For Peer Review Only
Neural Networks as a Tool for Constructing Continuous NDVI Time Series
from AVHRR and MODIS
Journal: International Journal of Remote Sensing
Manuscript ID: TRES-PAP-2007-0740.R1
Manuscript Type: Research Paper
Date Submitted by the
Author: 14-May-2008
Complete List of Authors: Brown, Molly; SSAI, NASA-Goddard Space Flight Center
Lary, David; University of Maryland Baltimore College, GEST
Vrieling, Anton; Joint Research Centre of the European Commission
Stathakis, Demetris; Joint Research Centre of the European
Commission
Mussa, Hamse; University of Cambridge
Keywords: AVHRR, MODIS, NDVI, NEURAL NETWORKS
Keywords (user defined): AVHRR, MODIS, NDVI
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing

For Peer Review Only
1
Neural Networks as a Tool for Constructing Continuous NDVI Time Series
from AVHRR and MODIS
Molly E. Brown1
David J. Lary2
Anton Vrieling3
Demetris Stathakis3
Hamse Mussa4
1 Science Systems and Applications, Inc., NASA Goddard Space Flight Center, MD, USA
Ph: 301-614-6616, Fax: 301-614-6015
Email: molly.brown@gsfc.nasa.gov
2 UMBC GEST, NASA Goddard Space Flight C, MD, USA
3 Joint Research Centre of the European Commission, Ispra (VA), Italy
4 Department of Chemistry, University of Cambridge, England
Revised, for International Journal of Remote Sensing
Page 1 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only
2
Abstract
The long term AVHRR-NDVI record provides a critical historical perspective on vegetation
dynamics necessary for global change research. Despite the proliferation of new sources of global,
moderate resolution vegetation datasets, the remote sensing community is still struggling to create
datasets derived from multiple sensors that allow the simultaneous use of spectral vegetation for
time series analysis. To overcome the non-stationary aspect of NDVI, we use an artificial neural
network (ANN) to map the NDVI indices from AVHRR to those from MODIS using atmospheric,
surface type and sensor-specific inputs to account for the differences between the sensors. The
NDVI dynamics and range of MODIS NDVI data at one degree is matched and extended through
the AVHRR record. Four years of overlap between the two sensors is used to train a neural network
to remove atmospheric and sensor specific effects on the AVHRR NDVI. In this paper, we present
the resulting continuous dataset, its relationship to MODIS data, and a validation of the product.
Keywords: Normalized difference vegetation index (NDVI), MODIS, AVHRR, Neural Networks
Page 2 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only
3
1.0 Introduction
Consistent, long term vegetation data records are critical for analysis of the impact of global change
on terrestrial ecosystems. Continuous observations of terrestrial ecosystems through time are
necessary to document changes in magnitude or variability in an ecosystem (Eklundh and Olsson,
2003; Slayback et al., 2003; Tucker et al., 2001). Satellite remote sensing has been the primary tool
for scientists to measure global trends in vegetation, as the measurements are both global and
temporally frequent. To extend measurements through time, multiple sensors with different design
and resolution must be used tog ther in the same time series. This presents significant problems as
sensor band placement, spectral response, processing, and atmospheric correction of the
observations can vary significantly and impact the comparability of the measurements (Brown et
al., 2006). Even without differences in atmospheric correction, vegetation index values for the
same target recorded under identical conditions will not be directly comparable because input
reflectance values differ from sensor to sensor due to differences in sensor design and spectral
response of the instrument (Miura et al., 2006; Teillet et al., 1997).
Several approaches have been taken to integrate data from multiple sensors. Steven et al. (2003),
for example, simulated the spectral response from multiple instruments and with simple linear
equations created conversion coefficients to transform NDVI data from one sensor to another.
Their analysis is based on the observation that the vegetation index is critically dependent on the
spectral response functions of the instrument used to calculate it. The conversion formulas the
paper presents cannot be applied to maximum value NDVI datasets because the weighting
coefficients are land cover and dataset dependent, reducing their efficacy in mixed pixel situations
Page 3 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only
4
(Steven et al., 2003). Trishchenko et al. (2002) created a series of quadratic functions to correct
for differences in the reflectance and NDVI to NOAA-9 AVHRR-equivalents (Trishchenko et al.,
2002). Both the Steven et al. (2003) and the Trishchenko et al. (2002) approaches are land cover
and dataset dependent and thus cannot be used on global datasets where multiple land covers are
represented by one pixel. Miura et al (2006) used hyper-spectral data to investigate the effect of
different spectral response characteristics between MODIS and AVHRR instruments on both the
reflectance and NDVI data, showing that the precise characteristics of the spectral response had a
large effect on the resulting vegetation index. The complex patterns and dependencies on spectral
band functions were both land cover dependent and strongly non-linear, thus we see that an
exploration of a non-linear approach may be fruitful.
In this paper we experiment with powerful, non-linear neural networks to identify and remove
differences in sensor design and variable atmospheric contamination from the AVHRR NDVI
record in order to match the range and variance of MODIS NDVI without removing the desired
signal representing the underlying vegetation dynamics. Neural networks are ‘data transformers’
(Atkinson and Tatnall, 1997), where the objective is to associate the elements of one set of data to
the elements in another. Relationships between the two datasets can be complex and the two
datasets may have different statistical distributions. In addition, neural networks incorporate a
priori knowledge and realistic physical constraints into the analysis, enabling a transformation from
one dataset into another through a set of weighting functions (Atkinson and Tatnall, 1997). This
transformation incorporates additional input data that may account for differences between the two
datasets.
Page 4 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only
5
Our objective in this paper is to demonstrate the viability of neural networks as a tool to produce a
long term dataset based on AVHRR NDVI that has the data range and statistical distribution of
MODIS NDVI. Previous work has shown that the relationship between AVHRR and MODIS
NDVI is complex and nonlinear (Brown et al., 2006; Gallo et al., 2003; Miura et al., 2006), thus
this problem is well suited to neural networks if appropriate inputs can be found. The impact of
atmospheric contamination, such as clouds, smoke, pollution and other aerosols, variations in soil
color and exposure through vegetation, and land cover type has a differential effect on AVHRR
data as compared to MODIS data. Here we explore how neural networks can be used to account
for these impacts and create an AVHRR NDVI dataset with similar characteristics as the MODIS
dataset. Overlapping years of observations are used to train the network. Examination of the
resulting MODIS-fitted AVHRR dataset both during the overlap period and in the historical dataset
enabled an evaluation of the efficacy of the neural net approach compared to other approaches to
merge multiple-sensor NDVI datasets.
2.0 Neural Networks
Neural networks are algorithms used for either classification or function approximation (Lippmann,
1987). A good introduction on neural networks is given by Lippmann (1987). Since their first
introduction, they have been used for almost two decades in remote sensing (Benediktsson et al.,
1990). The most commonly used type of neural network is the Multi-Layer Perceptron, of which
Kalman filters are one type. Artificial neural networks (ANN) are made up of input layers, hidden
layers and output layers.
Page 5 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only
6
The MLP neural network has an input layer where the data samples are fed, typically after being
normalized. The data from the input layer is then fed into a number of hidden layers, typically
either one or two. The choice of how many hidden layers and number of nodes per hidden layer
that should be used is currently an open research question (Stathakis, 2008). Several heuristics exist
to assist in selecting the number of nodes in the hidden layers, some of which developed explicitly
in the domain of remote sensing such as the Kanellopoulos – Wilkinson (1997) rule (Stathakis and
Vasilakos, 2006). Finally the hidden layers feed one or more input layers.
To summarize the ANN topology, a relation of x:y:z is frequently used. This implies a neural
network with x input nodes, one hidden layer with y hidden nodes and z output nodes (for example,
7:20:1). The neural network is trained by adjusting the values of the connections, called weights,
between nodes. The most commonly used training algorithm is back-propagation introduced by
Rumelhart et al. (1986). Several modifications to the original algorithm have greatly boosted
performance (Rumelhart et al., 1986). Neural networks can learn in an either supervised or
unsupervised mode depending on whether target vectors are presented along with input vectors or
not. In the supervised mode, several spectral bands (or in this study, time series) per data sample
are typically presented to the network. At the same time the desired output is also used to modify
the weights so that the deviation between actual and obtained output is minimized. Typically the
samples available, i.e. input and output vectors, are split in order to train the network and
independently validate the results. A three-set strategy has been proposed to offer a more objective
validation by Bishop (1995). According to this strategy three subsets are created, one of training,
one for validation and on for testing (Bishop, 1995).
Page 6 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only
7
One of the main advantages of neural networks is the fact that multiple sources, including non-
spectral, data can be used as input (Benediktsson et al., 1990; Stathakis and Kanellopoulos, 2008).
This is because neural networks make no assumptions, e.g. about statistical distributions, regarding
the input data. One of their main drawbacks is that they require experience in selecting values for
the numerous parameters that need to be set. Recent results show that global search methods can be
used to make near-optimal choices (Stathakis, 2008). Additionally, neural networks are often
accused of being black-box techniques because the knowledge learned can not be expressed in a
meaningful way. Several efforts have been made towards building transparent neural networks.
One way to do this is to deploy neuro-fuzzy methods (Stathakis and Vasilakos, 2006).
3.0 Data
This study uses global NDVI products derived from AVHRR and MODIS NDVI sensors at one
degree resolution and for a monthly time window. Ancillary files are used in this study to
determine the impact of clouds and other atmospheric effects on the vegetation measurement from
different sensors through time. We have restricted the number of inputs to six besides the AVHRR
NDVI to reduce redundancy and over-fitting of the neural network. These are three atmospheric
products from TOMS, a soil type map, a digital elevation model (DEM), and a land cover map.
3.1 NDVI datasets at one degree
AVHRR and MODIS NDVI products were downsampled to one degree resolution to reduce
processing time of the artificial neural network and to match the resolution of the atmospheric
Page 7 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only
8
TOMS inputs. To further reduce processing time, average monthly composites were made of the
two products. The spatial and temporal downsampling was done by averaging all pixels falling in a
one-degree cell for the two nearest periods in a month (MODIS products do not respect month
limits).
The maximum value AVHRR NDVI composites have an 8-km resolution (Holben, 1986; Tucker,
1979) and were from the NASA Global Inventory Monitoring and Modeling Systems (GIMMS)
group at the Laboratory for Terrestrial Physics (Brown et al., 2006; Tucker et al., 2005) from July
1981 to May 2004. A post-processing satellite drift correction has been applied to this dataset to
further remove artifacts due to orbital drift and changes in the sun-target-sensor geometry (Pinzon
et al., 2005). As a result of AVHRR's wide spectral bands, the AVHRR NDVI is more sensitive to
water vapor in the atmosphere than MODIS. An increase in water vapor results in a lower NDVI
signal, which can be interpreted as an actual change if no correction is applied (Pinheiro et al.,
2004; Pinzon, 2002). The maximum value composite should lessen these artifacts (Holben, 1986).
The GIMMS operational dataset incorporates AVHRR data from sensors aboard NOAA-7 through
14 with the data from the AVHRR on NOAA-16 and 17.
The Terra-MODIS 16 day L3 land surface NDVI product was selected. NDVI data for MODIS
was computed from the (White-Sky) Filled Land Surface Albedo Map Product, which is a value-
added product from the MODIS Atmospheres group. The global, one kilometer, 16 day MODIS
NDVI composites from February 2000 to December 2004 were used to create averaged one degree
monthly data for this analysis. The resulting one degree time series include only pixels with more
Page 8 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only
9
than 50% land and conforms to the ISCLSCP convention as described by Sellers et al. (1996).
(Sellers et al., 1996).
3.2 Ancillary datasets
To account for the differences between the AVHRR and MODIS data, we use four ancillary data
products in the neural network: TOMS Data which provides information on water vapor in the
atmosphere, soil maps, land cover maps and elevation. Each of these accounts for an aspect of the
sensor design differences and provide key information so that the neural network can work.
Preliminary work (not described here) demonstrated that the most important factors controlling the
relationship between the NDVI of MODIS and that of AVHRR are the surface reflectance, the land
surface type, aerosols and total ozone column. Variations in atmospheric contamination have direct
impact on the AVHRR NDVI used here because no atmospheric correction was implemented
during its processing, only volcanic aerosols and maximum value compositing (Tucker et al.,
2005). We know that ozone is a key atmospheric absorber of light in the visible region, and water,
as measured by aerosols, in the infrared. The AVHRR NDVI, calculated using the wide bands of
the instrument, will therefore be influenced by these elements.
The Nimbus-7 TOMS data is the only source of high resolution global information about the
atmospheric composition (and hence depression of AVHRR NDVI) for much of the AVHRR
record. As an instrument that measures the atmosphere back to 1981, TOMS has the advantage of
being co-located for much of its record on the same platform as AVHRR, which is particularly
important as the NOAA satellites from which the AVHRR NDVI are derived are subject to non-
linear orbital drift through time (McPeters et al., 1998). The TOMS data is from Version 8, includes
reflectance, aerosols and ozone measurements and is derived from three sensors: Nimbus 7, Meteor
Page 9 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only
10
3 and Earth Probe (Table 1). All three products are used in order to capture the impact of
atmospheric variations on the uncorrected AVHRR NDVI data. During the missing period of
1994-96, we use a climatology created by taking the median value of the preceding 2, 4, and 6
years and the following 2, 4, and 6 years. This approach was used as ozone has a quasi-biennial
oscillation (QBO). Although not optimal, this performed well and is required if we want to use
these datasets for a correction of the entire series.
The NASA Goddard Institute for Space Studies (GISS) soil type map is used to account for the
difference in sensitivity to underlying soil color from AVHRR and MODIS (Huete et al., 1994;
Huete and Tucker, 1991). The soil type map is at one degree resolution and contains 26 soil units,
and values for water and ice. The soil type data file was derived from the highest level of the FAO
soil units and is based on the work of Zobler (1986).
A one degree ‘surface type’ land cover dataset was created from the SPOT Global Land Cover
(GLC) 2000 dataset (Giri et al., 2004). Previous research has shown that variations in land cover
affect the strength of the impact of atmospheric thickness (Pinzon, 2002). This dataset has 22 land
cover classes based on the FAO land cover classification system. We aggregated the data to a one-
degree resolution using a vote procedure. We used the GLC2000 data instead of MODIS or
AVHRR-based land cover datasets as an independent surface classification for the ANN training.
We use a single land cover map to represent the land cover for the 25 year record. Even though we
acknowledge that land cover change may have occurred during this period, they are unlikely to
span an entire one by one degree pixel. The neural network uses this parameter to identify regions
Page 10 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only
11
with very low signal due to small amounts of vegetation. These regions are approximately static
through time globally.
A one degree DEM was used to ensure the identification and maintenance of mountainous regions
that may otherwise be confused with clouds or other atmospheric effects. This DEM was derived
from the USGS SRTM 90-m dataset, and has been aggregated to one degree using averaging.
3.3 Global Rainfall Data
We used Global Precipitation Climatology Centre (GPCC) rain gauge data from the Global
Precipitation Climatology Project (GPCP). These data were used to evaluate the ability of the
NDVI data products for capturing interannual vegetation dynamics related to rainfall. The GPCC
data are area-averaged and time-integrated precipitation fields based on surface rain gauge
measurements. The GPCC collects monthly precipitation totals received from the World Weather
Watch GTS (Global Telecommunication System) of the World Meteorological Organization
(WMO). The GPCC acquires monthly precipitation data from international/national meteorological
and hydrological services/institutions. Surface rain-gauge based monthly precipitation data from
6700 meteorological stations are analyzed over land areas and gridded using a spatial objective
analysis method (Rudolf et al., 1994).
4.0 Methods
4.1 Application of the ANN
Page 11 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only
12
When mapping AVHRR to MODIS NDVI using ANNs, factors that explain differences in the
sensors and their processing must be accounted for by the input variables. Here we use historical
data derived from the total ozone mapping spectrometer or TOMS, which is available with some
interruption back to 1978 (McPeters et al., 1998). The AVHRR is also more sensitive to
differences in background soil contamination than MODIS (Huete and Jackson, 1988), thus we use
a soil type map (Zobler, 1986), a DEM, and a land cover map to account for these differences (see
section 3 for a description of the datasets).
The neural network used here is a fully-connected feed-forward Multi-Layer Perceptron with
7:20:1 topology. Biases are connected to both hidden and output layers. The selection of the nodes
in the hidden topology conforms well to the Kanellopoulos – Wilkinson rule commonly used in
remote sensing. In this study we employed a feed-forward ANN with 20 nodes in a single hidden
layer using a Kalman filter training algorithm. The Kalman filter algorithm provides rapid
convergence for the weight estimation and is described by Lary and Mussa, (2004).
Besides the additional data sources, the neural net is trained with time-series data of AVHRR and
MODIS from the overlapping period of 2000-2003. Subsequently, the resulting weighting functions
were applied to the AVHRR data from 1982-2003, using the ancillary files. The functions enable
the correction of the entire dataset, enabling the production of an AVHRR dataset with similar
characteristics as the MODIS dataset. For simplicity, throughout this paper this new dataset will be
referred to as NNndvi, or the neural net corrected AVHRR NDVI. The result is an experimental
product, whose objective is to demonstrate how a seamless AVHRR to MODIS dataset may be
created. We do not assume that the method used is the only possible or even the most optimal
Page 12 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only
13
method, but one that can produce a far closer integration between the datasets than has been
demonstrated before using the actual processed data instead of modeled data. For this feasibility
demonstration we operated on the one degree scale at a monthly resolution to reduce processing
time of the neural net. The same training procedure could be conducted at a higher temporal and
spatial resolution with more computing time and/or for smaller areas.
4.2 Evaluation Methods
The obtained NNndvi dataset is evaluated in two ways to determine if it is closer to the target
MODIS NDVI than the original AVHRR dataset, and if it retains important interannual vegetation
dynamics that have previously been identified in the AVHRR data (Bounoua et al., 2000; Zeng et
al., 1999). First, time series for selected one degree boxes are presented to demonstrate the effect of
the neural net procedure on particular locations. Second, the NNndvi is compared to the GPCC
dataset to determine whether or not the correction has changed the relationship with observed
rainfall.
5.0 Results
Figure 1 shows a schematic representation of the neural net mapping of the AVHRR NDVI to the
MODIS NDVI during the years of overlap. Table 2 shows that the most important variable for
linking the two datasets is the AVHRR NDVI (as would be expected) followed by the surface
reflectance and total ozone column. In the TOMS data, the reflectance includes the degree of
Page 13 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only
14
cloudiness. Given the wide bands of the AVHRR sensor and the differences in processing, it is
expected that the TOMS reflectance is important in the correction (Cihlar et al., 2001).
Figure 2 shows the NDVI difference between the MODIS and AVHRR, and the MODIS and the
NNndvi by latitude band for a single image from December 2003. The biggest differences are in
the tropics which have high concentrations of atmospheric aerosols and water vapor that interfere
more with the AVHRR NDVI data than with the MODIS data (Huete et al., 2006). Another
substantial difference between the datasets is seen in the northern latitudes. The histogram is from
January, 2003, so the regions north of 40N have little active photosynthetic activity, the NDVI is
largely measuring differences in ground cover and atmospheric thickness. The GIMMS AVHRR
NDVI reports data over snow, ice, and during periods when there is no light, relying on the NDVI
to correctly record the very low photosynthetic activity during these months. MODIS NDVI data
incorporates much more sophisticated snow and ice detection, which results in large differences
between the AVHRR and MODIS data. Because we have inputs into the neural net that can
account for these differences (soil type, monthly changes in reflectivity), the differences between
MODIS and AVHRR are considerably reduced by the neural network processing.
Figures 3a and 3b show the spatial average of all pixels in the same latitudinal band for the
difference between the AVHRR and MODIS (3a) and NNndvi and MODIS (3b). The plots show
the significant improvement in the correspondence between the datasets in the tropics and in the
northern latitudes seen in Figure 2 is present in all years. Differences at the beginning and end of
the growing season in the far north are clearly seen. These differences will be significant to
scientists attempting to measure changes in phenology through time due to a warming climate. The
Page 14 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only
15
northern latitudes have experienced the largest degree of warming, thus these systematic
differences are important to both recognize and remove if a consistent, sensor-independent dataset
is to be developed.
The neural network process provides coefficients that were applied to the input data, to produce an
NDVI fit to MODIS from AVHRR back to 1982. Figure 4 shows the zonal averages of the
resulting dataset, displaying both seasonality and interannual variability as is expected. Table 3
shows the mean and standard deviation of the MODIS, AVHRR and NNndvi datasets. The mean
NNndvi is closer to the MODIS data than to the original AVHRR data. The differences in the
means can be seen in Figure 5, which shows the root mean square error (RMSE) in NDVI units
between the AVHRR - MODIS (Figure 5A), and the NNndvi – MODIS (B). The NNndvi dataset is
on average within 0.2 NDVI units of the MODIS data, removing the land-cover and regional
differences that can be seen in the top panel. The scatter above 0.2 RSME are seen in the map of
the RMSE in Figure 5B as being concentrated along the coastlines and where a sharp land-cover
gradient is located, such as along the Himalayas and Andes mountain ranges. This is likely to be
due to differences in the original land cover map between MODIS, AVHRR and TOMS and the
other ancillary datasets, as well as averaging procedures to make the one degree datasets. This
effect may be ameliorated by using a higher resolution, as at one degree much mixing of vegetated
and non-vegetated features occurs, particularly along sharp land cover and topographic features
which reduces the effectiveness of the neural network training.
Figure 6 shows the time series from MODIS, AVHRR, and the NNndvi from six selected one
degree pixels (Brown et al., 2006). These locations were selected from the Earth Observing System
Page 15 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only
16
land validation core sites described in Brown et al (2006) and were meant to display a range of
ecosystems and climates. The figure shows that the NNndvi is much closer to the MODIS series
than the original GIMMS AVHRR, particularly in areas with high humidity such as in the Cascades
of Washington state or Ji-Parana, Brazil. The NNndvi is higher than the GIMMS data, especially
during the winter months. In some regions where the match between MODIS and AVHRR was
fairly good originally, such as in the Harvard Forest, the fit between the datasets is extremely good.
Figure 7 shows the correlation coefficient, R, between the GPCC monthly gridded rainfall product
at one degree and the GIMMS AVHRR, NNndvi, and MODIS from 2000-2003. The maps in the
top two panels show that the NNndvi has a similar relationship with rainfall in semi-arid regions as
has been documented with the GIMMS data (Brown et al., 2004). It demonstrates that at one
degree, the correction maintains the datasets’ basic integrity and relationship with rainfall in semi-
arid zones. Panel D shows the histogram of the global correlation, showing a similar structure to
the data for the three datasets.
The results of this procedure are fairly robust, but they are not sufficiently good to be used for
scientific investigations. To determine if the data are usable immediately, we produced an anomaly
for August 2003 from each dataset versus the four year August mean for MODIS. Figure 8 shows
the histogram of the anomaly for August 2003 (when there was a major drought in Europe), which
shows the improvement of the NNndvi over AVHRR, but the data is still quite a bit different than
the MODIS data. Depending on the user requirements, this may be sufficiently similar. The bias in
the AVHRR has been removed so that the NNndvi is far more normally distributed. The Rp
statistic, a modified version of the Shapiro-Wilks test, measures the degree of normality of a dataset
Page 16 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only
17
by correlating the data with the standard normal distribution (Wilks, 1995). The Rp for the MODIS
anomaly shown in Figure 8 is 0.17, whereas the NNndvi anomaly has a value of 0.45, and the
AVHRR 0.47. So although the neural net correction has improved the data significantly, there are
still differences that are systematic for every pixel. The quality of the corrected data is significantly
better, however, as can be seen in Figure 9. The removal of cloud contamination in regions, such
as the Gulf of Guinea, that have always had depressed NDVI signal in the AVHRR dataset, is a
contribution that should not be underestimated.
6.0 Discussion
The lack of reliable climate observations throughout the AVHRR record is a major limitation in all
attempts to correct the AVHRR data to match the quality of the MODIS record. In order to remove
the systematic difference between the AVHRR and MODIS data due to atmospheric water vapor,
we need accurate observations of the amount of water vapor in the atmosphere at the time of data
acquisition. For AVHRR, the instrument that provides this data are derived from the Total Ozone
Mapping Spectrometer (TOMS) data (McPeters et al., 1998). TOMS data has its own problems
with data continuity and algorithms which may reduce the effectiveness of the neural network
because the issues may interfere with the NDVI differences we are trying to remove.
One reason for the lack of strong results in this experiment is the use of aggregated data. The
temporal mismatch between the 15 day AVHRR data, the 16 day MODIS data and the monthly
TOMS datasets has consequences that are difficult to identify. Although an effort was made to
minimize these problems through aggregation to the monthly time step, they may confound the
neural net. Aggregated data is much cleaner than daily observations, requires far less
Page 17 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only
18
computational effort (a key factor in running neural networks), and are the most widely used
products. In addition, daily data for the AVHRR NDVI and reflectances are currently not
available, thus they are not used here.
An effort is being made in the context of a NASA funded collaborative project called the Long
Term Data Record at the University of Maryland. In this project, daily AVHRR NDVI from
NOAA 7 through 14 (1981 to 1999) will be combined directly with MODIS data from 2000
onward. The data from the year 2003 will be used to relate the two datasets. The research
presented in this paper will illuminate the efforts of this project.
7.0 Conclusion
Remote sensing datasets are the result of a complex interaction between the design of a sensor, the
spectral response function, stability in orbit, the processing of the raw data, compositing schemes,
and post-processing corrections for various atmospheric effects including clouds and aerosols. The
interaction between these various elements is often non-linear and non-additive, where some
elements increase the vegetation signal to noise ratio (compositing, for example) and others reduce
it (clouds and volcanic aerosols) (Los, 1998). Thus, although other authors have used simulated
data to explore the relationship between AVHRR and MODIS (Trishchenko et al., 2002; van
Leeuwen et al., 2006), these techniques are not directly useful in producing a sensor-independent
vegetation dataset that can be used by data users in the near term.
There are substantial differences between the processed vegetation data from AVHRR and MODIS
[3, 7]. In order to have long data record that utilizes all available data back to 1981, we must find
Page 18 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only
19
practical ways of incorporating the AVHRR data into a continuum of observations that include both
MODIS and VIIRS. The results in this paper show that the TOMS data record on clouds, ozone
and aerosols can be used to identify and remove sensor-specific atmospheric contaminants that
differentially affect the AVHRR over MODIS. Other sensor-related effects, particularly those of
changing BRDF, viewing angle, illumination, and other effects that are not accounted for here,
remain important sources of additional variability. Although this analysis has not produced a
dataset with identical properties to MODIS, it has demonstrated that a neural net approach can
remove most of the atmospheric-related aspects of the differences between the sensors, and match
the mean, standard deviation and range of the two sensors. A similar technique can be used for the
VIIRS sensor once the data is released.
Page 19 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only
20
References
Atkinson, P.M. and Tatnall, A.R.L., 1997. Introduction: Neural networks in remote sensing. International
Journal of Remote Sensing, 18(4): 699 - 709.
Benediktsson, J.A., Swain, P.H. and Ersoy, O.K., 1990. Neural network approaches versus statistical
methods in classification of multisource remote sensing data. IEEE Transactions in Geoscience and
Remote Sensing, 28: 540-552.
Bishop, C., 1995. Neural Networks for Pattern Recognition. Oxford University Press.
Bounoua, L., Collatz, G.J., Los, S.O., Sellers, P.J., Dazlich, D.A., Tucker, C.J. and Randall, D.A., 2000.
Sensitivity of climate to changes in NDVI. Journal of Climate, 13(13): 2277-2292.
Brown, M.E., Pinzon, J. and Tucker, C., 2004. New Vegetation Index Dataset Available to Monitor Global
Change. EOS Transactions, 85(52): 565-569.
Brown, M.E., Pinzon, J.E., Didan, K., Morisette, J.T. and Tucker, C.J., 2006. Evaluation of the consistency
of long-term NDVI time series derived from AVHRR, SPOT-Vegetation, SeaWiFS, MODIS and
LandSAT ETM+. IEEE Transactions Geoscience and Remote Sensing, 44(7): 1787-1793.
Cihlar, J., Tcherednichenko, I., Latifovic, R., Li, Z. and Chen, J., 2001. Impact of Variable Atmospheric
Water Vapor Content on AVHRR Data Corrections over Land. IEEE Transactions Geoscience and
Remote Sensing, 39(1): 173-180.
Eklundh, L. and Olsson, L., 2003. Vegetation index trends for the African Sahel 1982-1999. Geophysical
Research Letters, 30(8).
Gallo, K.P., Ji, L., Reed, B.C., Dwyer, J. and Eidenshink, J.C., 2003. Comparison of MODIS and AVHRR
16-day normalized difference vegetation index composite data. Geophysical Research Letters, 31:
L07502-5.
Giri, C., Zhu, Z. and Reed, B.C., 2004. A comparative analysis of the Global Land Cover 2000 and MODIS
land cover data sets. Remote Sensing of Environment, 94: 123-132.
Holben, B., 1986. Characteristics of Maximum-Value Composite Images from Temporal AVHRR Data.
International Journal of Remote Sensing, 7(11): 1417-1434.
Huete, A., Justice, C. and Liu, H., 1994. Development of Vegetation and Soil Indices for MODIS-EOS.
Remote Sensing of Environment, 49: 224-234.
Huete, A.R., Didan, K., Shimabukuro, Y.E., Ratana, P., Saleska, S.R., Hutyra, L.R., Yang, W., Nemani,
R.R. and Myneni, R.B., 2006. Amazon rainforests green-up with sunlight in dry season.
Geophysical Research Letters, 33: L060405.
Huete, A.R. and Jackson, R.D., 1988. Soil and Atmosphere Influences on the Spectra of Partial Canopies.
Remote Sensing of Environment, 25: 89-105.
Huete, A.R. and Tucker, C.J., 1991. Investigation of soil influences in AVHRR red and near-infrared
vegetation index imagery. International Journal of Remote Sensing, 12(6): 1223-1242.
Lippmann, R.P., 1987. An introduction to computing with neural nets. IEEE ASSP Magazine: 4-22.
Los, S.O., 1998. Estimation of the Ratio of Sensor Degradation Between NOAA AVHRR Channels 1 and 2
from Monthly NDVI Composites. IEEE Transactions on Geoscience and Remote Sensing, 36(1):
206-213.
McPeters, R., Bhartia, P.K., Krueger, A., Herman, J., Wellemeyer, C., Seftor, C., Jaross, G., Torres, O.,
Moy, L., Labow, G., Byerly, W., Taylor, S., Swissler, T. and Cebula, R., 1998. Earth Probe Total
Ozone Mapping Spectrometer (TOMS) Data Product User’s Guide, National Aeronautics and Space
Administration, Greenbelt, MD.
Miura, T., Huete, A. and Yoshioka, H., 2006. An empirical investigation of cross-sensor relationships of
NDVI and red/near-infrared reflectance using EO-1 Hyperion data. Remote Sensing of
Environment, 100(2): 223-236.
Pinheiro, A.C., Privette, J.L., Mahoney, R. and Tucker, C.J., 2004. Directional Effects in a Daily AVHRR
Land Surface Temperature Data Set over Africa. IEEE Transactions in Geoscience and Remote
Sensing, 42(9): 1941-1954.
Page 20 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only
21
Pinzon, J., 2002. Using HHT to successfully uncouple seasonal and interannual components in remotely
sensed data, SCI 2002 Conference Proceedings Jul 14-18. SCI International, Orlando, Florida.
Pinzon, J., Brown, M.E. and Tucker, C.J., 2005. Satellite time series correction of orbital drift artifacts using
empirical mode decomposition. In: N. Huang and S.S.P. Shen (Editors), Hilbert-Huang Transform:
Introduction and Applications. World Scientific, Hackensack NJ, pp. 167-186.
Rudolf, B., Hauschild, H., Rueth, W. and Schneider, U., 1994. Terrestrial Precipitation Analysis:
Operational Method and Required Density of Point Measurements. In: M. Desbois and F.
Desalmand (Editors), Global Precipitations and Climate Change. NATO ASI Series. Springer-
Verlag, pp. 173-186.
Rumelhart, D.E., McClelland, J. and group, t.P.r., 1986. Parallel distributed processing: Explorations in the
microstructure of cognition, Vol. 1 - Foundations. MIT, 547 pp.
Sellers, P.J., Collatz, J., Hall, F.G., Meeson, B.W., Closs, J., Corprew, F., MeManus, J., Myers, D., Sun, K.-
J., Dazlich, D., Kerr, Y., Koster, R., Los, S., Mitchell, K. and Try, P., 1996. The ISLSCP Initiative I
Global Datasets: Surface Boundary Conditions and Atmospheric Forcings for Land-Atmosphere
Studies. Bulletin of the American Meteorological Society, 77(9): 1987-2005.
Slayback, D.A., Pinzon, J.E., Los, S.O. and Tucker, C.J., 2003. Northern hemisphere photosynthetic trends
1982-99. Global Change Biology, 9(1): 1-15.
Stathakis, D., 2008. How many hidden layers and nodes? International Journal of Remote Sensing,
submitted.
Stathakis, D. and Kanellopoulos, I., 2008. Global elevation ancillary data for land use classification using
granular neural networks. Photogrammetric Engineering and Remote Sensing, 74(1).
Stathakis, D. and Vasilakos, A., 2006. Comparison of several computational intelligence based classification
techniques for remotely sensed optical image classification. IEEE Transactions in Geoscience and
Remote Sensing(44): 8.
Steven, M.d., Malthus, T.J., Baret, F., Xu, H. and Chopping, M.J., 2003. Intercalibration of vegetation
indices from different sensor systems. Remote Sensing of Environment, 88: 412-422.
Teillet, M., Staenz, K. and Williams, D.J., 1997. Effects of Spectral, Spatial and Radiometric Characteristics
on Remote Sensing Vegetation Indices of Forested Regions. Remote Sensing of Environment, 61:
139-149.
Trishchenko, A.P., Cihlar, J. and Li, Z., 2002. Effects of spectral response function on surface reflectance
and NDVI measured with moderate resolution satellite sensors. Remote Sensing of Environment,
81: 1-18.
Tucker, C.J., 1979. Red and Photographic Infrared Linear Combinations for Monitoring Vegetation. Remote
Sensing of Environment, 8: 127-150.
Tucker, C.J., Pinzon, J.E., Brown, M.E., Slayback, D., Pak, E.W., Mahoney, R., Vermote, E. and El Saleous,
N., 2005. An Extended AVHRR 8-km NDVI Data Set Compatible with MODIS and SPOT
Vegetation NDVI Data. International Journal of Remote Sensing, 26(20): 4485-4498.
Tucker, C.J., Slayback, D.A., Pinzon, J.E., Los, S.O., Myneni, R.B. and Taylor, M.G., 2001. Higher northern
latitude normalized difference vegetation index and growing season trends from 1982 to 1999.
International Journal of Biometeorology, 45(4): 184-190.
van Leeuwen, W., Orr, B.J., Marsh, S.E. and Herrmann, S.M., 2006. Multi-sensor NDVI data continuity:
Uncertainties and implications for vegetation monitoring applications. Remote Sensing of
Environment, 100: 67-81.
Wilks, D.S., 1995. Statistical Methods in the Atmospheric Sciences, an Introduction. Academic Press, San
Diego.
Zeng, N., Neelin, J.D. and Lau, W.K.-M., 1999. Enhancement of Interdecadal Climate Variability in the
Sahel by Vegetation Interaction. Science, 286: 1537-1540.
Zobler, L., 1986. A world soil file for global climate modeling, NASA.
Page 21 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only
22
Captions
Table 1. Global datasets used in this paper.
Table 2. Statistics of the MODIS, AVHRR, and NNndvi datasets for 48 months of data (2000-
2003).
Figure 1. Schematic representation of the neural network used in this paper.
Figure 2. Graph showing the latitudinal means of the difference between MODIS, AVHRR and
NNndvi for January 2003. The figure highlights the zones where the neural net correction is the
strongest.
Figure 3. Zonal mean (averaged per latitude) of the difference between MODIS and AVHRR
(Panel A) and MODIS and NNndvi (Panel B) through time from 2000 to 2003.
Figure 4. Latitude-averaged mean of NNndvi from 1982 to 2003.
Figure 5. Root mean square error from MODIS-AVHRR (above) and the MODIS-NNndvi (below)
from 2000 to 2003 in NDVI units.
Figure 6. Time series plots of six latitude-longitude locations: A. Louga, Senegal (16, -16), Tigray
Ethiopia (14, 40), Bondville Illinois (10, -88), Cascades Washington (44,-122), Harvard Forest
Massachusetts (43,-72), and Ji-Parana Brazil (-11,-62).
Figure 7. Correlation coefficient of AVHRR, (A), NNndvi (B), and MODIS (C) vs GPCC rainfall
data. Panel D shows the histogram of the correlation coefficient of the NDVI vs gridded rainfall by
percent.
Figure 8. The August 2003 anomaly, defined as the difference between the MODIS, AVHRR and
NNndvi image for August 2003 and the mean of four August MODIS images (2000-2003).
Figure 9. Africa subset of one degree images for July 2002 for the AVHRR (A), NNndvi (B), and
the difference between the two (C).
Page 22 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only
23
Table 1.
Sensor AVHRR NDVI MODIS NDVI GPCC Rain
TOMS reflectivity, ozone and
aerosol
Data Source
GIMMS NDVIg
Operational Dataset
MODIS-Land and
Atmospheres Gridded Gauge data
NASA GSFC Ozone Processing
Team
Native Spatial
Resolution 8000 m 250 m 1 degree 26 km
Temporal
Resolution 15 day 16-day monthly Daily
Period
Available
July 1981 – present
(NOAA 7, 9,
11,14,16 and 17) Feb 2000 – present April 1986 – present
11/1978-5/1993(Nimbus 7)
5/1993-11/1994 (Meteor 3)
7/1996-12/2005 (Earth Probe)
Equatorial
Crossing ~9 AM - ~6 PM 10.30 AM NA ~9 AM - ~6 PM
Field of View
(FOV) ±55.4º ±55º NA ±55.4º
Table 2.
Element
Accumulated
weight
AVHRR NDVI 0.6
TOMS Reflectance 0.5
TOMS Column Ozone 0.3
Land Surface Type 0.3
TOMS Aerosol Index 0.2
Soil cover 0.2
Digital Elevation Model 0.2
Table 3.
Sensor NNndvi AVHRR MODIS
Global Mean
NDVI 0.4834 0.2982 0.4830
Global Std
NDVI 0.2384 0.2460 0.2522
Page 23 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only
24
Figure 1.
Page 24 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only
25
Figure 2.
Page 25 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only
26
Figure 3.
Page 26 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only
27
Figure 4.
Page 27 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only
28
Figure 5.
Page 28 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only
29
Figure 6.
Page 29 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only
30
Figure 7.
Page 30 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only
31
Figure 8.
−0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.40
500
1000
1500
2000
2500
3000
3500
4000
Anomaly
Nu
mb
er
of
P
ixe
ls
2003 NN − 4 year MODIS mean
2003 AVHRR − 4 year MODIS mean
2003 MODIS − 4 year MODIS mean
Page 31 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

For Peer Review Only
32
Figure 9.
Page 32 of 32
http://mc.manuscriptcentral.com/tres Email: IJRS-Administrator@Dundee.ac.uk
International Journal of Remote Sensing
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60