within pixels is more significant at

these times. Furthermore, the frozen soil is defined according to the temperature regime.


48

However, most of the water in the soil still remains in the liquid state when the soil

temperature is just below the soil freezing point, which shows similar dielectric properties

as the thawed soil and would result in misclassification between frozen and thawed soil.

Fig. 5 actual number of frozen days in China (a) and Map of geocryological regionalization and

classification in China (b) for the period from Oct. 1, 2002 to Sep. 31, 2003. We also conducted a grid-to-grid validation by the Kappa statistics using the map of

geocryological regionalization and classification in China (Zhou et al., 2000) as a reference

(Fig. 5b), a widely used method to measure the agreement between the reference data and

the classified result in grid format (Congalton, 1991). For comparability, we first obtained

the actual number of frozen days for one year—during the period from October 1, 2002 to

September 31, 2003—over China based on the pentad compositions by counting the frozen

days for each pixel (Fig. 5a). Then, the map of the frozen soil area was delineated by

assuming that the pixels that were frozen for more than 15 days should be seasonally frozen

soil or permafrost. The pixels that were frozen for less than 15 days represent short time

frozen soil (Zhou et al., 2000). The new frozen soil area map derived from the decision tree

classification result using the SSM/I data was compared with the reference map. The results

show that the overall classification accuracy was 91.66%, which was calculated from the

error matrix, and the Kappa index was 80.5%. The boundary between the frozen and thawed

soil in the new map (Fig. 5a) was consistent with the southern limit of seasonally frozen

ground in the reference map (Fig. 5b).


49

Fig. 6 Frequency histograms of the soil temperature and occurrence time for the misclassified pixels.

4.3 Validation Sites

Table 2. Stations used in algorithm development and validation (Wang et al., 2000, Zhou et al., 2000) Station Situation Altitude(m) Geocryological regionalization Landscape

AMDO 91.63ºE; 32.24ºN 4700 predominantly continuous permafrost subhumid alpine

MS3608 91.78ºE; 31.23ºN 4610 predominantly continuous and island permafrost subhumid alpine

MS3637 91.66ºE; 31.02ºN 4820 predominantly continuous and island permafrost subhumid alpine

D66 93.78ºE; 35.52ºN 4600 predominantly continuous permafrost semi-arid desert steppe

D105 91.94ºE; 33.07ºN 5020 predominantly continuous permafrost N/A

D110 91.88ºE; 32.69ºN 5070 predominantly continuous permafrost subhumid swamp

meadow

BJ 91.90ºE; 31.37ºN 4509 predominantly continuous and island permafrost N/A

Tuotuohe 92.43ºE; 34.22ºN 4535 predominantly continuous permafrost semi-arid alpine

Tazhong 83.4ºE; 39.0ºN 1099 desert desert

4.4 Auxiliary Measurements 4.5 Scaling 4.6 Data Protocols and Dissemination 4.7 Proposed Validation Tests 5. Ancillary Data

The daily F13 SSM/I brightness temperatures during the period from Oct. 1, 2002 to Sep.

30, 2003 were provided by the National Snow and Ice Data Center (NSIDC) at the

University of Colorado in the Equal Area Scalable Earth Grid (EASE-Grid) format. The

global level 3 products were used in this study, and the spatial resolution is 25 km. The

SSM/I radiometer passes over the same region twice daily at 6:00 (descending orbit) and

18:00 (ascending orbit) local time. Because the surface soil temperature at 6:00 local time

approximates the daily minimal surface temperature, the descending orbit data was selected


50

to capture the daily freeze/thaw cycle. The atmospheric influence was not corrected for the

SSM/I brightness temperature since it has an insignificant effect.

Due to the coarse spatial resolution of passive microwave remote sensing, “pure”

training samples from SSM/I data need to be collected to analyze the brightness temperature

characteristics of different land surface types and to determine the threshold of each node in

the decision tree. We selected four types of samples, including frozen soil, thawed soil,

desert and snow. The ancillary data used to ensure the purity of samples include the daily

MODIS snow cover product with 0.05º resolution (MOD10C1), the map of geocryological

regionalization and classification in China, and the Chinese land use map at 1:1,000,000

scale.

All the training samples were randomly selected according to the following criteria,

and a training sample corresponds to a SSM/I pixel. The frozen soil samples were selected

in the seasonally frozen ground region and the permafrost region from the map of

geocryological regionalization and classification in China from winter data. The thawed soil

samples were picked from the unfrozen region, and the short-term frozen ground region

from summer data. The desert samples came from the hinterland of Taklimakan according

to the Chinese land use map. The snow samples were determined if the snow fraction

derived from MODIS snow cover products was larger than 0.75 in a 25 km EASE-grid pixel.

The number of samples of frozen soil, thawed soil, desert and snow are 207, 317, 467 and

362, respectively.

The 4 cm deep soil temperatures observed by the Soil Moisture and Temperature

Measuring System (SMTMS) of the GEWEX-Coordinated Enhanced Observing Period

(CEOP) (http://monsoon.t.u-tokyo.ac.jp/ceop2/index.html) were used as validation data.

Table 1 shows the locations of the CEOP stations used in the paper.

6. Programming and Procedural Considerations 6.1 Programming Issues 6.2 Processing Issues 6.3 Quality Assurance References


51

Allison, I., Barry, R. G., & Goodison, B. E. (2001). Climate and Cryosphere (CliC) project science and co-ordination plan. WCRP-114/WMO/TD No.1053

Cao, M. S., Chang, A. C. T. (1997). Monitoring terrain soil freeze/thaw condition on Qinghai Plateau in spring and autumn using microwave remote sensing. Journal of Remote Sensing, 1(2): 139-144. in Chinese

Che, T., Li, X., Jin. R. et al. (2008). Snow depth derived from passive microwave remote sensing data in China. Annals of Glaciology, 49: 145-154

Congalton, R. G. (1991). A review of assessing the accuracy of classifications of remotely sensed data. Remote sensing of Environment, 37: 35-46

Dobson, M. C., Ulaby, F. T., Hallikainen, M. et al. (1985). Microwave dielectric behavior of wet soil- Part : four-component dielectric mixing models. IEEE Transactions on Geoscience and Remote Sensing, GE-23: 35-46

Edgeton, A. T., Stogryn, A., & Poe, G. (1971). Microwave radiometric investigations of snowpack. Final Rep. 1285R-4 of Contract 14-08-001

England, A. W. (1990). Radiobrightness of diurnally heated, freezing soil. IEEE Transactions on Geoscience and Remote Sensing, 28(4): 464-476

Fiore Jr, J. V. & Grody, N. C. (1992). Classification of snow cover and precipitation using SSM/I measurement: case studies. International Journal of Remote Sensing, 13(17): 3349-3361

Goodison, B. E., Brown, R. D. & Grane, R. G. (1998). EOS Science Plan: Chapter 6 Cryospheric System. NASA

Grody, N. C. & Basist, A. N. (1996). Global identification of snowcover using SSM/I measurement. IEEE Transactions on Geoscience and Remote Sensing, 34(1): 237-248

Grody, N. C. (1991). Classifiaction of snow cover and precipitation using the special sensor microwave imager. Journal of Geophysical Research, 96(D4): 7423-7435

He, W. Y. & Chen, H. B. (2006). Analyses of evolutional characteristics of a hailstorm precipitation from TRMM observation. Acta Meteorological Sinica, 64(3): 364-376. in Chinese

Hoekstra, P. & Delaney, A. (1974). Dielectric Properties of Soils at UHF and Microwave frequency. J. Geophys Res., 79: 1699-1708

Jin, R. & Li, X. (2002). A review on the algorithm of frozen/thaw boundary detection by using passive microwave remote sensing. Remote Sensing Technology and Application, 17(6): 370-375. in Chinese

Judge, J., Galantowicz, J. F. & England, A. W. et al. (1997). Freeze/thaw classification for prairie soils using SSM/I radiobrightnesses. IEEE Transaction On Geoscience and Remote Sensing, 35(4): 827-832

Kimball, J. S., McDonald, K. C., Keyser, A. R. et al. (2001). Application of NASA scatterometer (nscat) for determining the daily frozen and nonfrozen landscape of Alaska. Remote Sensing of Environment, 75: 113-126

Li, X., Cheng, G. D. & Jin, H. J. et al. (2008). Cryospheric change in China. Global and Planetary Change, 62 (34): 210-218, doi:10.1016/j.gloplacha.2008.02.001.

Neale, C. M. U., McFarland, M. J., Chang, K. (1990). Land-surface-type classification using microwave brightness temperature from the Special Sensor Microwave/Imager. IEEE Transactions on Geoscience and Remote Sensing, 28(5): 829-838

Smith, N. V., Saatchi, S. S., Randerson, J. T. (2004). Trends in high northern latitude soil freeze and thaw cycles from 1988 to 2002. Journal of Geophysical Research, 109, D12101, doi:10.1029/2003JD004472

Ulaby, F. T., Moore, R. K. & Fung, A. K. (1986). Microwave remote sensing: active and passive. Dedham MA: Artech House.


52

Wang S. L., Yang M. X., Toshio K. et al. (2000). Application of time-domain-reflectometer to researching moisture variation in active layer on the Tibetan Plateau. Journal of Glaciology and Geocryology, 22(1): 78-84. in Chinese

Zhang, T., Barry R. G., Knowles, K. Ling, F. & Armstrong R. L. (2003a) Distribution of seasonally and perennially frozen ground in the Northern Hemisphere, in Proceedings of the 8th International Conference on Permafrost, Zurich, Switzerland, edited by Phillips M., Springman S. M. & Arenson L. U., pp. 1289-1294, A. A. Balkema, Brookfield, Vt.

Zhang, T., Armstrong, R. L. & Smith, J. (2003b). Investigation of the near-surface soil freeze-thaw cycle in the contiguous United States: algorithm development and validation. Journal of Geophysical Research, 108(D22), doi: 10.1029/2003JD003530

Zhang, T. & Armstrong, R. L. (2001). Soil freeze/thaw cycles over snow-free land detected by passive microwave remote sensing. Geophysical Research Letters, 28(5): 763-766

Zhou, Y. W., Guo, D. X., Qiu, G. Q. et al. (2000). Geocryology in China. Beijing: Science Press. in Chinese

Zuerndorfer, B., England, A. W., Dobson, M. C. et al. (1990). Mapping freezing/thaw boundary with SMMR data. Agricultural and Meteorology, 52: 199-225

Zuerndorfer, B. & England, A. W. (1992). Radiobrightnesses decision criteria for freeze/thaw boundaries. IEEE Transaction On Geoscience and Remote Sensing, 30(1): 89-102

Acknowledgments

The work described in this publication has been supported by the EuropeanCommission (Call FP7-ENV-2007-1 Grant nr. 212921) as part of the CEOP-AEGIS project (http://www.ceop-aegis.org) coordinated by the Universityof Strasbourg, France.