Document Sample
                    ANALYSIS IN CASE OF CAMBODIA

                                Naoki Mitsuzuka, Wataru Ohira
                                Japan Wildlife Research Center
                   3-10-10 SHITAYA , TAITO-KU TOKYO 110-8676 JAPAN
                           T +81-3-5824-0960 F +81-3-5824-0961

KEYWORDS: Deciduous forest, MODIS, LMF, Phenology , time series NDVI


Analysis of forest situation in Cambodia is difficult only using data captured by moderate spatial
resolution optical sensor (such as Landsat TM, SPOT HRG, ASTER) . There is problem to identify
deciduous forest by satellite imagery captured in the dry season because of defoliation, but data
captured in the rainy season is covered by cloud. So that we developed new method to classify
evergreen forest, deciduous forest, flooded forest, farmland using MODIS (MODerate resolution
Imaging Spectroradiometer) time series NDVI (Normalized Difference Vegetation Index) data
composite each 8 days. After time series filtering (Local Maximum Fitting) to remove the influence of
cloud, characteristics of NDVI profile of each forest types are made clear. Classification by Forest
Index (effective accumulative NDVI ) and characteristics of NDVI profile was implemented and Forest
Classification Map of Cambodia was created. Producer’s Accuracy is 80% in evergreen forest and 92%
in deciduous forest. User’s accuracy is 76% in evergreen forest, and 70- 49% in deciduous forest.


Careful consideration in phenology is important in cases where the remote sensed data is used for the
purpose of understanding the forest distribution in monsoon affected region, such as Cambodia.
The seasons in Cambodia are roughly divided into rainy season and dry season. The data observed in
the rainy season by moderate resolution optical sensor (such as Landsat TM, SPOT HRG, and ASTER)
is affected by cloud cover, resulting in difficulty to use for analysis. Therefore, the data obtained in the
dry season should be used.
The evergreen forest shows forest-like reflection characteristics throughout the year, allowing the
understanding of the distribution areas by satellite imagery, irrespective of times during the dry season.
In case of deciduous forest, however, leaves fall in the dry season. Therefore, such forest appears in the
same situation as grassland or bare land on the data of optical sensor. In regards to deciduous forest, it is
necessary to take into full account which time of defoliation the data was observed in the dry season.
If the satellite data is observed in the very last stage of the rainy season (in late October), the deciduous
forest can be easily understood. However, in consideration of the number of revisit dates of the
moderate resolution optical sensor, the probability to be observed is low in this short period of time in
the subject areas, so the acquisition of the data in expected point of time is difficult.
As one of the answer to the problems which cannot be solved by only these moderate resolution optical
sensors, there is a method which employs high-revisit observation satellite data. In this study, the
time-series NDVI data set of the high-revisit observation sensor (MODIS) was used, so a more detailed
forest distribution grasping method utilizing the temporal resolution was implemented. The “temporal
resolution” is one of the advantage of MODIS data. We applied this point to classify the several forest


2.1 Truth data

Locations of some types of forests and farmlands were determined to be used as the truth for the
analysis of the method, according to the following information:
  ASTER data
  Forest distribution map created by the Cambodia Forestry Administration (2006)
  On-site observation (January 2008)

2.2 MODIS NDVI data time-series filtered

MODIS data were downloaded form LPDAAC’s site, as data set named MODIS/Terra Surface
Reflectance 8-Day L3 Global 250m SIN Grid V004 and V005. After download, using band 1 and 2 ,
NDVI value was calculated.
The LMF (Local Maximum Fitting) processing (Sawada 2005) was conducted, as a preprocessing, to
the NDVI of the MODIS. The LMF is a time series filtering processing for the high-revisit observation
satellite data, which allows NDVI value of each pixel to be compensated by model, and inappropriate
pixel value due to cloud and system noise to be corrected.

           Fig.1 MODIS NDVI 31.Oct 2005 Left: Original Right: LMF Processed


3.1 Forest Index: effective accumulative NDVI by threshold

Forest Index (FI) is obtained by accumulating NDVI values over 0.7 in a time span of one year (Ohira
The MODIS FI calculated in this study is obtained by accumulating values of over 218 which fall on
the original NDVI 0.7 as digital values. Therefore, the FI ranges from the minimum value 0 to the
maximum value 1,702.
A relationship between MODIS FI value and the forest classification of forest distribution map created
by the Cambodia Forestry Administration (FA) is shown in Figure 2. The FI in which each forest
classification represents ranges from the maximum value obtained by adding a half of standard
deviation to the mean value to the minimum value obtained by subtracting a half of standard deviation
from the mean value.
The forest classification by FA is composed of Evergreen forest, Semi-evergreen forest, Deciduous
forest, Wood and shrub land evergreen, Wood and shrub land dry, Bamboo, Other forests, and
From a viewpoint of this content, it can be believed that the forests, which are well-suited for the
classification by the time-series NDVI characteristics such as FI, are evergreen forest (including
bamboo), deciduous forest, and non-forest. Other forest classes are displayed on the satellite data as
mixed pixels such as tree, grass, and bare land. As a result, the pattern of the NDVI seasonal variations
becomes inconstant depending on the mixture fraction of those mixed pixels, which does not allow the
forest classification to be identified. In addition, it is predicted that the semi-evergreen forest can be
classified into either evergreen forest or deciduous forest, based on the majority of pixels of evergreen
trees or deciduous trees. Consequently, it would be difficult to set the semi-evergreen forest as single
Based on the above-mentioned data, a classification of general forest conditions was conducted,
according to the FI threshold. As criteria,
1: Evergreen forest in case of FI 628 or higher
2: Evergreen forest or deciduous forest in case of FI 268 or higher. However, pixels which overlap with
1 are judged to be evergreen forest
3: Non-forest in case of FI 248 or lower.
Applying these criteria to MODIS FI data, it confirmed that the forest distribution of Cambodia were
represented roughly. However, within three criteria, how the judgment is made when respective FI
values overlap should be discriminated not only by FI but also by additional information.

         Fig.2 Relationship between MODIS FI and Forest Category by Forest Administration, Cambodia

3.2 Characteristics of NDVI time-series profile to each vegetation type

As shown in Figures 3 and 4, the NDVI profile of seasonal variations shows discriminative forms
based on the classification of vegetation,.
                Fig.3 NDVI Profile of Deciduous forest           Fig.4 NDVI Profile of Evergreen forest

Forest characteristics of Cambodia appear clearly in particular in the period from the end of the rainy
season and until beginning of it. Therefore, attention was paid to the MODIS data in this term (22 time
points of MODIS data, from October 31 to April 14. In this paper, shorten FCT from “Term of Forest
Characteristics “).
The forest distribution map by FA is created by using the Landsat data of 2005 and 2006. Therefore, in
order to match the point of time of the data, the processing was implemented for the data of
above-mentioned term in 2005-2006.
Based on the MODIS NDVI in FCT, profiles by typical vegetation are shown in Figures 5 to 8.
As a result of investigation of profile of the plural spots falling on the four categories (deciduous forest,
evergreen forest, farmland, and flooded forest), the profile characteristics of FCT are summarized as

        Fig.5 NDVI Profile of      Fig.6 NDVI Profile of   Fig.7 NDVI Profile of     Fig.8 NDVI Profile of

          Deciduous forest          Evergreen forest        Farmland                  Flooded forest

In case of deciduous forest (Figure 5), NDVI 0.72 or higher is shown because leaves have not fallen at
the initial time point of FCT (31.Oct 2005). After this, defoliation advances, and the NDVI lowers
gradually. In the second half of FCT, the NDVI drops down to approximately 0.40, and then it recovers
gradually. At the final time point of FCT (14. Apr 2006), the NDVI recovers to be a value of 0.56 or
higher, however, showing lower value than that at the initial time point of FCT (31.Oct 2005).
In case of evergreen forest (Figure 6), the NDVI shows the mean value 0.8 or higher throughout FCT,
with a transition of amplitude (difference between maximum value and minimum value) of entire FCT
in the narrow range of approximately 0.2.
In case of farmland (Figure 7), although the NDVI shows a high value of 0.72 or higher in the
beginning of FCT, it decreases gradually. In the second half of FCT, the NDVI transits in the value of
0.4 or lower. Therefore, the NDVI mean value in the second half of FCT shows only 0.40 or lower. In
addition, unlike the first half, the amplitude shows only a small range of 0.16 or lower. Unlike in the
case that natural vegetation such as deciduous forest in Figure 5 which recovers as FCT comes close to
the end, the profile of farmland in Figure 7 shows the proof of management by humans. It is believed
that the NDVI starts to increase under irrigation in the rainy season.
In case of the flooded forest (Figure 8), it is difficult to judge unconditionally by only the profile
characteristics of NDVI value. Even the forest defined as a flooded forest, it has a situation in which
various densities and tree species are mixed. In addition, factors, such as time of flood length and flood
height, that are influenced by the distance from the water source also have an influence on the NDVI
profile characteristics. Those complex situations have a large impact on the data observed by the
satellite. As a result, it is difficult to encompass the conditions of NDVI profile in accordance with the
flooded forest. On the premise of this point, the profile characteristics peculiar to the flooded forest is
attempted to be considered. At the beginning of FCT, the NDVI shows a value of 0.17 or lower
because of flood. In addition, innate characteristics of the forest emerged as the water receded in the
second half of FCT, showing the maximum value of 0.7 or higher in only the second half of FCT.
Combining the characteristics by the vegetation of the above-mentioned NDVI time-series profiles
with the FI threshold forest classification in the above paragraph 3.1, the forest classification method
which takes advantage of the time-series NDVI has been designed.

3.3 Discrimination of the forest classification by FI and NDVI profile

Based on the knowledge in above paragraphs 3.1 and 3.2, the criteria of evergreen forest, deciduous
forest, flooded forest, and farmland are adopted. The program of this processing is created in macro
language of software for remote sensed data, providing the forest classification map data.
The forest classification map in Figure 10 is obtained as a result of judgment and classification, based
on these conditions.

          Fig.9 Forest Distribution Map by FA, Cambodia   Fig.10 Forest Classification Map by MODIS NDVI Profile
 and FI


Error Matrix expressing the classification accuracy is shown in Table 1. Producer’s Accuracy is 80% in
evergreen forest and 92% in deciduous forest. In the evergreen forest and deciduous forest zones
classified by FA, it can be evaluated that sufficiently - high accuracy is shown. On the other hand, with
respect to the User’s Accuracy, evergreen forest is 76% and deciduous forest is 49%. In categories of
FA, some forest types which cannot be classified using MODIS data are contained. Semi-evergreen
forest, Wood and shrub land dry, and Other forest are these. Assuming all these types are categorized in
deciduous forest of MODIS classification, User’s accuracy of deciduous forest class by MODIS is up
to 70%. In fact User’s accuracy of deciduous forest is between 70 and 49%.
                            Table.1 Error Matrix of Forest classification by MODIS data
*Class 1:Evergreen 3:Deciduous 9:Farmland 10:Flooded Forest
 G_1:Evergreen G_2:Semi-evergreen G_3:Deciduous G_4:Other forest G_5:Shrub,dry
 G_6:Shrub,evergreen G_7:Non-forest G_8:Bamboo

There are areas judged as non-forest or farmland in the MODIS classification, however, they are
displayed as deciduous forests on the forest distribution map by FA (Figure 13, in white). Considering
of MODIS NDVI profile 2001-2006 on these areas (Figure 11,12), some of them show high possibility
which indicates change of land use or forest situation . In our farther step of this study, we will research
about current forest situation in prior area for investigation to improve the classification method.

  Fig.11 MODIS NDVI Profile showing change                              Fig.12 MODIS NDVI Profile showing
   from forest to non-forest                                             change of land use

  Fig.13 prior area for investigation of forest situation (in white)

Land Processes Distributed Active Archive Center , MODIS Data site ,
Sawada,Y.,Mitsuzuka,N., and Sawada,H., 2005, Development of a Time-series Model Filter for
HighRevisit Satellite Data, Proceedings of the 2nd International VEGETATION Users Conference,
Office for Official Publication of the European Communities, PP83-89
Ohira,W.,Wada,Y.,Miyashita,Y.,Nonaka,I.,Mitsuzuka,N., Development of Technology to Analyze
Forest Dynamism in Eastern Part of Asia, 2004, Proceedings of the 24th Asian Conference on Remote