You are on page 1of 15

International Journal of Remote Sensing

ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/tres20

Mapping plant functional types in Northwest


Himalayan foothills of India using random forest
algorithm in Google Earth Engine

Ritika Srinet , Subrata Nandy , Hitendra Padalia , Surajit Ghosh ,


Taibanganba Watham , N. R. Patel & Prakash Chauhan

To cite this article: Ritika Srinet , Subrata Nandy , Hitendra Padalia , Surajit Ghosh , Taibanganba
Watham , N. R. Patel & Prakash Chauhan (2020) Mapping plant functional types in Northwest
Himalayan foothills of India using random forest algorithm in Google Earth Engine, International
Journal of Remote Sensing, 41:18, 1-14, DOI: 10.1080/01431161.2020.1766147

To link to this article: https://doi.org/10.1080/01431161.2020.1766147

Published online: 30 Jun 2020.

Submit your article to this journal

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at


https://www.tandfonline.com/action/journalInformation?journalCode=tres20
INTERNATIONAL JOURNAL OF REMOTE SENSING
2020, VOL. 41, NO. 18, 1–14
https://doi.org/10.1080/01431161.2020.1766147

Mapping plant functional types in Northwest Himalayan


foothills of India using random forest algorithm in Google
Earth Engine
Ritika Srineta, Subrata Nandy a, Hitendra Padaliaa, Surajit Ghosh a,b
,
Taibanganba Wathama, N. R. Patela and Prakash Chauhana
a
Indian Institute of Remote Sensing, Indian Space Research Organisation, Department of Space, Government
of India, Dehradun, India; bInternational Water Management Institute, Colombo, Sri Lanka

ABSTRACT
Plant functional types (PFTs) have been widely used to represent
the vegetation characteristics and their interlinkage with the sur­ ARTICLE HISTORY
rounding environment in various earth system models. The present Received 26 September 2019
study aims to generate a PFT map for the Northwest Himalayan Accepted 27 April 2020
(NWH) foothills of India using seasonality parameters, topographic
conditions, and climatic information from various satellite data and
products using Random Forest (RF) algorithm in Google Earth
Engine (GEE) platform. The seasonality information was extracted
by carrying out a harmonic analysis of Normalized Difference
Vegetation Index (NDVI) time-series (2008 to 2018) from Moderate
Resolution Imaging Spectroradiometer (MODIS) Terra surface
reflectance 8 day 500 m data (MOD09A1). For topographic informa­
tion, Shuttle Radar Topography Mission (SRTM) digital elevation
model (DEM) derived aspect and Multi-Scale Topographic Position
Index (MTPI) were used, whereas, for climatic variables, WorldClim
V2 Bioclimatic (Bioclim) variables were used. RF, a machine learning
classifier, was used to generate a PFT map using these datasets. The
overall accuracy of the resulting PFT map was found to be 83.33%
with a Kappa coefficient of 0.71. The present study provides an
effective approach for PFT classification using different well-
established, freely available satellite data and products in the GEE
platform. This approach can also be implemented in different eco­
logical settings by using various meaningful variables at varying
resolutions.

1. Introduction
Terrestrial vegetation plays an important role in regulating the exchange of energy and
mass between the land surface and the atmosphere (Xia et al. 2014; Fatichi et al. 2019;
Patel et al. 2019). It contributes to the global carbon cycle by regulating and storing
atmospheric carbon dioxide through various metabolic processes (Canadell et al. 2000;
Nandy et al. 2019). The structure and functions of vegetation in various terrestrial
ecosystems vary significantly from place to place, as it is associated with the prevailing

CONTACT Subrata Nandy subrato.nandy@gmail.com Indian Institute of Remote Sensing, Indian Space Research
Organisation, Department of Space, Government of India, Dehradun, India
© 2020 Informa UK Limited, trading as Taylor & Francis Group
2 R. SRINET ET AL.

climatic, edaphic and topographic conditions. The biophysical parameters like leaf area
index (LAI), fraction of absorbed photosynthetically active radiation (FAPAR), surface
roughness, and albedo have a strong control over the carbon, water and energy fluxes
between the vegetation and atmosphere (Castanho et al. 2013; Srinet, Nandy, and Patel
2019; Pillai et al. 2019). Hence, the distribution of vegetation holds an important place in
determining the spatial patterns of biogeochemical cycling (Sellers et al. 1996). Due to the
control of vegetation on climatic feedback, a reliable description of vegetation types and
their patterns is, thus, a key component in ecosystem models, including land surface
models (LSMs), dynamic vegetation models (DVMs) and earth system models (ESMs), with
relevance to climate predictions, carbon dynamics and ecosystem modelling (Poulter
et al. 2015).
Within ecosystem models, the vegetation distribution is generally represented in the
form of plant functional types (PFTs). PFTs can be defined as sets of plants exhibiting
similar responses to environmental conditions and having similar effects on the dominant
ecosystem processes (Walker 1992; Smith et al. 1997). PFTs vary with respect to phenol­
ogy, morphological and photosynthetic parameters and are usually constrained by an
individual bioclimatic range of tolerance defined by temperature thresholds (Dallmeyer,
Claussen, and Brovkin 2019). Detailed and accurate information regarding the distribution
of PFTs can contribute to improved predictive capabilities of the global and regional
carbon cycle, climate and ecosystem models (Sun and Liang 2008). Highly diverse terres­
trial ecosystems and the variability in morphological and spectral characteristics of PFTs
make its mapping a challenging task in large areas (Semenova and van der Maarel 2000).
Various studies have used remote sensing (RS) techniques in combination with differ­
ent approaches for PFTs classification. At the global level, Moderate Resolution Imaging
Spectroradiometer (MODIS) PFT map (MODIS Land Cover Type 5, MCD12Q1) is freely
available (https://lpdaac.usgs.gov/products/mcd12q1v006/). MCD12Q1 PFT product is
generated by relabelling or cross-walking MODIS land cover product, which follows the
International Geosphere-Biosphere Programme (IGBP) scheme (Friedl et al. 2010). Some
studies have used the existing land cover data sets to extract PFTs (Bonan et al. 2002;
Verant et al. 2004; Lapola et al. 2008; Poulter et al. 2011; Ottlé et al. 2013). The limitation
with this method is that the accuracies of the pre-existing land cover maps may affect the
accuracy of the PFT classification. Broad PFT categories can be most useful over large
geographic areas, while more narrowly defined types may be necessary for regional levels.
To produce PFT maps at regional levels, earlier studies have applied multisource eviden­
tial reasoning (Sun et al. 2008), hierarchical soft time-space classification (Cai et al. 2014)
and classification using machine learning algorithms like Support Vector Machine (SVM)
(Padalia et al. 2019). Both parametric and non-parametric classifiers have been used in the
past to classify RS images. However, the non-parametric classifiers do not make any
assumptions regarding frequency distribution and have, therefore, emerged as more
accurate and efficient methods for classifying remotely sensed data (Rodriguez-Galiano
et al. 2012; Belgiu and Drăguţ 2016).
The Northwest Himalayan (NWH) foothills of India have a unique climatic and topo­
graphic setting (Navalgund, Kumar, and Nandy 2019). The forests in the area are diverse
and provide various ecosystem services. It is essential to address the phenological
variability in the vegetation as well as the variation in topography and climate to prepare
an accurate PFT classification for the NWH foothills of India. The processing of long-term
INTERNATIONAL JOURNAL OF REMOTE SENSING 3

RS data, for time-series analysis at the regional level, requires high storage capability and
efficient computational power (Campos-Taberner et al. 2018). Google Earth Engine (GEE)
is a cloud-computing platform designed for geospatial analysis (Gorelick et al. 2017). With
the help of an application program interface (API), it enables users to access and analyse
a repository of publicly available geospatial datasets using various algorithms at Google’s
cloud computing and processing platform (Gorelick et al. 2017). GEE also provides various
classifiers for pixel-based classification, including machine learning algorithms, for map­
ping applications (Johansen, Phinn, and Taylor 2015; Azzari and Lobell 2017; Acharya et al.
2019; Mahdianpari et al. 2019). The present study aims to generate a PFT map of the NWH
foothills of India using Random Forest (RF) classification, utilizing the seasonality informa­
tion derived from MODIS Terra 8 day global surface reflectance time-series products
(MOD09A1) in combination with topographic information extracted from Shuttle Radar
Topography Mission (SRTM) digital elevation model (DEM) and climate information from
WorldClim V2 Bioclimatic (Bioclim) variables in GEE platform.

2. Materials and methods


2.1. Study area
The present study was carried in the NWH foothills (Figure 1) of India. The study area
is spread across Uttarakhand, Uttar Pradesh, Himachal Pradesh, Punjab, and Jammu
and Kashmir. The area has tropical to subtropical humid climate. The mean annual
temperature ranges from 20 to 25°C. The area experiences most of its rainfall from
south-west monsoon from June-end to September. The topography in NWH foothills
varies widely and comprises of flat to undulating terrain with hills and valleys. The
elevation in the study area ranges from 187 to 1300 m above mean sea level. The

Figure 1. Location of the study area.


4 R. SRINET ET AL.

vegetation present in the study area varies as per the climate and topographic
settings. In NWH foothills of India, the radiation and moisture regimes differ in the
northern and southern aspects, which has resulted in variation in the vegetation
compositions. The northern aspect areas of the foothills, which experience cooler
temperatures and retain more soil moisture, comprise of denser vegetation of moist
deciduous nature, whereas the southern aspect areas have drier vegetation, which
can endure higher temperatures and lower moisture conditions.
Major forest type groups present in the study area are Tropical moist deciduous (3 C)
and dry deciduous (5B) which can be further divided into Moist Shiwalik Sal Forest (3 C/
C2a), Moist Bhabar Sal Forest (3 C/C2b), Moist Tarai Sal Forest (3 C/C2 c), Dry Shiwalik Sal
Forest (5B/C1a), Dry plain sal forest (5B/C1b), Northern dry mixed deciduous (5B/C2), Dry
tropical riverain forest (5B/1S1), and Khair-Sissu forest (5B/1S2) (Champion and Seth 1968).

2.2. Data used


2.2.1. MODIS data
MODIS Terra surface reflectance 8 day 500 m (MOD09A1) data from 2008 to 2018 was
used for the present study. MOD09A1 provides atmospherically corrected estimates of the
surface spectral reflectance of Terra MODIS Bands 1 to 7 within 8 day composite period.
Normalized difference vegetation index (NDVI) was calculated using band 1 (red) and
band 2 (near-infrared) of MOD09A1 to obtain a time-series of 8 day NDVI data over the
study area in GEE.

2.2.2. Topographic data


As the ecosystem functioning is strongly influenced by the availability of soil moisture,
evapotranspiration, the intensity of insolation and shading (Whittaker 1956), there is
a need to take into account the aspect and topographic positioning. The aspect was
generated from SRTM DEM. To account for the topographic positioning, Global SRTM-
derived Multi-Scale Topographic Position Index (MTPI) (Theobald et al. 2015) was used.
This index is used to define the relative position of a location along a topographic
gradient (ridge top, middle slope or valley). It is calculated using elevation data for each
location subtracted by the mean elevation within a neighbourhood (Guisan, Weiss, and
Weiss 1999).

2.2.3. Climate data


To represent the climatic conditions in the study area, WorldClim V2 Bioclimatic (Bioclim)
variables (Fick and Hijmans 2017) were used. It comprises of 19 bioclimatic variables
representing annual trends (mean annual temperature, annual precipitation), seasonality
(annual range in temperature and precipitation), and extreme or limiting environmental
factors (temperature of the coldest and warmest month, and precipitation of the wet and
dry quarters).

2.2.4. Field data


GPS coordinates of various land cover classes were collected by field survey. The
collected points were spatially well distributed in the study area and were randomly
divided into training and testing sets. The training set, comprising of 500 points, was
INTERNATIONAL JOURNAL OF REMOTE SENSING 5

used to train the classifier whereas the testing set, comprising of 150 points, was
used for accuracy assessment.

2.3. Methodology
The classification scheme used in the study is developed keeping in mind the vegetation
type present in the study area. To map the PFTs of the study area, phenological trends,
climate regime, aspect, and topographic position were considered. The detailed metho­
dology is shown in Figure 2.

2.3.1. Seasonality parameters


Harmonic analysis of time-series of 8 day NDVI, calculated from MOD09A1 data, was
carried out to differentiate the vegetation based on their phenological behaviour. The
harmonic analysis permits a complex curve to be expressed as the sum of a series of sine
or cosine waves with an additive term (Rayner 1971).
Any periodic process can be considered as:
Xt ¼ A cosð2πωt þ φÞ (1)

for t = 0, ±1, ±2, . . .


where, Xt is a periodic process, ω is a frequency index, defined in cycles per unit time, A
is the amplitude of the function and φ is the phase which determines the start point of the
cosine function. Each wave can be defined by a unique amplitude and phase value. Using
trigonometric identity, Equation (1) can be written as:
Xt ¼ β1 cosð2πωtÞ þ β2 sinð2πωtÞ (2)

where, β1 ¼ AcosðφÞ and β2 ¼ AsinðφÞ are normally distributed random variables, and in
�1=2 � �
this case, amplitude A ¼ β1 2 þ β2 2 , and phaseφ ¼ tan 1 β2=β .
1

Figure 2. Methodology adopted for plant functional type mapping in northwest Himalayan foothills of
India.
6 R. SRINET ET AL.

The successive harmonic terms are summed to get a complex curve that accounts for
the total variation in the original time-series dataset. The amplitude represents the
maximum increase in the values whereas the phase value represents the period of that
increment. The spatial information on amplitude, phase and mean value was produced for
every pixel in the study area.

2.3.2. PFT Classification and accuracy assessment


The dataset used for classification included amplitude, phase and mean NDVI extracted from
NDVI time-series derived using MOD09A1, aspect, MTPI, and Bioclim variables. RF, a machine-
learning algorithm, was used for classification. It is an ensemble classifier that uses a set of
classification and regression trees for prediction. These trees are created by drawing a subset
of training samples using a bagging approach. About two-thirds of the samples (in-bag
samples) are used to train the trees whereas the remaining one third (out-of-the bag samples)
are used in internal cross-validation to estimate the performance of the RF model (Breiman
2001). To tune the algorithm, two input parameters are required, namely, NTree and MTry,
which represent the number of trees grown based on a bootstrap sample of the observations
and the number of variables fed to each predictor tree, respectively. For classification, each
pixel is evaluated against all the decision trees created in the ensemble and the class label is
assigned based on the majority of votes of the trees. The classification was carried out using
different values of NTree and MTry and the optimum value was decided based on the
classification accuracy. For MTry, NTree was fixed and several values of MTry (1/3rd, 1/2, 2/3rd
of the total number of input features, square root of the total number of input features and
the total number of input features) were tested and the optimum value was obtained. For
this optimum value of MTry, various values of NTree were tested (100, 200, 300, 500, 800 and
1000) for obtaining the optimum NTree value based upon the accuracy of classification. A set
of 500 training points was used to train the RF classifier. The classified map was tested for
producers, users and overall accuracies as well as Kappa coefficient was determined using
150 points. The accuracy of classified PFTs was also compared with MODIS PFT.

3. Results
3.1. Seasonality
The seasonal changes in NDVI values in the time-series represent the phenological cycle of the
vegetation. The harmonic analysis of NDVI time-series revealed the differences in the phase
and amplitude values of different types of vegetation. The phase represents the greening
period and the amplitude represents the peak of the greening period. The phase and
amplitude values in combination with mean NDVI values were able to present a good contrast
between different types of vegetation present in the study area. Figure 3 shows the season­
ality map of the study area, represented as a composite of phase, amplitude and mean NDVI
values.

3.2. PFT classification and accuracy assessment


The capability to differentiate the vegetation based on their seasonal behaviour was taken
into account and using this seasonality in combination with topographic and climatic
INTERNATIONAL JOURNAL OF REMOTE SENSING 7

Figure 3. Seasonality parameters extracted from NDVI time-series.

variables, a PFT classification was carried out using RF classifier. In the present study, MTry
values of 8, 12, 16, 5, and 24 were tested. This resulted in marginal or no influence on the
classification accuracies, hence the square root of the total number of variables, i.e., 5
were selected for MTry. NTree value of 800 was found to be suitable for the present study, as
the error did not increase beyond this point. Using these optimum MTry and NTree values in
RF classifier, PFT classification was carried out.
The PFT map (Figure 4) consisted of 5 classes including moist deciduous (MD), dry
deciduous (DD), shrubs or grasses (S/G), water bodies (W) and non-forest (NF), which
include crop and built-up areas. Figure 5 represents the statistical distribution of (a) phase,
(b) amplitude and (c) mean NDVI obtained from the harmonic analysis of NDVI time-series
for different PFT classes using box-and-whisker plots. The black horizontal lines within
boxes illustrate median values, boxes demonstrate the lower and upper quartiles, and
whiskers extend to the minimum and maximum values of the phase, amplitude and mean
NDVI. The median of phase and mean NDVI values obtained for MD class was compara­
tively higher than the other classes followed by DD, whereas the median of the amplitude
value of DD class was the highest. These values were able to present a separability within
the classes.
The overall accuracy of the PFT classification was found to be 83.33% with a Kappa
coefficient of 0.71 (Table 1). The producer’s accuracies for MD, DD, S/G, NF, and
W categories were found to be 97.14%, 79.55%, 68.42%, 75.76%, and 94.74%, respectively.
Whereas the user’s accuracies for the respective classes were 87.18%, 94.59%, 76.47%,
8 R. SRINET ET AL.

Figure 4. Plant functional types of northwest Himalayan foothills of India.

Figure 5. Box-and-whisker plot illustrating the distribution of (a) phase, (b) amplitude and (c) mean
NDVI obtained from the harmonic analysis of NDVI time-series for different PFT classes.

67.57%, and 90.00%. The error of omission for MD class was found to be only 2.86%
whereas for DD it was 20.45%. The error of commission was found to be 12.82% in MD,
which was higher in comparison with DD class (5.41%). The errors of omission for S/G, NF
and W were found to be 31.58%, 24.42%, and 5.26%, respectively whereas the error of
commission was found to be 23.53%, 32.43%, and 10%, respectively. The present classi­
fication focused on the forests present in the NWH foothills, so to compare it with MODIS
PFT (MCD12Q1), the classes defined in MODIS PFT for the study area were merged
accordingly to make both the maps comparable and accuracy assessment was carried
INTERNATIONAL JOURNAL OF REMOTE SENSING 9

Table 1. Error matrix and accuracy assessment of the classification.


Reference data
Class MD DD S/G NF W Total User’s accuracy (%)
MD 34 5 0 0 0 39 87.18
DD 0 35 0 2 0 37 94.59
S/G 0 0 13 4 0 17 76.47
NF 1 4 6 25 1 37 67.57
W 0 0 0 2 18 20 90.00
Total 35 44 19 33 19 150
Producer’s accuracy (%) 97.14 79.55 68.42 75.76 94.74
MD: Moist deciduous; DD: Dry deciduous; S/G: Shrubs/Grasses; NF: Non forest; W: Water
Overall classification accuracy = 83.33%; Kappa coefficient = 0.71

out. For the present study area, MODIS PFT showed an overall accuracy of 66% and the
Kappa coefficient was found to be 0.51.

4. Discussion
The Himalayan ecosystems are vulnerable to the changing climate and increasing pres­
sure, hence, the characterization of the diversity in functional types is essential to
represent the PFTs more accurately in the ecosystem models. The forest present in the
study area represents a unique amalgamation of variation in functional aspects including
phenology. Hence, it is required to develop reliable approaches to classify PFTs, which are
more representative of the region. MODIS PFT is widely accepted but some studies have
reported lower accuracy of this product in comparison to other methods of PFT classifica­
tion at regional levels (Sun et al. 2008; Cai et al. 2014). In earlier attempts to map PFTs at
the regional level, Sun et al. (2008) used multisource evidential reasoning and RS techni­
ques to produce PFT maps. Using this method, the accuracy of the PFT maps for four
U.S. states (57.61%, 54.95%, 63.65%, and 46.60%) were found to be better compared to
MODIS PFT (55.41%, 48.97%, 51.07%, and 44.19%). Though the accuracy of these maps
improved, this method could not outline a classification procedure for mapping over
larger geographic areas. Cai et al. (2014) used hierarchical soft time-space classification to
produce PFT maps at the regional level using MODIS NDVI and topographical information
from SRTM DEM. The overall accuracy of this PFT map was found to be 64.42%, which was
higher in comparison to the MODIS PFT map (57.68% overall accuracy). Padalia et al.
(2019) mapped the dominant PFTs in the alpine regions of western Himalaya using SVM
classification of Landsat 8 images with an overall accuracy of 82.71%.
The present study tried to address the variables, which can affect the variation of PFTs
in the study area. The major forests present in the study area differ phenologically,
therefore, harmonic analysis of NDVI time-series was carried out to highlight this variation.
Running et al. (1995) suggested that PFTs should be characterized by a variety of
variables, which include vegetation structure, phenology, and environmental conditions.
Sun and Liang (2008) also mentioned that site-specific environmental and ecological
setup such as climate and terrain are among the most important factors determining
the geographic distribution of PFTs. To represent the site-specific ecological conditions,
aspect, MTPI and Bioclim variables were used in this study. Aspect is an important
topographic factor, which can affect the forests through its direct influence on radiation
and moisture (Måren et al. 2015). The opposing slopes vary in their microclimate,
10 R. SRINET ET AL.

including light intensity, soil and air temperature, humidity, soil moisture and evapora­
tion, and duration of growing periods, which is closely associated with vegetation
structure and functioning (Holland and Steyn 1975).
In the present study, we used different well-established, freely-available satellite data
and products in the GEE platform. GEE has led to a significant reduction in the processing
time and overcoming the limitation of downloading, processing and storing a large
amount of satellite data for regional and global studies (Gorelick et al. 2017). Harmonic
analysis of the long-term NDVI dataset was carried out prior to classification. RF,
a machine-learning algorithm was applied in GEE platform to obtain a PFT map using
the relevant datasets. RF is known to be insensitive to noise and over-fitting (Gislason,
Benediktsson, and Sveinsson 2006; Belgiu and Drăguţ 2016), can handle high data
dimensionality and can achieve higher classification accuracies (Rodriguez-Galiano et al.
2012; Teluguntla et al. 2018; Tsai et al. 2018). A set of 500 spatially well-distributed points,
collected considering the spatial feature of the landscape, was used to train the RF
classifier. The PFT map, thus produced, was validated with ground truth observations.
The error of omission was found to be higher in the DD type compared to MD. On
analysing the error matrix, it can be stated that while 97.14% of MD pixels were correctly
identified, however, only 87.18% of the area under this class was actually MD type. On the
other hand, only 79.55% of DD pixels were correctly identified whereas 94.59% of the area
under this class was DD type.
The present study is mainly focussed on the differentiation between the major PFTs
present in the NWH foothills of India. The MD type present in the NWH foothills of India
are dominated by sal (Shorea robusta). The phenology of sal has been much discussed due
to its evergreen to deciduous nature (Kushwaha and Nandy 2012). Sal forests present in
some patches of the study area display deciduous behaviour during the dry season (April-
May) whereas in some places it shows semi-evergreen to evergreen behaviour, which can
lead to misclassification. On the other hand, the forest present in the DD type experiences
leaflessness during dry and hot weather conditions (March – May) and the soil is exposed
(Champion and Seth 1968), which may lead to the mixing of some DD pixels with NF class.
On closer comparison with MODIS PFT, it was observed that the majority of area under
MD PFT has been classified as evergreen broadleaf trees whereas in some areas it is
defined as deciduous broadleaf trees. The MD and DD types vary greatly in their pheno­
logical behaviour and functioning. Therefore, the characterization of these forests in the
ecosystem models based on the MODIS PFT may lead to errors in predictions of carbon,
water, and energy fluxes at regional levels. To overcome this problem, the seasonality
parameters were considered, which was able to present a contrast between the MD and
DD PFTs based on the amplitude and phase values obtained from the harmonic analysis
of time-series NDVI.

5. Conclusions
PFTs are considered as an important parameter to understand the climate change feed­
back to the plant community. The mapping of PFTs using RS data is an evolving field of
research. Accurate mapping of PFTs at the regional level can contribute to the better
ability of ecosystem models to simulate biogeochemical cycles and processes at regional
to global scales. Using the global PFT maps at regional levels without testing the
INTERNATIONAL JOURNAL OF REMOTE SENSING 11

accuracies for simulations by ecosystem models may give rise to an error in the char­
acterization of the vegetation, and hence the error in modelling. Thus, the present study
attempted to generate the PFT map of NWH foothills of India – one of the vulnerable areas
due to climate change. It provides an efficient approach for PFT classification, which uses
different freely available satellite data and products in the GEE platform. Here, RF classifier
was used to generate the PFT map using seasonality parameters derived from NDVI time-
series, topographic and climatic variables. The study revealed that seasonality parameters
in combination with topographic and climatic variables presented a good contrast
between different PFTs in the study area. The accuracy of classification using this method
was found to be higher than the freely available MODIS PFT product. The study provides
a methodology to determine the geographic distribution of PFTs by considering major
driving factors of PFT distribution, which can be implemented in GEE using various
meaningful variables and at varying resolutions in different ecological settings. It also
highlighted the utility of the GEE platform to process and analyse a large volume of long-
term time-series data to get accurate PFT map at a regional level.

Acknowledgements
The present study was carried out as a part of Soil-Vegetation-Atmosphere-Flux (SVAF) of National
Carbon Project (NCP) supported by ISRO-Geosphere-Biosphere Programme. Authors are thankful to
the MODIS Science team for the Science Algorithms, the Processing Team for producing MODIS
data, and the GES DAAC MODIS Data Support Team for making MODIS data available to the user
community. The authors would also like to thank the Google Earth Engine team for providing cloud-
computing resources. Thanks are also due to the anonymous reviewers for their valuable sugges­
tions, which helped us to improve the manuscript.

Disclosure statement
No potential conflict of interest was reported by the authors.

Funding
This work was supported by the Indian Space Research Organisation.

ORCID
Subrata Nandy http://orcid.org/0000-0003-4127-4035
Surajit Ghosh http://orcid.org/0000-0002-3928-2135

References
Acharya, P., C. Biradar, M. Louhaichi, S. Ghosh, S. Hassan, H. Moyo, and A. Sarker. 2019. “Finding
a Suitable Niche for Cultivating Cactus Pear (Opuntia Ficus-indica) as an Integrated Crop in
Resilient Dryland Agroecosystems of India.” Sustainability 11 (21): 5897. doi:10.3390/
su11215897.
Azzari, G., and D. B. Lobell. 2017. “Landsat-based Classification in the Cloud: An Opportunity for
a Paradigm Shift in Land Cover Monitoring.” Remote Sensing of Environment 202: 64–74.
doi:10.1016/j.rse.2017.05.025.
12 R. SRINET ET AL.

Belgiu, M., and L. Drăguţ. 2016. “Random Forest in Remote Sensing: A Review of Applications and
Future Directions.” ISPRS Journal of Photogrammetry and Remote Sensing 114: 24–31. doi:10.1016/
j.isprsjprs.2016.01.011.
Bonan, G. B., S. Levis, L. Kergoat, and K. W. Oleson. 2002. “Landscapes as Patches of Plant Functional
Types: An Integrating Concept for Climate and Ecosystem Models.” Global Biogeochemical Cycles
16 (2): 5–1. doi:10.1029/2000GB001360.
Breiman, L. 2001. “Random Forests.” Machine Learning 45: 5–32. doi:10.1023/A:1010933404324.
Cai, D., Y. Guan, S. Guo, C. Zhang, and K. Fraedrich. 2014. “Mapping Plant Functional Types over
Broad Mountainous Regions: A Hierarchical Soft Time-space Classification Applied to the Tibetan
Plateau.” Remote Sensing 6 (4): 3511–3532. doi:10.3390/rs6043511.
Campos-Taberner, M., Á. Moreno-Martínez, F. García-Haro, G. Camps-Valls, N. Robinson, J. Kattge,
and S. Running. 2018. “Global Estimation of Biophysical Variables from Google Earth Engine
Platform.” Remote Sensing 10 (8): 1167. doi:10.3390/rs10081167.
Canadell, J. G., H. A. Mooney, D. D. Baldocchi, J. A. Berry, J. R. Ehleringer, C. B. Field, S. T. Gower, et al.
2000. “Commentary: Carbon Metabolism of the Terrestrial Biosphere: A Multitechnique Approach
for Improved Understanding.” Ecosystems 3 (2): 115–130. doi:10.1007/s100210000014.
Castanho, A. D. A., M. T. Coe, M. H. Costa, Y. Malhi, D. Galbraith, and C. A. Quesada. 2013. “Improving
Simulated Amazon Forest Biomass and Productivity by Including Spatial Variation in Biophysical
Parameters.” Biogeosciences 10 (4): 2255–2272. doi:10.5194/bg-10-2255-2013.
Champion, H. G., and S. K. Seth. 1968. A Revised Forest Types of India. Delhi: Manager of Publications,
Government of India.
Dallmeyer, A., M. Claussen, and V. Brovkin. 2019. “Harmonising Plant Functional Type
Distributions for Evaluating Earth System Models.” Climate of the Past 15: 335–366.
doi:10.5194/cp-2018-41.
Fatichi, S., C. Pappas, J. Zscheischler, and S. Leuzinger. 2019. “Modelling Carbon Sources and Sinks in
Terrestrial Vegetation.” New Phytologist 221 (2): 652–668. doi:10.1111/nph.15451.
Fick, S. E., and R. J. Hijmans. 2017. “WorldClim 2: New 1km Spatial Resolution Climate Surfaces
for Global Land Areas.” International Journal of Climatology 37 (12): 4302–4315. doi:10.1002/
joc.5086.
Friedl, M. A., D. Sulla-Menashe, B. Tan, A. Schneider, N. Ramankutty, A. Sibley, and X. Huang. 2010.
“MODIS Collection 5 Global Land Cover: Algorithm Refinements and Characterization of New
Datasets.” Remote Sensing of Environment 114 (1): 168–182. doi:10.1016/j.rse.2009.08.016.
Gislason, P. O., J. A. Benediktsson, and J. R. Sveinsson. 2006. “Random Forests for Land Cover
Classification.” Pattern Recognition Letters 27 (4): 294–300. doi:10.1016/j.patrec.2005.08.011.
Gorelick, N., M. Hancher, M. Dixon, S. Ilyushchenko, D. Thau, and R. Moore. 2017. “Google Earth
Engine: Planetary-scale Geospatial Analysis for Everyone.” Remote Sensing of Environment 202:
18–27. doi:10.1016/j.rse.2017.06.031.
Guisan, A., S. B. Weiss, and A. D. Weiss. 1999. “GLM versus CCA Spatial Modeling of Plant Species
Distribution.” Plant Ecology 143: 107–122. doi:10.1023/A:100984151.
Holland, P. G., and D. G. Steyn. 1975. “Vegetational Responses to Latitudinal Variations in Slope
Angle and Aspect.” Journal of Biogeography 2 (3): 179–183. doi:10.2307/3037989.
Johansen, K., S. Phinn, and M. Taylor. 2015. “Mapping Woody Vegetation Clearing in Queensland,
Australia from Landsat Imagery Using the Google Earth Engine.” Remote Sensing Applications:
Society and Environment 1: 36–49. doi:10.1016/j.rsase.2015.06.002.
Kushwaha, S. P. S., and S. Nandy. 2012. “Species Diversity and Community Structure in Sal (Shorea
Robusta) Forests of Two Different Rainfall Regimes in West Bengal, India.” Biodiversity and
Conservation 21 (5): 1215–1228. doi:10.1007/s10531-012-0264-8.
Lapola, D. M., M. D. Oyama, C. A. Nobre, and G. Sampaio. 2008. “A New World Natural Vegetation
Map for Global Change Studies.” Anais Da Academia Brasileira De Ciências 80 (2): 397–408.
doi:10.1590/S0001-37652008000200017.
Mahdianpari, M., B. Salehi, F. Mohammadimanesh, S. Homayouni, and E. Gill. 2019. “The First
Wetland Inventory Map of Newfoundland at a Spatial Resolution of 10 M Using Sentinel-1 and
Sentinel-2 Data on the Google Earth Engine Cloud Computing Platform.” Remote Sensing 11 (1):
43. doi:10.3390/rs11010043.
INTERNATIONAL JOURNAL OF REMOTE SENSING 13

Måren, I. E., S. Karki, C. Prajapati, R. K. Yadav, and B. B. Shrestha. 2015. “Facing North or South: Does
Slope Aspect Impact Forest Stand Characteristics and Soil Properties in a Semiarid trans-Himalayan
Valley?” Journal of Arid Environment 121: 112–123. doi:10.1016/j.jaridenv.2015.06.004.
Nandy, S., S. Ghosh, S. P. S. Kushwaha, and A. S. Kumar. 2019. “Remote Sensing-based Forest
Biomass Assessment in Northwest Himalayan Landscape.” In Remote Sensing of Northwest
Himalayan Ecosystems, edited by R. R. Navalgund, A. S. Kumar, and S. Nandy, 285–311.
Singapore: Springer.
Navalgund, R. R., A. S. Kumar, and S. Nandy. 2019. Remote Sensing of Northwest Himalayan
Ecosystems. Singapore: Springer.
Ottlé, C., J. Lescure, F. Maignan, B. Poulter, T. Wang, and N. Delbart. 2013. “Use of Various Remote
Sensing Land Cover Products for Plant Functional Type Mapping over Siberia.” Earth System
Science Data 5 (2): 331–348. doi:10.5194/essd-5-331-2013.
Padalia, H., A. Bushra, G. Singh, S. Nandy, I. D. Rai, and G. S. Rawat. 2019. “Spatial Patterns of Plant
Functional Types and Environmental Proxies of Plant Richness in Alpine Region of Western
Himalaya, India.” Biodiversity and Conservation 28 (8–9): 2221–2244. doi:10.1007/s10531-018-
1664-1.
Patel, N. R., H. Padalia, S. P. S. Kushwaha, S. Nandy, T. Watham, J. Ahongshangbam, R. Kumar,
V. K. Dadhwal, and A. S. Kumar. 2019. “CO2 Flux Tower and Remote Sensing: Tools for Monitoring
Carbon Exchange over Ecosystem Scale in Northwest Himalaya.” In Remote Sensing of Northwest
Himalayan Ecosystems, edited by R. R. Navalgund, A. S. Kumar, and S. Nandy, 313–327. Singapore:
Springer.
Pillai, N. D., S. Nandy, N. R. Patel, R. Srinet, T. Watham, and P. Chauhan. 2019. “Integration of Eddy
Covariance and Process-based Model for the Intra-annual Variability of Carbon Fluxes in an Indian
Tropical Forest.” Biodiversity and Conservation 28 (8–9): 2123–2141. doi:10.1007/s10531-019-
01770-3.
Poulter, B., P. Ciais, E. Hodson, H. Lischke, F. Maignan, S. Plummer, and N. E. Zimmermann. 2011.
“Plant Functional Type Mapping for Earth System Models.” Geoscientific Model Development 4 (4):
993–1010. doi:10.5194/gmd-4-993-2011.
Poulter, B., N. MacBean, A. Hartley, I. Khlystova, O. Arino, R. Betts, S. Bontemps, et al. 2015. “Plant
Functional Type Classification for Earth System Models: Results from the European Space
Agency’s Land Cover Climate Change Initiative.” Geoscientific Model Development 8: 2315–2328.
doi:10.5194/gmd-8-2315-2015.
Rayner, J. N. 1971. An Introduction to Spectral Analysis. London: Pion.
Rodriguez-Galiano, V. F., B. Ghimire, J. Rogan, M. Chica-Olmo, and J. P. Rigol-Sanchez. 2012. “An
Assessment of the Effectiveness of a Random Forest Classifier for Land-cover Classification.” ISPRS
Journal of Photogrammetry and Remote Sensing 67: 93–104. doi:10.1016/j.isprsjprs.2011.11.002.
Running, S. W., T. R. Loveland, L. L. Pierce, R. R. Nemani, and E. R. Hunt Jr. 1995. “A Remote Sensing
Based Vegetation Classification Logic for Global Land Cover Analysis.” Remote Sensing of
Environment 51 (1): 39–48. doi:10.1016/0034-4257(94)00063-S.
Sellers, P. J., D. A. Randall, G. J. Collatz, J. A. Berry, C. B. Field, D. A. Dazlich, C. Zhang, et al. 1996.
“A Revised Land Surface Parameterization (Sib2) for Atmospheric GCMs. Part I: Model Formulation.”
Journal of Climate 9 (4): 676–705. doi:10.1175/1520-0442(1996)009<0676:ARLSPF>2.0.CO;2.
Semenova, G. V., and E. van der Maarel. 2000. “Plant Functional Types–a Strategic Perspective.”
Journal of Vegetation Science 11 (6): 917–922. doi:10.2307/3236562.
Smith, T. M., T. M. Smith, H. H. Shugart, and F. I. Woodward. 1997. Plant Functional Types: Their
Relevance to Ecosystem Properties and Global Change. Vol. 1. Cambridge: Cambridge University
Press.
Srinet, R., S. Nandy, and N. R. Patel. 2019. “Estimating Leaf Area Index and Light Extinction Coefficient
Using Random Forest Regression Algorithm in a Tropical Moist Deciduous Forest, India.”
Ecological Informatics 52: 94–102. doi:10.1016/j.ecoinf.2019.05.008.
Sun, W., and S. Liang. 2008. “Methodologies for Mapping Plant Functional Types.” In Advances in
Land Remote Sensing, edited by S. Liang, 369–393. Dordrecht: Springer.
14 R. SRINET ET AL.

Sun, W., S. Liang, G. Xu, H. Fang, and R. Dickinson. 2008. “Mapping Plant Functional Types from
MODIS Data Using Multisource Evidential Reasoning.” Remote Sensing of Environment 112 (3):
1010–1024. doi:10.1016/j.rse.2007.07.022.
Teluguntla, P., P. Thenkabail, A. Oliphant, J. Xiong, M. K. Gumma, R. G. Congalton, K. Yadav, et al.
2018. “A 30-m Landsat-derived Cropland Extent Product of Australia and China Using Random
Forest Machine Learning Algorithm on Google Earth Engine Cloud Computing Platform.” ISPRS
Journal of Photogrammetry and Remote Sensing 144: 325–340. doi:10.1016/j.isprsjprs.2018.07.017.
Theobald, D. M., D. Harrison-Atlas, W. B. Monahan, and C. M. Albano. 2015. “Ecologically-relevant
Maps of Landforms and Physiographic Diversity for Climate Adaptation Planning.” PloS One 10
(12): e0143619. doi:10.1371/journal.pone.0143619.
Tsai, Y., D. Stow, H. Chen, R. Lewison, L. An, and L. Shi. 2018. “Mapping Vegetation and Land Use
Types in Fanjingshan National Nature Reserve Using Google Earth Engine.” Remote Sensing 10 (6):
927. doi:10.3390/rs10060927.
Verant, S., K. Laval, J. Polcher, and M. De Castro. 2004. “Sensitivity of the Continental Hydrological
Cycle to the Spatial Resolution over the Iberian Peninsula.” Journal of Hydrometeorology 5 (2):
267–285. doi:10.1175/1525-7541(2004)005<0267:SOTCHC>2.0.CO;2.
Walker, B. H. 1992. “Landscape to Regional-scale Responses of Terrestrial Ecosystems to Global
Change.” Abstract from IGBP symposium, Stockholm, reprinted in IGBP Newsletter 13, 18.
Whittaker, R. H. 1956. “Vegetation of the Great Smoky Mountains.” Ecological Monographs 26 (1):
1–80. doi:10.2307/1943577.
Xia, J., J. Chen, S. Piao, P. Ciais, Y. Luo, and S. Wan. 2014. “Terrestrial Carbon Cycle Affected by
Non-uniform Climate Warming.” Nature Geoscience 7 (3): 173. doi:10.1038/ngeo2093.

You might also like