Professional Documents
Culture Documents
Ashoka Trust for Research in Ecology and the Environment, No 659, 5th A Main Road, Hebbal, Bangalore, 560024, India
Stockholm Resilience Centre, Stockholm University, SE-106 91 Stockholm, Sweden
Department of Biology, University of Massachusetts, Boston, MA 02125, USA
d
Department of Genetics and Plant Breeding, University of Agricultural Sciences, GKVK, Bangalore, 560065, India
b
c
a r t i c l e
i n f o
Article history:
Received 7 April 2008
Received in revised form 24 December 2008
Accepted 26 December 2008
Keywords:
Mapping biodiversity and ecosystem services
Biodiversity surrogate
Tropical forest
Remote sensing
a b s t r a c t
There is an urgent need for techniques to rapidly and periodically measure biodiversity and ecosystem
services over large landscapes. Conventional vegetation classication and mapping approaches are based on
discrete arbitrary classes which do not capture gradual changes in forest type (and corresponding biodiversity and ecosystem services values) from site to site. We developed a simple multi-date NDVI based
Mahalanobis distance measure (called eco-climatic distance) that quanties forest type variability across a
moisture gradient for complex tropical forested landscapes on a single ecologically interpretable, continuous
scale. This Mahalanobis distance, unlike other distance measures takes into account the variability in the
reference class and shared information amongst bands as it is based on the covariance matrix, and therefore
is most useful to summarize ecological distance of a pixel to a reference class in multi-band remotely sensed
space In this study we successfully apply this measure as a surrogate for tree biodiversity and ecosystem
services at two nested scales for the Western Ghats Bio-diversity hotspot. Data from over 500 tree-plots and
forest type maps was used to test the ability of this remotely sensed distance to be a surrogate for abundance
based tree-species compositional turn-over and as a continuous measure of forest type and ecosystem
services. Our results suggest a strong but scale dependant relationship between the remotely-sensed distance measure and oristic distance between plots. The multi-date NDVI distance measure emerges as very
good quantitative surrogate for forest type and is a useful complement to existing forest classication
systems. This surrogate quanties forest type variability on a single, continuous quantitative scale and has
important applications in conservation planning and mapping and monitoring of hydrologic and carbon
storage and sequestration services.
2008 Elsevier Inc. All rights reserved.
1. Introduction
Continuing biodiversity and ecosystem services loss urgently
requires techniques to rapidly assess and monitor changes in biodiversity and ecosystem services (Balvanera et al., 2001; Menon & Bawa,1997;
Margules & Pressey, 2000; Ramesh et al., 1997; Stork & Samways, 1995).
Assessment of changes in biodiversity at a large spatial scale often
depends upon the use of surrogate information to detect trends in
general patterns because direct information on changes in the distribution of individual species or species assemblages is hard to collect
and is time consuming (Ferrier, 2002; Faithe et al., 2004). Remotely
sensed imagery combined with limited sampling on the ground has the
potential to ascertain spatial and temporal variation in biodiversity over
large spatial scales (Couteron et al., 2005; Nagendra, 2001; Rosenzweig
& Abramsky, 1993; Tuomisto et al., 2003). However, the potential of
Corresponding author. Ashoka Trust for Research in Ecology and the Environment,
No 659, 5th A Main Road, Hebbal, Bangalore, 560024, India.
E-mail address: jagdish.krishnaswamy@gmail.com (J. Krishnaswamy).
0034-4257/$ see front matter 2008 Elsevier Inc. All rights reserved.
doi:10.1016/j.rse.2008.12.011
remotely sensed imagery in mapping various components of biodiversity has not yet been fully utilized or demonstrated effectively in a
rigorous manner (Couteron et al., 2005; Lillesaeter, 1982; Nagendra &
Gadgil, 1999; Nagendra, 2001; Sanchez-Azofeifa & Rivard, 2001) with a
few notable exceptions usually in a single type of forest and at a single
spatial scale (e.g. Rocchini, 2007; Tuomisto et al., 2003).
Existing vegetation and habitat classication and mapping
approaches have several severe limitations. Forests are not mosaics of
discrete categories of vegetation types, but are a continuously changing
terrain of biological diversity and corresponding ecosystem services. In
addition, seasonal tropical forests have varying degrees of deciduousness and tree-density which inuences a range of ecological and hydroecological processes. Every site or pixel should potentially have a distinct
value on an easily interpretable ecological or ecosystem service distance
scale. Earlier methods of mapping or classication did not reect this.
However, fuzzy, sub-pixel unmixing, and other soft classication methods have the capability to reect the variability (Brown, 1997; Foody,
1996a,b; Lobell & Asner, 2004) and increasingly the utility of a continuous measure based on remote sensing is becoming apparent (e.g.
858
Asner et al., 2005; Carlson et al., 2007; Puzzolo et al., 2003; Rocchini,
2007; Tuomisto et al., 2003). There is thus an increasing need and
interest in linking biodiversity across a landscape to a well-dened and
interpretable remotely sensed continuous distance measure that is related to eco-climatic variability, and takes into account the variability of
the reference class. In this study, the rst set of criteria is met by adopting
the Mahalanobis distance measure and the second criteria by using
multi-season Normalized Difference Vegetation Index (NDVI) data for
dening the eco-climatic distance.
The Mahalanobis distance is a distance in multi-dimensional space
from a point to the centroid of a multi-dimensional variable sample. It
is superior to other distance measures since it takes into account the
variability of the reference class and the covariance matrix accounts for
shared information in correlated variables, both of which are common
feature of multi-band remotely sensed data. Formally, the Mahalanobis
distance from a group of values with mean and covariance matrix for
a multivariate vector to a point or pixel is dened as:
q
DM x = x T 1 x
"
#
$
%
= 1 ; 2 ; 3 ; 4 ; N p ; x = x1 ; x2 ; x3 ; N ; xp
Fig. 1. The eco-climatic distance map and location of ground measurement plots for two different landscape scales in the Western Ghats hotspot in India. The scale in reverse also spans
above ground carbon storage in biomass from approximately 50 Mg C ha 1 for very high distance values to over 250 Mg C ha1 at sites with very low distance values, in evergreen forests.
859
resolution of the tree-plot data (plot size) and pixel size of the remotely
sensed data. Each scale is also dened by distinct forest types that are
spread across a strong gradient of moisture availability but share common species (or abundances of species). The nested ne scale component
(landscape scale 1) was located in the Biligiri Rangaswamy Temple (BRT)
Wildlife Sanctuary, in Karnataka, described in detail later in this section.
At the broad landscape scale (landscape scale 2) our study used
independent ground and remotely sensed data for a 25,000 km2 region
(Latitude 11 32 48.31 N to 13 2 21.83 N and Longitude 75 7 38.98
E to 77 48 8.41 E) spread over the Western Ghats hotspot. Various
forest types such as evergreen, moist deciduous, dry deciduous, scrub
forests and grass lands occur in this area which span a rainfall gradient
from over 2000 mm to 500 mm yr 1. The natural forests in this area are
spread over 15,223 km2 of which 3423 km2 is evergreen forests. The
elevation ranges from 70 m to 2205 m. 3828 km2 of area is under
protection (Rodgers et al., 2002). Approximately six million people
inhabit this landscape (Census, 2001).
The comparison of the continuous remotely sensed measure to
existing, independently forest type classication was done for two
important protected areas in the Western Ghats bio-diversity hotspot:
BRT Wildlife sanctuary, and Bhadra Wildlife Sanctuary and Tiger Reserve (Fig. 2). These two protected areas are described briey below:
Biligiri Rangaswamy Temple (BRT) Wildlife Sanctuary, (latitude 11 43
to 12 08 N and longitude 77 to 77 16 E), approximately 540 km2 in size
and elevation from 600 to 1800 m. The major forest types as per the
Champion and Seth (1968) classication in the sanctuary are evergreen,
grassland, moist deciduous, dry deciduous and scrub forests. This corresponds approximately to a rainfall gradient from 1800 to 500 mm yr 1. The
reserve has an indigenous Soliga tribal community with a long history
of past shifting cultivation as well as high levels of current dependence
on forest product extraction.
Bhadra Wildlife Sanctuary was expanded and made into the present wildlife sanctuary covering an area of 492 km2 in 1974 and is now
a tiger reserve. Bhadra Tiger Reserve is located in Karnataka, India
(Longitude 75 29 E to 75 47 E and Latitude 13 22 to 13 47 N) with
the altitude ranging from 670 m to 1870 m. Bhadra receives an annual
rainfall varying spatially from less than 1000 to 2500 mm. The
vegetation in Bhadra Reserve is primarily moist deciduous forest with
bamboo and patches of dry deciduous occurring on the northern
fringes. At the higher elevation in Bababudangiri, Coffee plantations
dominate, but there are patches of a third type of forest known as the
Shola wet evergreen forests and some grassland patches occur.
2.2. Sources of data
2.2.1. Tree-plot data
The ground data used was independently obtained for each of the
two landscape scales. For landscape scale 1, data from an earlier study
(Murali et al., 1998) was used. For this study, a network of plots was
established in and adjacent to the BRT sanctuary. The study area was
divided into 2 2 km grid cells and at the centre of each cell, an
80 5 m plot was laid (Bawa et al., 2002; Murali et al., 1998). All trees
greater than 10 cm diameter at breast height (DBH) were enumerated
for each cell. The number of cells or plots that fell within the sanctuary
was 125. The total number of tree species was 127 across eight forest
types, (Ramesh & Menon, 1997).
For landscape scale 2, we used ground data from the network of 0.1 ha
plots established by the Forest Survey of India (Shivraj et al., 2000). These
plots are spread over an area of 15,000 km2 within landscape scale 2. All
tree species greater than 10 cm DBH were recorded. The Forest Maps of
South India (Pascal et al., 1982, 1992) were used to classify the vegetation
of landscape scale 2 into 18 types. At each scale the vegetation type to
which each plot belonged was determined by its location on these forest
maps. The forest types are basically a function of bio-climate and span a
strong gradient of moisture availability mainly related to rainfall and
860
Fig. 2. Comparison of eco-climatic distance maps with conventionally classied forest bio-climatic maps for two protected areas (BRT and Bhadra) with diverse forest types and
rainfall regimes.
length of dry season (Barboni et al., 2003; Champion & Seth,1968; Pascal,
1982; Prasad et al., 2005). Only plots for which recorded locations could
be veried using meta-data such as topographic sheet number and site
description and which fell within a distinct forest vegetation type were
nally used. A total of 464 plots with 320 species were included in the
study.
In summary, the density of plots is about 1 per 3.94 km2 for the ner
scale and 1 per 32.33 km2 for the coarser, broad scale. The ne scale had
regularly spaced plots, whereas the broad scale had irregularly spaced
plots (Fig. 1).
2.2.2. Remotely sensed data
The Indian Remote Sensing (IRS) satellite's LISS III images which had
23.5 m spatial resolution acquired on 2nd February 1998, 16th November
1998 and 15th April 1999 were used for the landscape scale 1. For
landscape scale 2, IRS WiFS images (188 m spatial resolution) acquired on
14th December 2000, 15th January 2001 and 23rd March 2003 were
used. These dates approximately span the period with maximum seasonal variation in eco-physiological and eco-climatic conditions. Due to
cloud cover and large area coverage, it was difcult to get enough cloud
free images for same year.
2.2.3. Climate data
The annual potential evapotranspiration (PET) at 30 min resolution
was derived from the Global data set (Ahn & Tateishi, 1994). Rainfall
data for every 5 km resolution pixel was obtained from a kriged layer
of rainfall generated from secondary rainfall data from stations in and
around the Western Ghats available from the National Institute of
Hydrology, Belgaum and Water Resources Development Organization,
Bangalore.
2.3. Data processing and analyses
To summarize, the ne scale study covered an area of about 500 km2,
plot size of 0.04 ha and a pixel resolution of 23.5 m whereas the
corresponding values are 15,000 km2, 0.1 ha and 188 m for the broader
scale study.
2.3.1. Eco-climatic distance
At each spatial scale, a new RS based distance measure called ecoclimatic distance was used to compute the distance in NDVI space for
each pixel, tree-plot or group of tree plots pooled by forest type/moisture
regime from the reference forest type class. We dened a Mahalanobis
distance to the reference class in multi-band space using the four NDVI
derived layers.
In this study the reference class vector and the pixel vector x
consists of four values (p = 4) based on three multi-date NDVI:
1. Mean of NDVI from three dates (1.end of wet-season, 2. beginning
of dry season, 3. late dry season)
2. Coefcient of variation (CV) of NDVI from the three dates
3. NDVI difference 12
4. NDVI difference 23
These four layers derived from multi-date NDVI span the period of
maximum phenological and eco-climatic variation. The rst band 1
above is a surrogate for tree density and canopy foliage biomass, and 2
above is an excellent surrogate for degree of deciduousness as related to
length of dry season, plant available moisture availability and evapotranspiration (Barboni et al., 2003; Krishnaswamy et al., 2004; Prasad
et al., 2005) and 3 and 4 are related to intra-seasonal phenological and
plant available moisture trends (Krishnaswamy et al., 2004).
Rainfall or moisture is the principal environmental inuence on
forest type in the Western Ghats (Barboni et al., 2003; Gunnel, 1997;
Pascal, 1982; Prasad et al., 2005) and we dened the reference forest
type class at one end of the rainfall/moisture gradient: evergreen forest
type. The evergreen forest type is at one end of the principal environmental gradient, is relatively stable with respect to NDVI over the dry
season, has the highest evapotranspiration, and is also the most diverse
in tree-species, and so it should be a good candidate for a reference
forest/eco-service type. The metric is based on the energy absorbed/
reected by photosynthesising foliage in the forest canopies. Plant
photosynthesis is constrained by seasonal uxes in water availability. So,
the connection with climate is that rainfall is the dominant input to the
soil water balance and the availability of water to plants. So, the metric is
more of a plant-ecophysiology distance or a plant phenology metric
as the link with climate is inferred. In addition, as this metric is related to
leaf area index, canopy cover, tree density and degree of deciduousness,
it is also a measure of an evapotranspiration and soil erosion protection
gradient.
Several pixels from this class were identied and chosen for each
landscape scale. The values of the four NDVI layers or variables for
these reference pixels dene the mean vector and covariance matrix
861
for the reference class. An S-PLUS (S-Plus, 2000) macro was written to
compute the distance from each pixel to the reference class in multi
dimensional space.
2.3.2. Forest type classication comparison
Digitized Forest type maps for BRT and Bhadra were processed
from the French Institute Bio-Climatic Map sheets. For each forest type
identied in the map, the corresponding eco-climatic pixel data were
used to generate a Box and Whiskers plot for each site.
2.3.3. Analyses of tree-plot data pooled and averaged by bio-climatic/
moisture regime/forest type
The remotely sensed eco-climatic distance was averaged by bioclimatic or moisture regime/forest type using the tree-plot locations. In
the case of the landscape scale 1 (ne), the mean value for a 7 7 window
placed around each plot was used whereas in the landscape scale 2
(broad), the exact value of the eco-climatic distance measure corresponding to the WiFS pixel (0.4 ha) within which the plot (0.1 ha) was located
was used directly. The window was used for the ner scale study rather
than the original pixel resolution to ensure that the smaller plot would be
located within a particular pixel window given the geo-referencing error.
We nally compared the map of eco-climatic distance with the
most detailed and widely used existing maps of vegetation and habitat
types produced by the French Institute at Pondicherry, India at both
landscape scales.
2.3.4. Tree species compositional distance
At each landscape scale, data from the plots within each distinct
forest type were pooled.
A macro was written in S-Plus software that computed the mean
number of trees for each species for a given vegetation type. Only those
vegetation types that had at least ve plots corresponding to it were
selected. There were four such vegetation types for landscape scale 1, 15
types respectively for landscape scale 2. We used the BrayCurtis or
Czekanowski's coefcient or Srenson (abundance) (Bray & Curtis, 1957;
EstimateS, 2000). We computed compositional distances to the reference class in EstimateS 6.0 software (EstimateS, 2000). The reference
plots were selected from the dense evergreen forest type. There were 23
reference plots for landscape scale 1 and 50 plots for landscape scale 2.
Each distance measure was computed as 1-similarity.
2.3.5. Regressions
The log transformed eco-climatic distance of each vegetation type to
the reference evergreen class was regressed against the tree species
compositional distance for those classes. The independent variable is the
log transformed eco-climatic distance and the dependant variable was
the tree species compositional distance of a vegetation type to the reference evergreen class. Regressions were done for each landscape scale. We
tested the eco-climatic distance with all four different measures of
dissimilarity distances as described above.
Image processing software Idrisi Kilimanjaro (Idrisi, 2003) and
Erdas 8.3.1 were used to process the satellite images.
2.3.6. Tree-plot level analyses
The analyses described above were done by averaging the tree-plot
data within a bio climatic/rainfall regime type. Another set of analyses
(see below) was done using the unpooled tree-plot data at both
landscape scales.
2.3.7. Quantilequantile plots
The pair-wise BrayCurtis tree-species compositional or oristic
distances were plotted against the corresponding pair-wise log (ecoclimatic) distances for both landscape scales. As these pair-wise distances are not independent of each other, and the number of data points
is very large, conventional statistical methods (e.g. regression, linear or
non-linear with specic parametric distributional assumptions for the
862
1 + w EP0
B
C
ET = @
" #1 AP:
1 + w EP0 + EP0
NDVI is a very good surrogate for Leaf Area Index (LAI), which is
strongly related to available soil moisture at rooting depth (Donohue
et al., 2007). Variability of soil moisture from site to site as summarized
in the parameter w in the above model can therefore be calibrated with
NDVI or eco-climatic distance using ground information on vegetated
sites with extreme low and high values of w (0.5 to 2, Zhang et al.,
2001). Using this calibration, and spatially explicit data on rainfall and
potential evapotranspiration, we could predict and map green-water
(evapotranspiration) and blue-water (rainfall-evapotranspiration)
uxes. Subtracting predicted evapotranspiration from rainfall gives the
blue-water ux for every pixel.
Table 1
Mantel's correlation coefcients between remotely sensed eco-climatic distance and
tree-species compositional distance
Region and landscape scale
Western Ghats (Broad)
BRT (Fine)
Mantel rm coefcients
Pooled and averaged
Un-pooled tree-plots
0.70 (n = 16)
0.63(n = 5)
0.10 (n = 464)
0.0085 (n = 125)
Signicant values (p b0.05) are marked and number of samples, n are indicated.
863
Fig. 4. The scatter plot of the eco-climatic distance measure against the tree-species compositional distance to reference evergreen forest at two landscape scales. The eco climatic distance
for different vegetation types to the reference evergreen class is solely derived from remotely sensed data and shown on a log scale. The tree-species compositional distance to the reference
evergreen class is calculated from measurements from ground plots using the BrayCurtis/Sorenson abundance based index and is pooled by forest type (bio-climatic regime) and averaged.
Each data point represents a pooled average for a particular forest type. These plots indicate that oristic distance of different forest types to evergreen forest is correlated with remotely
sensed distance of that forest type to evergreen forest.
forest types derived from the Champion and Seth and the French Institute
classication system for two protected areas reveals a close correspondence (Fig. 2). The eco-climatic distance measure effectively captures the
continuum of bio-climatic variability across environmental gradients
including eco-tones. In addition, it captures variability across the
landscape in two ecosystem functions related to ecosystem services,
evapotranspiration and carbon storage and sequestration.
864
Fig. 5. Examples of habitat or niche of three large mammals from the Western Ghats (Chital
deer (Axis axis), Sambar deer (Cervus unicolor) and Lion-Tailed Macaque (Macaca silenus)
and Western Ghats endemic bird species dened on the remotely sensed eco-climatic
distance measure. The range for each species was dened approximately using locational
information available from published studies.
5. Conclusion
Our results are noteworthy for three reasons.
First, the remotely sensed eco-climatic distance successfully quanties habitat and forest variability, with low values corresponding to
more moist, denser, more evergreen forest habitats with high evapotranspiration, and high carbon storage and higher values to more open
dry deciduous and scrub habitats with low evapotranspiration and
lower carbon storage (Figs. 1 and 2). Thus it is has quantied forest type
and ecosystem services on a single continuous numerical scale.
Furthermore we have already shown in a previous study (Das et al.,
2006) and further demonstrated it in this paper, that the RS based metric
865
Fig. 6. Spatially explicit (25 km2 resolution) (a) green-water and (b) blue-water uxes in mm per year in a part of the Western Ghats predicted by the Zhang, Walker and Dawes model
using the eco-climatic distance based prediction of plant-available moisture index. Protected areas are overlayed. An ideal measure of net hydrologic service would be a weighted
average of blue and green water uxes, adjusted for rainfall. The scale for green-water uxes above also approximately covers above ground carbon storage from less than
approximately 50 Mg C ha 1 to over 250 Mg C ha 1.
866
867
Rodgers, W. A., Panwar, H. S., & Mathur, V. B. (2002). Wildlife protected area network in
India: a review (executive summary). Dehradun, India: Wildlife Institute of India.
Rosenzweig, M. L., & Abramsky, Z. (1993). How are diversity and productivity related? In
R. Ricklefs, & D. Schluter (Eds.), Species diversity in ecological communities: historical
and geographical perspectives (pp. 5265). Chicago: University of Chicago Press.
Sanchez-Azofeifa, G. A., & Rivard, B. (2001). Advancing frontiers in tropical biology: the
future of remote sensing research in tropical environments. In K. N. Ganeshaiah, R.
Uma Shaanker, & K. S. Bawa (Eds.), Tropical ecosystems: structure, diversity and human
welfare (pp. 143144). New Delhi: Oxford & IBH Publishing Co.
Seidler, T. G., & Plotkin, J. B. (2006). Seed dispersal and spatial pattern in tropical trees.
PLoS Biology, 4(11), e344. doi:10.1371/journal.pbio.0040344
Shivraj, B., Barve, N., Kiran, M. C., Uma Shaanker, R. U., & Ganeshaiah, K. N. (2000).
Mapping of forests based on biological diversity to identify conservation sites: a
case study from Udupi and South Canara districts of Karnataka. Journal of Indian
Institute of Science, 80, 531536.
S-Plus (2000). S-PLUS 2000 Professional Release 1. Cambridge, MA: Mathsoft Inc.
Stork, N. E., & Samways, M. J. (1995). Inventorying and monitoring. In V. H. Heyward (Ed.),
Global biodiversity assessment (pp. 453543). New York: Cambridge Press.
Tuomisto, H., Poulsen, A. D., Ruokolainen, K., Moran, R. C., Quintana, C., Cell, J., et al. (2003).
Linking oristic patterns with soil heterogeneity and satellite imagery in Ecuadorian
Amazonia. Ecological applications, 13, 352371.
Waide, R. B., Willig, M. R., Steiner, C. F., Mittelbach, G., Gough, L., Dodson, S. I., et al. (2000).
The relationship between productivity and species richness. Annual Review of
Ecology and Systematics, 30, 257300.
Webb, L. J. (1968). Environmental relations of the structural types of Australian rainforest
vegetation. Ecology, 49, 296311.
Webb, L. J., Tracy, J. G., Williams, W. T., & Lance, G. N. (1970). Studies in the numerical
analysis of complex rainforest communities. V. A comparison of the properties of
oristic and physiognomic-structural data. Journal of Ecology, 58, 203232.
Wilk, M. B., & Gnanadesikan, R. (1968). Probability plotting methods for the analysis of
data. Biometrika, 55, 117.
Zhang, L., Dawes, W. R., & Walker, G. R. (2001). The response of mean annual evapotranspiration to vegetation changes at catchment scale. Water Resource Research, 37,
701708.