DEVELOPMENT OF SEMI-AUTOMATED SYSTEM BASED ON

PALSAR DATA FOR FOREST MAPPING AT REGIONAL SCALE

N. Longépé 1, P. Rakwatin 1, O. Isoguchi 1, M. Shimada 1, Y. Uryu 2

1
Earth Observation Research Center, Japan Aerospace Exploration Agency,
2-1-1 Sengen, Tsukuba, Ibaraki 305-8505 Japan
nicolas.longepe@jaxa.jp
2
World Wildlife Fund,
Kantor Taman A9 Unit A2, Mega Kuningan, Jakarta, Indonesia.

Abstract

This study investigates the relevancy of PALSAR sensor onboard the ALOS
satellite for regional land cover classification. Our study site is the tropical
rainforest in Indonesia, specifically the entire Borneo and Sumatra islands.
These islands covering an area of about 1.18 millions km2, an adequate
methodology should be investigated in order to tackle the radiometric
imbalance across tracks but also the intrinsic variability of the imaged natural
media. Starting from the orthorectified 50m mosaic products freely available on
the Internet (HH and HV channels acquired in 2007 and 2009), an algorithm
based on the Support Vector Machines is implemented and performed in a non-
supervised way in the aim to produce a global land cover map at 50m
resolution across the Indonesian islands. The results are compared with a
Landsat-based estimation available over the entire Sumatra Island. At this
scale, the validation is a delicate issue since the Landsat-based classification
may have an accuracy of the order of 85-90%. The agreement is equal to
82.9% for the forest/non forest map which is remarkable since this assessment
is performed over more than 171 millions pixels. Some enhancements are
suggested for the SVM-based classifier.

1. Introduction

Carbon emissions accounting has become an activity of international

importance. Any systems able to monitor at global scale the biomass decay
(carbon released to the atmosphere) or regrowth (terrestrial uptake) are of high
interest to assess the anthropogenic carbon flows. Even though the radar-
based remote sensing community does not cast doubt on the usefulness of L-
band SAR data for forest monitoring, real evidences still need to be presented
when moving to large, if not, global scale applications. Begun in 2000, the
ALOS Kyoto and Carbon (K&C) Initiative is a global, multi-disciplinary research
initiative of the Japan Aerospace Exploration Agency (JAXA), aiming to produce

1
remotely-sensed products which can support international conservation
Conventions, Carbon cycle science and natural Conservation (CCC) such as
the UNFCCC activities on REDD [1]. The K&C Initiative is based on three
themes related to the global biomass: forests, wetlands, deserts and semi-arid
regions, and a fourth one dealing with the generation of regional ALOS
PALSAR mosaics. Indonesia has experienced one of the highest deforestation
rates in the world with an average of 1.78 Mha/yr in the 90s, reduced to 0.71
Mha/yr from 2000 to 2005 [2]. Driven mainly by pulp and paper and oil palm
industries, the Borneo and Sumatra islands have been drastically changing
since the 80s. In this research study, we highlight the role of the PALSAR-
based mosaics for the generation of forest maps over these two islands in a
first step, before moving to the entire tropical rainforest belt in the future. It
should be mentioned that this study focuses on natural forest which does not
include forest plantations.

Many classification techniques have been used in the SAR community for
forest monitoring ranging from the simplest ones (Threshold, Maximum-
likelihood) to enhanced methodologies (Neural Network, Random Forest,
Support Vector Machines). In a parallel study, we showed that a Support Vector
Machine (SVM) based approach can be efficient for land cover classification at
regional scale [3]. An optimized set of textural parameters was computed based
on the so-called SVM-based Recursive Features Elimination (RFE) process.
Using this optimized information plus the radiometric channels (HH and HV
polarizations), the SVM classifier can be efficiently applied resulting in an
accuracy of about 70.0 and 86.5 % over the Riau Province (Sumatra) at 50m
resolution (40.8 million pixels) when, respectively, 6 and 2 classes are
discriminated. This process just needs one dual-polarization dataset and could
be applied to generate annual land cover maps over the tropical rainforest.
However, the generation of the textural parameters and their integration into the
SVM classifier are highly demanding for memory and processor capabilities.
The use of such a methodology still needs to be further refined to be fully
operational at global scale with higher resolution (25/10 meters).

The aim of this paper is mainly to investigate the capabilities of the PALSAR
K&C mosaics freely available on the JAXA’s website [4] for the forest
classification over the tropical rainforest in Indonesia. Based on our previous
experience with the SVM approach, we decided to extend the target area in this
research (from 11 million ha in Riau to more than 118 million ha for
Borneo/Sumatra islands). Two aspects will result from this enlargement: the
radiometric imbalance across tracks and the intrinsic variability of the imaged
natural media will have to be particularly addressed. In the first Section, the
PALSAR based mosaic products which have been generated in the framework
of the K&C project will be described with a particular interest on the geometric
and radiometric information. Section 3 presents the SVM methodology and its
implementation for this project. Results are shown for the entire Borneo and
Sumatra islands in the last Section. Accuracy assessment is performed for the
Sumatra case with the comparison of a Landsat-based estimation provided by

2
the WWF Indonesia. Some recommendations for further enhancements are
finally given.

2. PALSAR mosaics using calibrated path images

The generation of continental mosaics at JAXA has been initiated in the late
90’s through the Global Rain Forest and Boreal Forest Mapping (GRFM/GBFM)
projects in which 100m spatial resolution mosaics of the entire tropical and
boreal zones were generated using data acquired by the L-band SAR-based
Japanese Earth Resources Satellite (JERS-1). At that time, different attempts
have been simultaneously carried out [5, 6, 7]. The K&C Initiative project forms
its continuation and extension being logically based on the use of the ALOS
PALSAR sensor. When dealing with SAR-based mosaics, two aspects have to
be carefully addressed: the geometric and the radiometric accuracies.

The co-registration of neighboring strips in slant range geometry being difficult

due to the difference of incidence angle (especially over mountainous areas),
most of the mosaicking processes involve an ortho-rectification based
methodology. The approach developed by JAXA [5] is based on “strip” data
which have an azimuth length exceeding thousands of kilometers. Reducing the
number of associated image numbers is the key general concept to avoid
radiometric and geometric problems. However, this method requires the
implementation of a long-strip SAR processor. Due to the joint effects of the
movement of the platform and Earth’s rotation, the Doppler parameters have to
be perfectly and continuously estimated all along the time acquisition. In
addition to this, the continuous check on the Pulse Repetition Frequency (PRF)
and the Sampling Windows Start Time (SWST) is one of the issue when
dealing with long-strip imaging. For the PALSAR sensor onboard ALOS, the
PRF changes approximately every 2000km to fully cover the variation of the
Doppler bandwidth along the latitude. This aspect has to be considered to
ensure a continuous image in the azimuthal direction when integrating the
range-compressed data with two different PRFs. The processing steps to
generate the slant-range SAR path are described in [8]. Then, all these strips
are ortho-rectified using the 90m SRTM DEM, projected to an individual geo-
reference coordinate system (LGCS) and finally re-adjusted successively from
West to East (for ascending orbit) to a global coordinate system by means of
cross-correlation of overlapped regions.

Beyond these geometric aspects, the radiometric quality is of high interest as

well. During the calibration and validation phase, the PALSAR sensor has been
evaluated using both natural and artificial calibration targets. The Amazon
forest was used since the Normalized Radiometric Cross Section (NRCS) has
been shown to be almost constant over this natural media with a small
seasonal variation of 0.27 dB. In addition, 572 target-calibration points were
collected from the Corner Reflectors deployed all over the globe. As a result,
the radiometric stability of PALSAR is of the order of 0.76 dB [9]. In the case of

3
the Sumatra island, more than 20 strips are used to cover the entire area for
2009 (see Fig. 1). The acquisition dates range from June 13 to October 15.
PALSAR being highly stabilized and sensitive to the surface backscatter
change (e.g. moisture), some offsets may appear between neighboring strips of
gamma-naught. Even if these features represent the true radiometric
information, they can be detrimental for post-processing techniques such as
segmentation and classification. In order to obtain uniform γ0 values across the
entire mosaic, a gain balancing method has been applied [8].

Fig. 1: PALSAR mosaic (HH: red-blue, HV:green) of Sumatra (Indonesia) and Malaysia
overlaid into GoogleEarth. The PALSAR strips of approximately 70km swath width are
illustrated in red polygons.

It should be noted that the version of the SAR mosaic processor used in this
study provides 64-look orthorectified images which are not corrected for slope
effect1. Being relatively too “smooth”, textural information have not been
computed in this study and the radiometric information will be solely used for
this version.

3. Methodology

Given the radiometric information provided by the dual-polarized PALSAR

mosaics in 2007 and 2009, a classification methodology based on the Support
Vector Machines (SVM) is outlined in this section. The methodology
implemented for this study is composed of 5 steps: 1) a classical One-Against-
All (OAA) SVM-based training, 2) a slight optimization on the independent SVM
classifiers, 3) a pre-clustering operation and the SVM application on a seed-

1
currently available on the JAXA’s website (as of Sept. 2010)

4
cluster basis 4) a merging process and 5) a relaxation method using contextual
information.

1) About 10 000 pixels are carefully selected over the Borneo island given a
sparse ground truth dataset and our experience. This set of pixels equally
contains Nc = 9 land cover types which mainly cover the western Indonesian
islands. For this paper, two dual-polarized mosaics are used (2007 and 2009)
resulting in a total of 4 layers as input of the SVM classifiers. After the training
phase generating the binary classifier functions f c (c=1… Nc) in the case of a
OAA approach, Nc outputs f c (x) could be produced [10-12].

2) However, SVM considers independent classifiers f c and the SVM-based

outputs are directly compared without considering the competence of the
classifiers. Several methods exist to associate them in a consistent manner in
the framework of the “One-Against-All” approach. Their concepts consist to use
some error-correcting codes to improve the multi-class classification [13-14]. In
our study, an original and simple procedure is used. The error assessment will
be performed with the ground truth dataset. Starting from all the Nc = 9
classifiers, the worst classifier is selected using the Producer Accuracy (PA) as
a criteria. The classifier is then optimized to maximize the overall accuracy by
the integration of a bias adjusting the values of the selected classifier. This
operation is successively performed over all the classifier, starting from the
worst to the best classifier. When all the classifiers have been optimized once,
this operation is reiterated until the number of iteration has reached the user-
defined maximum.

3) The third step consists in computing some seed-clusters of small size (20
pixels or so). SVM requires significant CPU resources and may be relatively
slow to process. Applying this technique to small clusters (instead of pixels) will
significantly reduce the CPU consumption. One additional drawback of SVM is
that it does not consider the spatial location of the pixels in the final decision.
The classical SVM-based methodology is independently applied pixel by pixel.
However, two neighbouring pixels are not totally independent. Beyond the
classical approach using a moving boxcar and a majority vote [15], a clustering
method permits the generation of clusters with different sizes/shapes. In this
paper, we use a two-fold approach. First of all, the k-means method partitions
the HH/HV channels in 2009 into 30 groups, which are used to identify
homogeneous regions within the image. The different homogeneous areas are
then successively merged using a L2 norm metric until all the regions become
large enough (20 pixels). This method is usually called the “average group
linkage”. The Nc outputs f c (x) are computed over each homogeneous seed-
cluster.

4) Continuous areas should be computed from the final classifier: neighboured

pixels are to be considered. In order to produce an even smoother final
decision, each output of the SVM classifier is used as a criterion (with a L1 norm
metric) for merging again the seed-clusters previously computed in step 3) until

5
reaching large-size clusters. Depending on the land cover type and its
associated SVM output, the minimum size of clusters may vary (few hundreds
to one thousand). For all the Nc = 9 classifiers, the SVM-based output is finally
averaged over the computed large clusters.

5) Finally, the last step is the relaxation method which attributes a final decision.
Since the segmentation is operated for each SVM-based output, clusters do not
necessarily match between classifiers. A pixel-based relaxation method is
operated, attributing the most likely class to each pixel. This process is also
weighted with contextual information such as altitude, distance with water
2
bodies and so on . For example, it is unlikely to retrieve mangrove forest at high
altitude which is obviously not the case along the coast. Also the distinction
between mountainous forest and lowland forest is operated with the altitude
information.

4. Results

For this study, it is assumed that the natural media across the Borneo and the
other Indonesian islands are very similar, both for their bio- and geo-physical
characteristics. The SVM-based classifiers have been trained once and for all
using the ground truth dataset available over Borneo. Based on the
methodology described in Section 3, some classification maps are shown in this
Section.

Starting with the Borneo island which is about 736 000 km2, the results in Fig. 2
seem relatively good in a general manner. Most of the land cover types are
consistently discriminated. However, the accuracy of such a map has not been
assessed up to now. The available ground truth is quite sparse consisting in
several spots located across the island. In addition, it is partially used during the
training phase. Over-training could induce an excellent accuracy over the
training samples and their surroundings but with a low generalization capability.
Such a ground truth is not satisfactory enough to give an insight on the real
accuracy of the classification. That is why the same SVM rules have been
applied over the Sumatra Island where a forest/non forest “ground truth” is
available. Forest classification has been manually performed based on Landsat
images acquired in 2008/2009. More details can be found in [16]. The classes
previously defined over Borneo have been merged to match those of the
Landsat-based estimation in Sumatra.

2
Here, water bodies are detected using a simple threshold-based approach.

6
Fig. 2: Classification (v1.21) across the Borneo island in 2009 at 50m resolution using
PALSAR mosaic products.

The PALSAR and Landsat-based forest classifications and its associated

confusion matrix are shown in Fig. 3 and Table 1, respectively. The accuracy
assessment is performed on a total of about 171.3 millions pixels.

Class \ Truth Non Forest Natural Forest

Non Forest 25.481.837 ha 2.439.537 ha
83.7 % 19.1 %
Natural 4.952.175 ha 10.320.468 ha
Forest
16.3 % 80.9 %

Table 1: Confusion matrix for the Sumatra forest map

7
Fig. 3: Forest maps of Sumatra by the WWF Landsat-based classification (left) and by
the proposed methodology (v1.2) at 50m using PALSAR mosaic products (right). The
legend is: dark green – natural forest over dry land, mid-green – natural forest over
swamp areas, light green – mangrove.

The agreement between these 2 estimations is relatively good with an overall

accuracy of 82.9 %, the Producer Accuracy for the forest and the non forest
classes being equal to 80.9 % and 83.7 %, respectively. As confirmed by the
classification maps across Sumatra, it seems that this classification
methodology slightly underestimates some forested areas. This is particularly
true over mineral soil. Also some misclassifications of “Natural forest on
swampy” as “Natural forest” still occur. It might be explained by the fact that
peat areas in Borneo and Sumatra may be different, in terms of depth, soil
properties, or soil moisture. This aspect has to be further investigated in the
future. However, the presented methodology combining the PALSAR mosaics
and the Support Vector Machines is relatively robust and consistent. The
capability of PALSAR sensor for large scale forest monitoring has been
demonstrated. It is worth emphasizing the fact that the SVM-based
classification rules retrieved from a sparse training dataset in Borneo can be
successfully applied across the Borneo and Sumatra islands. Finally, Fig.4
shows the classification results over the entire Indonesia and Malaysia.

8
 Fig. 4: Classification (v1.2) across Indonesia (Sumatra, Kalimantan, Sulawesi, Java)
and Malaysia in 2009 at 50m resolution using PALSAR mosaic products.

5. Conclusion

In the framework of the Kyoto and Carbon initiative led by the Japan Aerospace
Exploration Agency, dual-polarized PALSAR mosaics have been generated to
support information needs raised by international environmental Conventions,
Carbon Cycle Science and Conservation of the environment (CCC). In this
study, they are directly used to generate 50m land cover and forest
classification maps. Time series of HH/HV channels in 2007 and 2009 are
solely used as input of a Support Vector Machines methodology. Once the
SVM approach has been trained at local scale, the presented methodology is
performed in an automatic manner. Results are relatively good considering the
scale of the retrieved maps and the fact that no other data source is used. The
potential of the PALSAR sensor with its consistent spatial and temporal
acquisition strategy is clearly demonstrated.

This paper should be considered as an update of the version introduced in [17]

and presents the current status of this methodology (as of September 2010).
Many enhancements are still possible and will be integrated in the future. In a
previous study, it has been shown that the textural information can be of high
interest for the discrimination of land cover types even if their uses are
detrimental for CPU time consumption [3]. The relevance of textures will be all
the more relevant with the use of higher resolution dataset (25m, 10m ...). The
use of pattern detection could be useful to detect man-made structure such as
canals or roads which are highly linked to industrialized plantations. This

9
information could be easily added as input of the SVM-based classifiers or
simply used as contextual information during the post-processing step.

Acknowledgment

This work has been undertaken within the framework of the JAXA Kyoto and
Carbon Initiative. The authors would like to thank the engineers at the Remote
Sensing Technology Center of Japan for processing the mosaic products,
especially T. Ohtaki and H. Okumura. They also acknowledge D. Hoekman
(Wageningen University) for providing the ground truth dataset in Borneo.

References

[1] United Nations Framework Convention on Climate Change, 2005, Reducing

emissions from deforestation in developing countries: approaches to stimulate
action-draft conclusions proposed by the president.

[2] M.C. Hansen, S.V. Stehman, P.V. Potapov, B. Arunarwati, F. Stolle, and K.
Pittman, 2009, Quantifying changes in the rates of forest clearing in indonesia
from 1990 to 2005 using remotely sensed data sets. Environmental Research
Letters, 4(3).

[3] N. Longépé, P. Rakwatin, O. Isoguchi, M. Shimada, Y. Uryu, and K.

Yulianto, 2010, Assessment of ALOS PALSAR 50m orthorectified FBD data for
regional land cover classification by Support Vector Machines. IEEE Trans.
Geosci. Remote Sens., under review.

[4] JAXA EORC, 2009, http://www.eorc.jaxa.jp/ALOS/en/kyoto/kyoto_index.htm.

Official Kyoto and Carbon initiative project website.

[5] M. Shimada and O. Isoguchi, 2002, JERS-1 SAR mosaics of Southeast Asia
using calibrated path images. Int. J. Remote Sens., 23:1507–1526.

[6] P. Siqueira, S. Hensley, S. Shaffer, L. Hess, G. Mc-Garragh, B. Chapman,

and A. Freeman, 2000, A continental-scale mosaic of the Amazon basin using
JERS-1 SAR. IEEE Trans. Geosci. Remote Sens., 38(6):2638–2644.

[7] G. De Grandi, P. Mayaux, Y. Rauste, A. Rosenqvist, M. Simard, and S.S.

Saatchi, 2000, The Global Rain Forest Mapping Project JERS-1 radar mosaic
of tropical Africa: Development and product characterization aspects. IEEE
Trans. Geosci. Remote Sens., 38(5):2218–2233.

10
[8] M. Shimada and T. Ohtaki. Generating continent scale high quality SAR
mosaic datasets: Application to PALSAR data for global monitoring, 2010, IEEE
J.Sel. Applied Earth Obs. Remote Sensing (JSTARS), under review.

[9] M. Shimada, O. Isoguchi, T. Tadono, and K. Isono. PALSAR radiometric

calibration and geometric calibration, 2009, IEEE Trans. Geosci. Remote Sens.,
47:3915–3932.

[10] V. Vapnik, 1998, Statistical learning theory. Wiley- Interscience,.

[11] N. Cristianini and J. Shawe-Taylor, 2000, An introduction to support vector

machines (and other kernel-based learning methods). Cambridge University
Press.

[12] C. J. C. Burges, 1998, A tutorial on support vector machines for pattern

recognition. Data Mining Knowledge Discovery, 2(2):121–167.

[13] R. Rifkin and A. Klautau, 2004, In defense of one-vs-all classification. J.

Machine Learning Research, 5:101–141.

[14] S. Escalera, 0. Pujol, P. Radeva, P, 2010, the Decoding Process in

Ternary Error-Correcting Output Codes. IEEE Pattern Analysis and Machine
Intelligence, 32:120-134.

[15] Y. Tarabalka, J.A. Benediktsson, and J. Chanussot. Spectral-spatial

classification of hyperspectral imagery based on partitional clustering
techniques, 2009, IEEE Trans. Geosci. Remote Sens., 47(8):2973–2987.

[16] Y. Laumonier, Y. Uryu, M. Stuwe, A. Budiman, B. Setiabudi, and O.

Hadian, 2010, Eco-floristic sectors and deforestation threats in Sumatra:
identifying new conservation area network priorities for ecosystem-based land
use planning. Biodivers Conserv.

[17] N. Longépé, P. Rakwatin, O. Isoguchi, M. Shimada, Y. Uryu, K. Yulianto,

2010, Large scale classification using PALSAR/ALOS data: case of the tropical
rainforest in Indonesia. Proc. ESA Living planet symposium, June 2010.

11