Phenology based Classification Model for Vegetation Mapping using IRSWiFS

1
Phenology based Classification Model for
Vegetation Mapping using IRS-WiFS Shefali
Agrawal, Sarnam Singh, P.K.Joshi and
P.S.Roy Indian Institute of Remote Sensing, 4
Kalidas Road, Dehradun   Introduction   The
establishment and implementation of procedures
for vegetation classification has a long history
using remotely sensed data from various sensors
such as Landsat-MSS and TM, SPOT-XS, IRS-LISS-III
and NOAA AVHRR. The techniques used for NOAA
AVHRR based land cover classification are similar
to those by other digital multispectral image,
the only difference being the analysis uses a
high frequency multi temporal data set. The
earlier demonstrations of the suitability of
AVHRR data set for large area land cover mapping
have been reported by Tucker et al.(1985) for
land cover classification of Africa and by
Townshed et al. (1987) for south America land
cover classification. These investigations showed
that resampled AVHRR Global Area Coverage (GAC)
data were an alternative to Landsat data for
large area land cover mapping due to lower data
volume, cost and higher temporal frequency. As a
result a number of studies have been reported
using coarse resolution AVHRR data to map on
national to continental scales by using different
classification algorithms viz. supervised and
unsupervised methods on multi temporal data sets.
  Multi temporal remote sensing data are widely
acknowledged as having significant advantages
over single date imagery (Townshed et al., 1985).
Mapping of land cover can be improved by using
variations in phonological patterns of
vegetation. Phenological differences are also
very useful in the detection of large-scale
vegetation disturbances. The use of multitemporal
images not only results in higher classification
accuracy but also gives consistent accuracy in
all classes. Use of multitemporal data is
especially advantageous in areas where vegetation
or land use changes rapidly. This offers many
opportunities for more complete vegetation
description than could be achieved with a single
image. For example, the differences between
evergreen and deciduous trees can be highlighted
by the fact that the former may appear quite
uniform throughout the year, whereas the latter
varies widely between leaf-on and leaf-off
periods. The discriminant power of multitemporal
observations is based on their characterization
of seasonal dynamics of vegetation growth
(phenology).  
2
 Classification Methods   Traditional
classification methods such as supervised and
unsupervised methods of multispectral image
classification depends on spectral reflectance.
Land cover classification using remote sensing is
based on the assumption that different types of
land cover have distinct reflectance properties.
The unique spectral properties of a land cover
class is governed by canopy geometry, leaf
densities, colors, optical properties and
moisture content, shadow components,
transpiration rates, and non-vegetated
reflectances. These factors contribute to the
reflectance in each pixel, and the total number
of pixels within each class offer a set of mean
values and variances for classification. These
mean values and variances represent the central
tendency and spread of a class respectively and
together are referred as spectral signature. The
signatures for each class are collected and
subjected to a statistical classifiers
(supervised and unsupervised), which assign each
pixel on the image to one of the classes
according to some form of best-match
algorithm.   Vegetation indices (VIs) and derived
metrics have been extensively used for monitoring
and detecting vegetation and land cover change
(deFries et al. 1995). The development of
vegetation indices is based on differential
absorption, transmittance, and reflectance of
energy by the vegetation in the red and
near-infrared regions of the electromagnetic
spectrum (Jensen 1996). Various studies have
indicated that only Normalised Difference
Vegetation Index (NDVI) is least affected by
topographic factors. Vegetation indices condense
the data from two (or more) spectral bands into
one level of information. Vegetation indices are
especially advantageous with multi date data
sets. Various multidate vegetation indices are
clustered to classify broad areas (usually at
continental scale) according to the seasonalities
of their greenup/senescence sequences. Therefore
compared to land cover classification using
single date data, multitemporal datasets are
often found to improve the accuracy of
classification. Further the classification can
also be improved by using the phenological
metrices derived from NDVI viz. maximum, minimum,
amplitude , average and time integrated NDVI,
which can be used as a layer or added band for
classification in combination with a rule-based
approach for determining cover types. Maximum
NDVI is the maximum measurable NDVI recorded
during the year and is normally associated with
the peak of green during the growing season and
the corresponding lowest NDVI value recorded
during the year is referred as the minimum NDVI.
Mean NDVI is the maximum NDVI value obtained for
each recording period during the growing season
divided by the total number of periods.   The
Principal Component Analysis (PCA) is one of the
best the best known data reduction techniques
where in multispectral imagery is transformed
into a lesser number of principal component image
bands. PCA reduces the dimensionality of a data
set containing large number of interrelated
variables, while still retaining as much as
possible the variation present in the data set.
This reduction is achieved by transforming to a
new set of variables, the principal components,
which are uncorrelated, and which are ordered so
that the first few retain most of the variation
present in all of the original variables.
Therefore, Principal Component images always
contain most of the original input image variance
in a lesser number of bands.  
3
On applying unstandardized PCA of time series
NDVI it was observed that Principal Component 1
could be interpreted as the time-integrated NDVI
over the entire three year period representing
the typical greenness of the continent (Eastman,
1992). Principal Component 2 is interpreted as a
change component, representing winter/summer
seasonal effect. Principal Components 3 and 4 are
also essentially seasonal, but represent areas
where the timing of greenup is different than
that for component 2. Higher order components are
interpreted as sensor artifacts or relatively
short-term meteorological effects. As is common
in PCA, the interpretation of higher order
components becomes progressively difficult and it
is not clear how many components are significant
in terms of information (Jackson, 1993). Present
Methodology   In the study an attempt has been
made to classify vegetation over northeastern
part of India using distinct phenological growth
stages and spectral characteristic at mesoscale.
Multi date IRS Wide field Sensor (WiFS) data has
been used for this purpose. IRS-WIFS data with
two spectral bands red (0.62-0.68?m) and infrared
(0.77-0.86?m) at a spatial resolution of 188m and
temporal resolution for 3-5 days meets the
requirement of vegetation mapping at regional and
continental scale using phenological variability
in vegetation. In the present analysis temporal
vegetation characteristics over a five month
period at different phonological stages are
analysed by considering three datasets
corresponding to maturity (December, January and
February), senescent (March) and leaf fall
(April) periods.   The satellite data was first
corrected for atmospheric effects due to
scattering using dark pixel subtraction
technique. The data was then geometrically
rectified using control points and all the images
from different months were co-registered. To use
the different aspect of vegetation phenology for
classification, the multidate data sets was
subjected to various analytical procedures viz.
Vegetation Indices and Principal Component
Analysis. Maximum, minimum, mean and amplitude
NDVI were calculated on different season data.
The resulting NDVI images were subjected to grey
level scaling in order to segregate vegetation
types into broad categories based on NDVI values.
The multidate data set of February, November and
December was compressed into three principal
components.   Land use/cover characterization
was attempted first by using unsupervised
classification technique on the raw data layers
in combination with the maximun NDVI data.
K-means algorithm was run on maximum NDVI value
and the raw bands of November. Each cluster was
assigned a preliminary cover type label taking
care of the spatial pattern and spectral or multi
temporal statistics of each class on comparison
with ancillary data and extensive ground truth.
Ancillary data included descriptive land cover
information, NDVI profiles and class
relationships to the other land cover classes.
The classes were then grouped into broad classes
using a convergence of evidence approach. The
snow and cloud classes were masked out. The
unsupervised classification was followed by post
classification refinement for the coherent set of
classes.  
4
Results and Discussion   The unique climatic
condition of northeast India supports luxuriant
vegetation growth resulting in extensive forest
cover. Different types of forest have been
identified in northeastern region by Champion and
Seth (1968). According to the latest satellite
based survey report of Forest Survey of India
(FSI), northeastern region has 164359 Km2 of
forest, approximately 25 of the total forest
cover in the country (Anonymous, 1997). However,
due to human activities such as shifting
cultivation have brought considerable change in
the ecological status of the forests. Shifting
cultivation (locally called jhumming) is the
single factor responsible for forest and land
degradation. About 0.45 million families in this
region cultivate 10,000 sq. km forests annually
affecting approximately 44,000 Km2 of forest area
(Singh, 1990).   In the present study, attempts
were taken to stratify forest using the temporal
data set and to observe the phenological
variation among the different types of forest in
different regions. The temporal NDVI images
provide the rhythmic growth of vegetation and
hence able to distinguish the same species type
occurring in different biogeographical or
climatic conditions. Even the abandoned shifting
cultivation areas, which have attained good
growth of tree species or bamboo, have been
identified in the different regions using
temporal data set. Four different time period
NDVI images are considered as representative for
seasonal changes. The NDVI values varies from
1.0 to 1.0. However, values for land surfaces
were ranging from -1.0 to 0.992 (February),
0.357 to 0.994 (March), - 0.352 to 0.576
(April)- 0.449 to 0.758 (November) and - 0.492
to 0.748 (December). The maximum NDVI image has
been computed to represent the maximum foliage
cover in the study period.   Temporal plots were
selected for each landuse class and analyzed for
the study area. NDVI values obtained from the
vegetation index product/image for different
cover types were assessed. The representative
sites selected for each cover type indicate the
internal variation of the NDVI response of the
cover type. For each location area averaged NDVI
value was assessed. The NDVI images showed the
foliage cover in the respective season. The
maximum NDVI image represents the maximum foliage
cover or greenness in the study period for each
forest legends (Figure 1). The coniferous
locations have comparatively low NDVI values
throughout the study period striking uni-modal
peak in December. The broad-leaved forest is
having small peak during month of December with a
steep decline in values from February to March.
The semievergreen forest is having moderate NDVI
values with small peak during December and March.
The moist deciduous types are having high
photosynthetic activity during the study period
with highest during March. The secondary forest
(abandoned jhum gt10 yrs) shows the bimodal NDVI
values during the study period with high value in
March. From the preliminary analyses it is
apparent that different cover types exhibit
characteristic NDVI curves. The non-forest
classes viz. degraded grasses/shrubs and
agriculture showed almost similar pattern of NDVI
values throughout the year having high foliage
curve/photosynthetic activity during December.
The Bamboo jhum (5-10 yrs) showed high NDVI
values in the month of December and decline in
March. Because of high cloud cover during March,
the NDVI values do not follow the trend of
temporal variation i.e. phenology. The monitoring
of the crop development through growing period is
possible for the agricultural areas. However in
case of northeastern India such an approach will
be rather difficult due to recurring cloud
cover.  
5
In the present case the PCs of the data set
consisting of February, November and December
1998 data were studied. The scenes with high
cloud cover were rejected to overcome the
contamination. The inverse principal component
was carried out to visualise the data in RGB.
The false color composite of the first three PC
images was found to be informative over the raw
data sets and NDVI images. The PC1 was containing
information from all the bands of the three
dates. The PC 2 is collective information from
the IR bands of the each image hence supporting
the vegetation and land cover. The PC3 was having
contribution of IR and R band of November data.
The FCC of PC images provides this discrimination
between forest and non-forest inspite of healthy
foliage cover. (Figure 2) Within the forest
classes the discrimination among the forest types
is also highlighted. The broad-leaved and
coniferous evergreen forests were discriminated
as per the NIR response. The semi evergreen
patches were found to be enhanced and
distinguished. The fresh jhum patches and
abandoned jhum classes were identified as
different classes. The maximum NDVI image gives
intermixing among the northern healthy forest of
Arunachal Pradesh and southern part of northeast
i.e. Mizoram, which is dominated by the abandoned
jhum. The discriminating loading factors of the
PC2 represented the abandoned jhum and degraded
forest. Within non-forest classes, the
agriculture patches and tea gardens were
distinguished. The seasonal and permanent water
bodies were also clearly distinguished.   This
region is endowed with vast natural resources in
the form of tropical evergreen/semi-evergreen,
subtropical evergreen forest, moist deciduous,
temperate broad-leaved forest, temperate conifer
forest, alpine grasslands/scrub and secondary
forest and fresh water streams, rivers and lakes
(Figure 3). The forest cover area estimated is
about 42.24 of the geographical area. The forest
cover recorded by FSI is found overestimated in
comparison to previous and present studies
carried out using satellite remote sensing. This
is attributed to the fact that the maps prepared
by using visual interpretation are unable to
separate abandoned jhum and is grouped with
open/degraded forest. However in the present
analysis this class could be separated as
abandoned jhum (5-10 years). The total forest
cover of northeastern region including jhum is
worked out as 55.06 that is almost equal to
forest cover reported by FSI i.e.
54.02.   Conclusion   Accuracy assessment of the
classification was performed by using confusion
matrix. From the error matrix of the classified
forest cover it was observed that among the
various forest classes, evergreen and moist
deciduous forest showed relatively low users
accuracy. The above classes got mixed with
degraded forest and patches of agriculture. The
non-forest classes have shown higher accuracy
except agriculture and jhum (lt10 yrs). It may be
due to intermixing with moist deciduous forest.
The fresh jhum (lt10 yrs) normally occurred in
various stages of succession and various cover
type combinations. The overall accuracy was
82.15 and Kappa statistics was 80.03 in
agreement (Khat coefficient 0.80).  
6
The present study highlights the use of NDVI and
metrices derived from it and use of other
enhancement techniques like principal component
analysis for land use/land cover mapping. The
NDVI has been found related to green leaf
activity and as such provides a useful means to
monitor the vegetation cover/phenology. Its
effectiveness lies in its discrimination ability
among forest types and major crops and other land
cover classes. The accuracy have been found to be
satisfactory (accuracy 80 to 87) to perform
forest cover assessment, mapping and
delineation.   Phenological derived metrices viz
maximum, mean, minimum NDVI, integrated NDVI in
combination with raw data layers on multitemporal
data sets were also applied for vegetation cover
mapping in other regions (Gujarat, Himachal
Pradesh and Madhya Pradesh) of India and was
found to be satisfactory for land cover
characterization at regional and global scales.
  References   Champion, H.G. and Seth, S.K.,
1968. In A revised survey of forest types of
India, New Delhi Govt. Publication.   deFries,
R., M. Hansen, J. Townshend, 1995. Global
discrimination of land cover types from metrics
derived from AVHRR pathfinder data, Remote
Sensing of Environment 54(3) 209-222.
  Eastman, J.R., 1992. Time series map analysis
using standardized principal components.
ASPRS/ACSM/RT 92 Technical Papers, Vol. 1 Global
Change and Education. Aug. 3-8, Wash. D.C., pp.
195-204.   Jensen, J.R. 1996. Introductory
Digital Image Processing A Remote Sensing
Perspective, Prentice Hall, New Jersey,
316p   Roberts Miles, Well Chris, Doyle, Thomas
W., 1994. Component analysis for interpretation
of time series NDVI imagery, ASPRS/ACSM.   Roy,
P.S., Sarnam Singh, Agrawal Shefali, Joshi, P.K.,
2001 Assesment of Forest cover in North east
India and Northern Myanmar- Potential of Indian
Remote sensing satellite (IRS-1C WiFS) Data.
IIRS-JRC Report.   Singh, S., Agrawal S., Joshi,
P.K., and Roy, P.S., 1999. Biome Level
Classification of Vegetation in Western India- An
application of Wide Field View Sensor (WiFS).
Joint workshop of ISPRS Working Groups I/1,I,3
and IV/4 Sensors and Mapping from Space,
Hannover(Germany) 27-30 Sept. 1999  
7
Singh, G., 1990. Soil and water conservation in
India, In Proceeding Symposium on Water Erosion,
Settlement and Resource Conservation, March 25,
Deharadun. Systems of the northeastern hill
region of India. Agro-ecosystem,
7,11-25. Townshend, J.R.G., Golf, T.E., and
Tucker, C.J., 1985, Multispectral Dimensionality
of Images of Normalized Difference Vegetation
Index at Continental Scales, IEEE Transaction on
Geoscience Remote Sensing, 23,888-895.
    Townshend, J., Justice, C., and Kalb, V.,
1987, Characterization and Classification of
South American Land Cover Types Using Satellite
data, International Journal of Remote Sensing,
8,1189-1207.   Tucker, C.J., Townshend, R.G., and
Goff, T., 1985, Continental land cover
classification using NOAA-7 AVHRR data, Science,
227, 369-375.
8
Figure 1
Maximum NDVI Image North East India
Image MAX (February, March, April, November,
December)
Projection Lambert Conformal Conic
9
Figure 2
False Color Composite of Principal Components -
North East India
PC1PC2PC3
Projection Lambert Conformal Conic
10
Figure 3
Forest Cover Map North East India Level II
Projection Lambert Conformal Conic