a combined approach for estimating vegetation cover

Upload: zafeersaqib

Post on 09-Apr-2018

219 views

Category:

Documents


0 download

TRANSCRIPT

  • 8/7/2019 A combined approach for estimating vegetation cover

    1/11

    Computers & Geosciences 32 (2006) 12991309

    A combined approach for estimating vegetation cover in urban/

    suburban environments from remotely sensed data

    Chen Yunhao, Shi Peijun, Li Xiaobing, Chen Jin, Li Jing

    College of Resources Science and Technology, Beijing Normal University, Beijing 100875, China

    Received 24 August 2005; received in revised form 23 November 2005; accepted 24 November 2005

    Abstract

    The spatio-temporal distribution of vegetation is an important component of the urban/suburban environment.

    Therefore, correct estimation of vegetation cover in urban/suburban areas is fundamental in land use studies. In this study,

    the potential of extracting fractional vegetation cover (FVC) from remotely sensed data and ground measurements is

    explored. Based on the assumption that pixel has a mosaic structure, sub-pixel models for FVC estimation are first

    introduced. Then a combined approach of using different sub-pixel models for FVC estimation based on land cover

    classification is proposed. The experimental result, derived from a case study in Haidian district, Beijing, indicates that the

    accuracy of FVC estimation using the proposed method can be up to 80.7%. The results suggest that this method may be

    generally useful for FVC estimation in urban and suburban areas.

    r 2005 Elsevier Ltd. All rights reserved.

    Keywords: Remote sensing; Vegetation cover; Land use classification; Urban spread

    1. Introduction

    Recent estimates indicate that over 45% of the

    worlds human population now lives in urban areas,

    with this figure rising to over 60% projected by 2030

    (United Nations, 1997). Monitoring urban environ-

    ment change will therefore become increasingly

    important as the number and proportion of urbanresidents continue to increase. The spatio-temporal

    distribution of vegetation is a fundamental compo-

    nent of the urban/suburban environment (Small,

    2001). Research has shown that vegetation influ-

    ences urban environmental conditions and energy

    fluxes by selective reflection and absorption of solar

    radiation. The presence of fractional vegetation

    cover (FVC), which is defined as the percentage of

    vegetation occupying a unit area, in urban areas

    may also influence air quality and even human

    health (Wagrowski and Hites, 1997).

    Table 1 shows the different methods used now in

    FVC estimation. Field measurements such as

    methods of sampling, instrumental methods andocular estimation are the traditional methods used

    to determine the FVC (Zhou et al., 1998; White

    et al., 2000). This approach is usually too expensive,

    and remote sensing is an effective tool for observing

    the distribution and fraction of the vegetation cover.

    Two categories of methods, namely spectral mixture

    analysis (SMA) and vegetation indices (VIs), are the

    most frequently used techniques to estimate FVC

    (Camacho-de et al., 2004).

    ARTICLE IN PRESS

    www.elsevier.com/locate/cageo

    0098-3004/$- see front matterr 2005 Elsevier Ltd. All rights reserved.

    doi:10.1016/j.cageo.2005.11.011

    Corresponding author.

    E-mail address: [email protected] (C. Yunhao).

    http://www.elsevier.com/locate/cageohttp://www.elsevier.com/locate/cageo
  • 8/7/2019 A combined approach for estimating vegetation cover

    2/11

    There are several SMA methods (Adams et al.,

    1986; Smith et al., 1990; Settle and Campbell, 1998;

    Oki et al., 2004) that allow a user to estimate

    coverage at a sub-pixel level. However, the spectral

    radiance (end member) of each category covered

    within a pixel must be precisely known before

    carrying out the unmixing method (Quarmby et al.,

    1992; Foody and Cox, 1994). Despite major

    advances made in the field of remote sensing, the

    number of parameters that interact in the problem isstill greater than the intrinsic dimension of the

    spectral data that are provided by these sensors

    (Camacho-de et al., 2004). Thus, traditional SMA

    methods are difficult to apply to regions with urban

    and suburban areas with a homogeneous land

    surface.

    The methods of FVC estimation using VI data

    can be grouped into two sub-categories: empirical

    models and transformed models (Table 1). In the

    situation of empirical models, the FVC is derived

    from established empirical relationships between

    FVC data obtained on the ground and vegetation

    indices. Graetz et al. (1988) measured FVC in a

    semi-arid soil region, and proposed a linear regres-

    sion formulation to estimate the FVC. Many other

    empirical FVC models have been presented accord-

    ing to the features of specified study areas using

    multi-scale remotely sensed data (Dymond et al.,

    1992; Wittich and Hansing, 1995; Purevdorj et al.,

    1998). These empirical models can produce a

    reliable estimation of the FVC in specified areas,

    but are often difficult to use because of their

    dependency on good ground data.

    The main aim of transformed models is to

    establish the relationship between the vegetation

    index and vegetation fraction to estimate the FVC

    directly. Gutman and Ignatov (1998) presented

    uniform-pixel model and mosaic-pixel models

    (including the dense vegetation model, non-dense

    vegetation model, variable density vegetation mod-

    el) according to probable vegetation distributed

    features in pixels, and estimated global FVC from

    the NOAA AVHRR data using the dense vegeta-tion model. Extensive fieldwork is unnecessary in

    applying the transformed methods, but their accu-

    racy is often lower than that of empirical models.

    Thus, improving the accuracy of transformed

    models has become a key problem in the develop-

    ment of these methods.

    The main objective of this study is to develop a

    new approach for urban/suburban FVC estimation

    from remotely sensed data and field surveys,

    combining the dense vegetation and non-dense

    vegetation models. Two steps are involved in this

    approach: choosing appropriate sub-pixel models,

    and determining parameters for these models

    (especially in the non-dense vegetation model). A

    quantitative validation is to examine the applic-

    ability of the combined approach to the estimation

    of the vegetation fraction in suburban of Beijing

    using Landsat thematic mapper (TM) data.

    2. Theoretical basis

    NDVI is an integrative reaction parameter of

    vegetation types, cover fraction and growth status

    ARTICLE IN PRESS

    Table 1

    Comparisons between different FVC estimation methods

    Methods Advantage Disadvantage Used image

    Field

    measurements

    Sampling Need much manpower and

    financial resources; dependency

    on personal experience

    Digital camera data

    Instrumental method Higher accuracy

    Ocular estimation

    Spectral mixture

    analysis

    Linear mixture models Calculation simple with

    reasonable accuracy

    Cannot broadly apply to the

    urban and suburban

    TM/ Spot/ Ikonos

    Nonlinear mixture

    models

    Vegetation indices Empirical models Reasonable accuracy According to the features of

    specified study areas

    TM/ Spot

    Transformed models Less fieldwork with

    reasonable accuracy

    Dependency on good ground

    data

    NOAA/TM/Spot

    C. Yunhao et al. / Computers & Geosciences 32 (2006) 129913091300

  • 8/7/2019 A combined approach for estimating vegetation cover

    3/11

    in a unit pixel. Its value is determined by leaf area

    index (LAI) (vertical density) and FVC (horizontal

    density) (Price, 1992). Many empirical models have

    shown that there is a close relationship between

    FVC and NDVI, but the same NDVI signal may

    result from different sub-pixel vegetation structures(vertical and horizontal densities). Hence, to explore

    the numerical relationship between NDVIand FVC

    further, it is necessary to analyze the sub-pixel

    structure of an image pixel. According to different

    land cover structures, pixels can be divided into two

    types: uniform pixels and mosaic pixels (Gutman

    and Ignatov, 1998). When a pixel is fully covered by

    green vegetation, it is considered that the pixel is a

    uniform pixel, whereas a pixel partly covered by

    green vegetation is a mosaic pixel. This research

    simplified the sub-pixel types of mosaic pixels

    into: dense vegetation and non-dense vegetation,illustrated in Table 2.

    2.1. Uniform pixel

    In Table 2, depending upon the pixel type,

    relationships between the vegetation index (NDVI),

    fractional vegetation cover (FVC) and leaf area

    index (LAI) can be described as follows:

    In the situation of a uniform pixel, the pixel is

    fully covered by green vegetation (FVC 1) with acertain vertical density (Table 2, top). Hence, the

    value of NDVI is determined mainly by the LAI.

    The equation of LAI and NDVI is presented based

    on a modified Beers law (Baret and Guyot, 1991):

    NDVI NDVI1 NDVI1 NDVI0

    expk LAI, 1

    where NDVI0 and NDVIN are the signals of bare

    soil (LAI-0) and dense green vegetation (LAI-N),

    respectively, and k is the extinction coefficient.

    2.2. Mosaic pixel

    Mosaic pixels are partly covered by green

    vegetation, so the NDVI observed is a weighted

    average of the vegetated (NDVIg) and non-vege-

    tated (NDVI0) parts. According to the features of

    mosaic pixels on the vegetation coverage, two sub-

    pixel models can be described as follows:

    2.2.1. Dense vegetation model

    In this model, an assumption is made that

    the vegetation type is single and the density

    of the vegetated part of the pixel is very high(LAI-N, and NDVI-NDVIN

    ) correspondingly,

    so that

    NDVI FVC NDVI1 1 FVCNDVI0.

    (2)

    Hence, the FVC yields

    FVC NDVI NDVI0

    NDVI1 NDVI0. (3)

    2.2.2. Non-dense vegetation modelSimilar to the dense vegetation model, an

    assumption is made that the vegetation type is

    single but the density of vegetated part of the pixel is

    low LAI51. A combination of Eqs. (1) and (2)

    yields

    NDVI FVC NDVIg 1 FVCNDVI0, (4)

    ARTICLE IN PRESS

    Table 2

    Schematic representation of sub-pixel models for FVC estimation

    Pixel types Sub-pixel structure of vegetation Schematic map Formulation of FVC

    Uniform pixel Uniform full vegetation fg 1, NDVI value is determined by LAI mainly

    Mosaic pixel Dense vegetation fg NDVINDVI0

    NDVI1NDVI0

    Non-dense vegetation fg NDVINDVI0

    NDVIgNDVI0

    Gutman and Ignatov (1998), modified.

    C. Yunhao et al. / Computers & Geosciences 32 (2006) 12991309 1301

  • 8/7/2019 A combined approach for estimating vegetation cover

    4/11

    where

    NDVIg NDVI1 NDVI1 NDVI0

    expk LAI. 5

    Thus,

    FVC NDVI NDVI0

    NDVIg NDVI0. (6)

    From Eqs. (2)(6) we note that calculation of

    FVC is more complicated in the non-dense than in

    the dense vegetation model. In the former, not only

    NDVI0 and NDVIN, but also the extinction

    coefficient k and LAI, must be determined.

    3. Model

    3.1. Model selection

    It is straightforward to estimate the FVC with the

    dense vegetation model, which is suitable for

    estimating the FVC of a region covered by a single

    vegetation type with high vertical density. The non-

    dense vegetation model is suitable for estimating the

    FVC of a region covered by a single vegetation type

    with low vertical density. However, it is difficult to

    determine the parameters in this model.

    Taking into account the spatial resolution ofLandsat TM data, it is reasonable to choose a sub-

    pixel model among the uniform and mosaic-pixel

    models for urban and suburban FVC estimation.

    Considering the characteristics of the above-men-

    tioned sub-pixel models and any constraints, an

    appropriate sub-pixel model based on land cover

    classification should improve the precision of

    estimation. The models that follow were derived

    from our study and represent a range of vegetation

    structure.

    Orchard: single vegetation type with high vertical

    density, (LAI-N). Hence, the dense vegetation

    model is adopted.

    Woodland: complicated vegetation types with

    high vertical density. Accounting for limited spatial

    resolution of image, the dense vegetation model is

    chosen approximately.

    Meadow: is a single vegetation type and the value

    of LAI is suitable for the non-dense vegetation

    model. Hence, the non-dense vegetation model is

    adopted.

    Cropland: our study area is in the outskirts of

    Beijing. Vegetable gardens and lawns take up most

    of the land area, and winter wheat and broomcorn

    are rare, so the non-dense vegetation is adopted.

    City zone: its sub-pixel character is complex and

    arrays of trees are the main green vegetation part:

    except for large lawn areas where the non-dense

    vegetation model is applied, the dense vegetationmodel is adopted.

    3.2. Determination of parameters

    The parameters common to the dense and non-

    dense vegetation models are: NDVI0 and NDVIN,

    which are defined by Sellers et al. (1996, 1997) as the

    lower and upper 25% NDVI for each biome.

    Furthermore, it is required to determine k and LAI

    in the non-dense vegetation model. k is the light

    extinction coefficient, characterizing the vegetation

    type (Choudhury et al., 1994):

    k lnAPAR=PAR

    LAI, (7)

    where PAR relates to the visible part of the

    spectrum between 0.4 and 0.7mm, where chloro-

    phyll absorbs solar radiation, namely PAR is thus a

    fraction of the incoming solar radiation. APAR is

    absorbed PAR by the vegetation.

    The k values for meadow and cropland were

    determined using Eq. (7), and the parameters used

    for calculation of k were observed by field survey.PAR and APAR were taken by a SUNSCAN

    Canopy Analysis System and the LAI measurement

    was taken by an LAI-2000 plant canopy analyzer

    (PCA). The operation instructions for the PAR and

    LAI instrument were followed carefully to ensure

    that each point was measured accurately. Each

    PAR, APAR and LAI measurement represents an

    average of 10 PCA readings which were taken in a

    sample.

    Directly measured LAI may provide a reliable

    result, however, it is difficult to estimate a regional

    LAI by using this method, because there of the

    problem of transferring the data from a point scale

    to a surface scale (Chen et al., 2001). Price (1993)

    developed a two-stream approximation model for

    LAI estimation from remotely sensed data. Para-

    meters needed in the estimation of LAI are

    estimated by using a scattergram of visible and

    near-infrared reflectance measurements of vegeta-

    tion, along with the reflectance measurements

    derived from remotely sensed data. The LAI can

    be obtained using Prices method, but this requires

    knowledge of a number of constants: (1) the soil line

    ARTICLE IN PRESS

    C. Yunhao et al. / Computers & Geosciences 32 (2006) 129913091302

  • 8/7/2019 A combined approach for estimating vegetation cover

    5/11

    constants (see Fig. 3A), a and b, are required (under

    some conditions, information about the soil line

    may be identified from a scattergram of remotely

    sensed data); (2) the coefficients c1 and c2 used to

    describe the attenuation of radiation as it passes

    through successive layers of leaves, which differbetween vegetation types; (3) the reflectance values

    for dense vegetation rN

    , should also be estimated

    from remotely sensed data.

    Price (1993) has proposed an approach to identify

    parameters needed from remotely sensed data. The

    relationship between LAI and DNi, the measured

    values from a satellite sensor, can be written as

    (i 1,2 relating to TM band 3 and TM band 4

    respectively)

    DNsi

    DNie2ciLAI r21i DN1i1 e

    2ciLAI

    1 r21ie2ciLAI DNir21i1 e2ciLAI=DN1i,

    (8)

    where the DNNi can be estimated from remotely

    sensed data, for rN1 the value 0.05 and for rN2 the

    value 0.7 can be used. Putting Eq. (8) into the soil

    line equation, which can be obtained from satellite

    data in visible and near-infrared, we can express

    LAI as a function of DNi and c1, c2

    DNs2 a0DNs1 b

    0, (9)

    where a0 and b0 can be regressed from the

    scattergram of visible and near-infrared reflectance

    measurements.

    Although the functional equation of LAI and

    DNi can be expressed as a polynomial equation

    and be solved for certain values of c1, c2, a

    numerical solution is generally satisfactory, so a

    look-up table of LAI values corresponding to an

    array of DNi values is constructed, followed by

    interpolation.

    ARTICLE IN PRESS

    Fig. 1. Location of study area.

    C. Yunhao et al. / Computers & Geosciences 32 (2006) 12991309 1303

  • 8/7/2019 A combined approach for estimating vegetation cover

    6/11

    3.3. Study area and data

    Haidian district, an area in the northeast of Beijing,

    was chosen to examine the effect of the FVC

    estimation method (Fig. 1). Its hypsography is low

    in the east, high in the west, and average horizon levelis about 50 m. Of the total area, plains occupy 75%,

    and mountains and hills occupy 25%. Land cover

    types and the FVC are variable over the study area.

    A Landsat TM scene image of the study area,

    which was acquired on 19 May 2004 (path and row

    numbers are 123 and 32), was adopted as the main

    data source. After image preprocessing, land cover

    classification, model selection for FVC estimation,

    and FVC calculations for the Haidian district were

    carried out. The flowchart for the FVC estimation is

    illustrated in Fig. 2.

    4. Methodology

    4.1. Geometric registration

    Image-to-map registration was carried out using

    1:50 000 topographic maps. Geometric accuracy is

    less than 1 pixel at all land areas.

    4.2. Land cover classification

    Supervised classification of the TM image was

    carried using the maximum likelihood algorithm in

    ERDAS 8.7 applied to the six non-thermal bands.

    The training sites were defined using a false colourcomposite of channels 4, 3 and 2 displayed as red,

    green and blue, respectively. Three to five training

    areas were randomly chosen for each class. The

    following classes were used: meadow, cropland,

    woodland, orchard, urban, water zone, and bare

    area. The correctness of the land cover types was

    evaluated in relation to the ground data collected by

    field survey in study area in May, 2004. The field

    data were positioned with a GPS and classified by

    visual estimation. The digital interpretation indi-

    cates a total accuracy of 85.8% (Table 3).

    4.3. FVC estimation

    Sub-pixel models for the FVC estimation were

    selected based on the land cover classification. A

    value of 0 is assigned to the value of the FVC in

    water zones and bare areas, and the non-dense

    vegetation model was applied to estimate meadow

    ARTICLE IN PRESS

    Model selection for vegetation fraction based onland cover classification and field survey

    Woodland Orchard Urban Meadow Cropland Water zone Bare area

    Dense vegetation model Non-dense vegetation model

    A two stream approximation model

    Urban vegetation fraction map

    TM image

    Geometric registration/ Study zone selection

    Land cover classification

    Field survey

    Fig. 2. Flowchart for FVC estimation based on remotely sensed data and field survey.

    C. Yunhao et al. / Computers & Geosciences 32 (2006) 129913091304

  • 8/7/2019 A combined approach for estimating vegetation cover

    7/11

    and cropland. Meanwhile, the dense vegetation

    model was adopted for woodland, orchard and

    urban sites. The parameters for these models are

    illustrated in Table 4. In the non-dense vegetation

    model, the two-stream approximation method is

    used to determine the LAI. The scattergram of the

    TM band 4 (infrared) and TM band 3 (visible) is

    given in Fig. 3A. Constants needed for the LAI

    estimation are: DNs11 19, DNs2

    1 9; DNs12 102,

    DNs2

    2 95 for the soil line; and DNN

    1 16,

    DNN2 122 for the dense vegetation, where the

    subscript 1, 2 refers to the grey level value of the TM

    band 3 and 4, respectively. Two coefficients, c1 and

    c2, are chosen as 0.6 and 0.21 for the TM data,

    which are applicable to cropland and meadow with

    low vertical density (Price, 1993). Using these

    parameters, a look up table, relating LAI to DN1and DN2, is constructed and shown in Fig. 3B. The

    distribution of the LAI map for study area can be

    obtained by interpolation from the look up table.

    Fig. 4A shows the distribution of the FVC map

    for Haidian district on May 19, 2004. The spatial

    distribution features of the FVC map may be

    summarized as follows: (1) Generally, the FVC is

    low over the whole region. Only in the western

    narrow mountainous regions and hills does the

    vegetation cover exceed 50%, (2) in the city zone,

    only the FVC of scattered parks can achieve

    4055%, the other regions are mainly 1025%.

    This indicates that green space protection is lacking

    during urbanization of the Haidian district,

    (3) crops in the study area are younger in the

    beginning of May, so the FVC of northern cropland

    is low.

    5. Results and discussions

    5.1. Results analysis

    To analyze accuracy of above-mentioned FVC

    estimation quantitatively, we carried out ground

    FVC measurements in the middle ten days of May,

    2004. In order to evaluate the effect of accuracy

    improvement by using the combined approach, the

    ARTICLE IN PRESS

    Table 3

    Error matrix calculated for the supervised classification results

    Ground data Class Users accuracy

    Image 1 2 3 4 5 6 7 Total

    1 16 2 1 0 0 0 0 19 0.842 1 14 3 0 0 0 0 18 0.78

    3 1 2 26 1 0 0 0 30 0.87

    4 0 0 2 35 3 0 0 40 0.88

    5 0 0 0 2 22 1 0 25 0.88

    6 0 0 0 1 0 4 0 5 0.8

    7 0 0 0 0 0 0 4 4 1

    Total 18 18 32 39 25 5 4 141

    Producers accuracy 0.89 0.78 0.81 0.9 0.88 0.8 1 TA 0.858

    Classes: 1 meadow grassland; 2 cropland; 3 orchard; 4 woodland; 5 urban; 6 water zone; 7 bare area. Users and Producers accuracy are

    shown in last column and last row, respectively. TA is total accuracy.

    Table 4

    Parameters used for the FVC estimation in Haidian district, Beijing

    Land cover types Sub-pixel types NDVI0 NDVIN k

    Meadow Non-dense vegetation model 0.05 0.646 0.82

    Cropland Non-dense vegetation model 0.05 0.656 0.94

    Woodland Dense vegetation model 0.05 0.718

    Orchard Dense vegetation model 0.05 0.713

    Urban Dense vegetation model 0.05 0.627

    Water zone, bare soil 0

    C. Yunhao et al. / Computers & Geosciences 32 (2006) 12991309 1305

  • 8/7/2019 A combined approach for estimating vegetation cover

    8/11

    FVC was also calculated using the dense vegetation

    model, and the result is shown in Fig. 4B.

    Forty-five sample patches of different land-cover

    type were selected, including meadows (8 patches),

    croplands (6 patches), woodlands (12 patches),

    orchards (7 patches) and urban (12 patches). The

    smallest sample patch was 30 30 m. The FVC field

    survey was carried out with a digital camera

    following the procedure described by Calera et al.

    (2001) and using the line-intercept method (Duncan

    et al., 1993). Table 5 presents a comparison of the

    FVC values observed on the ground and those

    estimated by the model.

    The analysis showed that estimated FVC values are

    highly correlated with those observed on the ground,

    with an average accuracy of 80.7%. We compared the

    use of a single model (the dense vegetation model)

    with the combined approach. Due to the dense model

    results are already available for the woodland,

    orchard and urban land cover types, the dense model

    results would only have to be compiled for the

    meadow and cropland types (Table 5).

    Table 5 shows that the accuracy of meadow and

    cropland FVC using the dense model are 73 and

    68%, whereas the FVC using the combined

    approach are 86 and 78%, respectively. The results

    show that the combined model as compared to the

    dense model is more reasonable over the hetero-

    geneous land surface.

    Fig. 5 presents a comparison of FVC values

    observed on the ground and those estimated by the

    combined approach. Calculated FVC had a bias

    ARTICLE IN PRESS

    Dense vegetation model Combined approach

    0.9

    0.7~0.8

    0.8~0.9

    0.1~0.2

    0.4~0.5

    0.2~0.3

    0.5~0.6

    0.3~0.4

    FVC

    N 40009

    E 116003

    N 40009

    E 116023

    N39035E 116003

    N39035E 116023

    N 40009

    E 116003

    N 40009

    E 116023

    N39035E 116003

    N39035E 116023

    N

    0 5 km 0 5 km

    N

    (a) (b)

    Fig. 4. Estimated FVC in Haidian District, Beijing.

    Fig. 3. (A) Scattergram of Landsat/TM scene for Haidian District; (B) Look-up table relating LAI to digital values of TM3 and TM4.

    C. Yunhao et al. / Computers & Geosciences 32 (2006) 129913091306

  • 8/7/2019 A combined approach for estimating vegetation cover

    9/11

    (mean estimated FVCobserved FVC) of 6.89%.

    Regression analysis was carried out to find the

    linear trend and bias of the curve from the 1:1 line.

    The analysis showed that estimated FVC are highly

    correlated with those observed (r2

    0:83, n 45).Comparison of estimated FVC and observed data

    demonstrates that this approach may be an im-

    portant tool for estimating FVC.

    5.2. Comparison with soil adjusted vegetation index

    (SAVI)

    Huete (1988) suggested soil adjusted vegetation

    index (SAVI) based on NDVI and various observed

    to eliminate the effect from background soils:

    SAVIrNIR rredrNIR rred

    1 L, (10)

    where L is a constant that is empirically determined

    to minimize the vegetation index sensitivity to soil

    background reflectance variation. L varies with

    vegetation density, ranging from 0 for very high

    vegetation cover to 1 for very low vegetation cover.

    For intermediate vegetation cover ranges, L is

    typically around 0.5 (Schowengerdt, 1997). The

    SAVI index can be used to estimate the amount of

    pixel mixing between vegetation and non-vegetation

    portions. The objectives of the FVC approach and

    the SAVI approach are similar, but they are

    different in the following respects:

    (1) The aim of the SAVI is to calculate a vegetation

    index with minimizing the influence of soil on

    canopy vegetation reflectance. However, the

    objective of the FVC approach is to develop a

    new method for fractional vegetation cover

    estimation;

    (2) this paper demonstrated that the FVC estima-

    tion approach is suitable for urban/suburban

    ARTICLE IN PRESS

    Table 5

    Error analysis based on field survey

    Sample

    no.

    Woodland (%) Urban (%) Orchard (%) Meadow (%) Cropland (%)

    Obs. Est. 1 Err. 1 Obs. Est. 1 Err. 1 Obs. Est. 1 Err. 1 Obs. Est. 1 Err. 1 Est. 2 Err. 2 Obs. Est. 1 Err. 1 Est. 2 Err. 2

    1 53 42 0.21 18 13 0.28 44 37 0.16 53 41 0.23 44 0.17 30 36 0.20 22 0.272 75 68 0.09 14 18 0.29 75 63 0.16 39 31 0.21 29 0.26 31 22 0.29 19 0.39

    3 37 30 0.19 14 10 0.29 36 41 0.14 44 52 0.18 58 0.32 43 32 0.26 32 0.26

    4 43 51 0.19 26 23 0.12 43 52 0.21 40 45 0.13 27 0.33 52 46 0.12 40 0.23

    5 51 44 0.14 20 25 0.25 45 36 0.20 51 48 0.06 41 0.20 33 41 0.24 47 0.42

    6 33 37 0.12 29 37 0.28 37 45 0.22 33 38 0.15 48 0.45 34 26 0.24 21 0.38

    7 74 82 0.11 15 12 0.20 71 61 0.14 74 67 0.09 60 0.19

    8 62 73 0.18 13 9 0.31 52 46 0.12 40 0.23

    9 44 52 0.18 20 17 0.15

    10 33 38 0.15 26 21 0.19

    11 49 41 0.16 23 16 0.30

    12 28 36 0.29 20 15 0.25

    Avg1 0.17 0.24 0.17 0.14 0.27 0.22 0.32

    Avg2 0.193

    Obs. observed value, Est. 1 estimated value 1 using the combined approach, Est. 2 estimated value 1 using the dense vegetation model, Err.1 error using the combined approach, Err. 2 error using the dense vegetation model, Avg1 the absolute mean errors for each cover type,

    Avg2 the absolute mean error over all samples using the dense vegetation model.

    y = 0.9045x + 5.0635

    R2 = 0.8321

    0

    10

    20

    30

    4050

    60

    70

    80

    90

    0 20 40 60 80 100

    Estimated Vegetation Fraction Values (%)

    ObservedVegetationF

    ractionValues(%)

    Fig. 5. Relationship between FVC values observed and estimated

    from remotely sensed data.

    C. Yunhao et al. / Computers & Geosciences 32 (2006) 12991309 1307

  • 8/7/2019 A combined approach for estimating vegetation cover

    10/11

    areas with a homogeneous land surface. The

    SAVI approach is generally a superior method

    for estimating a green vegetation index over

    rural areas at broad spatial scales;

    (3) since vegetation density is usually unknown, it is

    difficult to optimize the SAVI index (Liang,2004). On the other hand, the SAVI was shown

    to be sensitive to NIR variations induced by

    sensor/sun geometry. In comparison to the

    SAVI, the FVC approach had reasonable

    parameters based on land cover classification.

    Thus the new approach has been shown to be a

    good estimator of the urban FVC.

    6. Conclusions

    In this study, a new method for fractional

    vegetation cover estimation in urban/suburban

    areas is derived. This method takes into account

    the sub-pixel vegetation structure, and deals with

    the parameters of the dense and non-dense vegeta-

    tion models specifically. Through a case study in

    Haidian district, Beijing, a good agreement (the

    accuracy of FVC estimation is about 80.7%) is

    obtained between estimated and the ground-based

    measurements, which is higher than using the dense

    vegetation model only. This combined approach

    may be an effective tool in monitoring the fractionalvegetation cover in urban and suburban areas.

    Acknowledgments

    This work was supported by the Natural Science

    Foundation of China (no. 40201036), the National

    Key Developing Program for Basic Sciences of

    China (no. 2006CB701302) and the Foundation of

    Key Lab of Resources Environment and GIS of

    Beijing, Capital Normal University. We are grateful

    to the anonymous referees for their thoughtful andhelpful comments.

    References

    Adams, J.B., Smith, M.O., Johnston, P.E., 1986. Spectral mixture

    modeling: A new analysis of rock and soil types at the viking

    lander 1 site. Journal of Geophysical Research 91, 80988112.

    Baret, F., Guyot, G., 1991. Potentials and limits of vegetation

    indices for LAI and APAR assessment. Remote Sensing of

    Environment 35, 161173.

    Calera, A., Martinez, C., Melia, J., 2001. A procedure for

    obtaining green plant cover: relation to NDVI in a case study

    for barley. International Journal of Remote Sensing 22,

    33573362.

    Camacho-de, C.F., Garcia-hare, F.J., Gilabert, M.A., Melia, J.,

    2004. Vegetation cover seasonal changes assessment from TM

    imagery in a semi-arid landscape. International Journal of

    Remote Sensing 25, 34513476.

    Chen, J., Chen, Y., He, C., Shi, P., 2001. Sub-pixel model forvegetation fraction estimation based on land cover classifica-

    tion. Journal of Remote Sensing 5, 416422.

    Choudhury, B.J., Nizam, U.A., Sherwood, B.I., 1994. Relations

    between evaporation coefficients and vegetation indices

    studied by model simulations. Remote Sensing of Environ-

    ment 50, 117.

    Duncan, J., Stow, D., Franklin, J., Hope, A., 1993. Assessing the

    relationship between spectral vegetation indices and shrub

    cover in the Jornada Basin, New Mexico. International

    Journal of Remote Sensing 14, 33953416.

    Dymond, J.R., Stephens, P.R., Newsome, P.F., 1992. Percent

    vegetation cover of a degrading rangeland from SPOT.

    International Journal of Remote Sensing 13, 19992007.

    Foody, G.M., Cox, D.P., 1994. Sub-pixel land cover compositionestimation using a linear mixture model and fuzzy member-

    ship functions. International Journal of Remote Sensing 15,

    619631.

    Graetz, R.D., Pech, R.R., Davis, A.W., 1988. The assessment

    and monitoring of sparsely vegetated rangelands using

    calibrated Landsat data. International Journal of Remote

    Sensing 9, 12011222.

    Gutman, G., Ignatov, A., 1998. The derivation of the green FVC

    from NOAA/AVHRR data for use in numerical weather

    prediction models. International Journal of Remote Sensing

    19, 15331543.

    Huete, A.R., 1988. A soil adjusted vegetation index (SAVI).

    Remote Sensing of Environment 25, 295309.

    Liang, S., 2004. Quantitative Remote Sensing of Land Surfaces,

    first ed. Wiley, New York, NY 560pp.

    Oki, K., Uenishi, T.M., Omasa, K., 2004. Accuracy of land cover

    area estimated from coarse spatial resolution images using an

    unmixing method. International Journal of Remote Sensing

    25, 16731683.

    Price, J.C., 1992. Estimating vegetation amount from visible and

    near infrared reflectance. Remote Sensing of Environment 41,

    2934.

    Price, J.C., 1993. Estimating leaf area index from satellite data.

    IEEE Transactions on Geoscience and Remote Sensing 31,

    727734.

    Purevdorj, T., Tateishi, R., Ishiyama, T., 1998. Relationships

    between percent vegetation cover and vegetation indices.International Journal of Remote Sensing 19, 35193535.

    Quarmby, N.A., Townshend, J.R.G., Settle, J.J., 1992. Linear

    mixture modelling applied to AHVRR data for crop area

    estimation. International Journal of Remote Sensing 13,

    415425.

    Schowengerdt, R.A., 1997. Remote Sensing Models and Methods

    for Image Processing, second ed. Academic Press, San Diego,

    CA, pp. 179187.

    Sellers, P.J., Los, S.O., Tucker, C.J., Justice, C.O., Dazlich, D.A.,

    Collatz, G.J., Randall, D.A., 1996. A revised land

    surface parameterization (Sib2) for atmospheric GCMS. Part

    II: the generation of global fields of terrestrial biophysical

    parameters from satellite data. Journal of Climate 9,

    706737.

    ARTICLE IN PRESS

    C. Yunhao et al. / Computers & Geosciences 32 (2006) 129913091308

  • 8/7/2019 A combined approach for estimating vegetation cover

    11/11

    Sellers, P.J., Dickinson, R.E., Randall, A., Betts, A.K., Hall,

    F.G., Berry, J.A., Collatz, G.J., Denning, A.S., Mooney,

    H.A., Nobre, C.A., Sato, N., Field, C.B., Henderson-Sellers,

    A., 1997. Modeling the exchanges of energy water and carbon

    between continents and the atmosphere. Science 275,

    502509.

    Settle, J., Campbell, N., 1998. On the errors of two estimators ofsub-pixel fractional cover when mixing is linear. IEEE

    Transactions on Geoscience and Remote Sensing 36,

    163170.

    Small, C., 2001. Estimation of urban vegetation abundance by

    spectral mixture analysis. International Journal of Remote

    Sensing 22, 13051334.

    Smith, M.O., Ustin, S.L., Adams, J.B., Gillespie, A.R., 1990.

    Vegetation in deserts: I. a regional measure of abundance

    from multispectral images. Remote Sensing of Environment

    31, 126.

    United Nations, 1997. Prospects for Urbanization. ST/ESA/

    SER.A/166, Sales No. E.97. XIII.3.

    Wagrowski, D.M., Hites, R.A., 1997. Polycyclic aromatic

    hydrocarbon accumulation in urban, suburban and rural

    vegetation. Environmental Science and Technology 31,

    279282.

    White, M.A., Asner, G.P., Nemani, R.R., Rrivette, J.L.,Running, S.M., 2000. Measuring fractional cover and leaf

    area index in arid ecosystem: digital camera, radiation

    transmittance, and laser altimetry methods. Remote Sensing

    of Environment 74, 4557.

    Wittich, K.P., Hansing, O., 1995. Area-averaged vegetative cover

    fraction estimated from satellite data. International Journal of

    Biometeorology 38, 209215.

    Zhou, Q., Robson, M., Pilesjo, P., 1998. On the ground

    estimation of vegetation cover in Australian rangelands.

    International Journal of Remote Sensing 19, 18151820.

    ARTICLE IN PRESS

    C. Yunhao et al. / Computers & Geosciences 32 (2006) 12991309 1309