TREE SPECIES CLASSIFICATION USING ERS SAR AND MODIS NDVI IMAGES
. Markus Tórmà" ', Juho Lumme*, Ulla Pyysalo", Niina Patrikainen, Kari Luojus"
Institute of Photogrammetry and Remote Sensing, Helsinki University of Technology, P.O.Box 1200, FIN-02015 HUT, Finland
- (Markus. Torma, Juho.Lumme, Ulla.Pyysalo)@hut.fi
5 Laboratory of Space Technology, Helsinki University of Technology, P.O.Box 3000, FIN-02015 HUT, Finland -
- (patrikai, kari.luojus)@avasun.hut.fi
KEY WORDS: Forestry, Land Cover, Classification, Fusion, Optical, SAR, Multitemporal
ABSTRACT:
À set of ERS SAR and optical MODIS-images were classified to land cover and tree species classes. Different methods for pixel and
decision based data fusion were tested. Classifications of featuresets were carried out using Bayes rule for minimum error. The
results were not very successful, the classification accuracies of land cover classes varied from 43% to 75%, depending on the used
features and classes. The decision based data fusion method, where the a'posteriori probabilities representing the proportions of
different land cover classes of low resolution classification are used as a'prior probabilities in high resolution classification looks
promising. Using this method, the increase of overall and classwise accuracies can be more than 10 and 25 Yo-units, respectively.
1. INTRODUCTION
Forest assessment deals with the methods of obtaining
information on forest resources: estimation of growing stock,
growth and health of the forest. That information is a basis for
decisions of the forest industry, the official forest policy and the
forest owners. For countries such as Finland, where 30% of
exports is based on forestry products and the percentage of the
forest area (76%) is the highest in the world, development of
inventory methods are a necessity.
The national forest inventory of Finland was the first national
inventory in the world to use satellite images (Tomppo 1991). It
is based on the use of optical data like Landsat images.
Unfortunately weather conditions limit the use of optical data.
For example, here in Finland summertime is usually quite
cloudy. There are usually only several days in summer when
large area of Finland is cloud-free, and during wintertime it is
dark also daytime and snow everywhere. These facts have lead
to investigate the use of SAR-images in forest inventory.
The previous single frequency / polarization spaceborne SAR-
instruments have not been that successful from land cover or
forest classification point of view (Herold etal, 2004,
Kurvonen etal, 2002, Tórmá, 1999). This is due to the low
information content of individual images and noise which is
difficult to remove. Utilization of texture (Kurvonen et.al.,
2002, Tôrmä, 1999), coherence (Gaveau et.al., 2003, Tôrmä,
1999) or polarization information (Randon et.al., 2001) improve
results. The instruments in the new Envisat satellite seem
promising due to the multipolarization SAR and low resolution
Optical instrument MERIS. So, the natural alternative to
enhance the information obtained using Envisat SAR is to fuse
it with MERIS. Unfortunately, our project has not yet received
any Envisat-images, so we have made our early experiments
using ERS SAR and MODIS-images.
The aims of this study are
* toevaluate the value of ERS SAR-intensity images in
land cover and tree species classification,
* toselect the best texture features for classificatiion
and their benefits,
* study different factors like the use of temporal data,
the effect of the soil of forest stand, the effect of the
age of forest stand,
* compare pixelwise and standwise classification, and
* study the possibilities to enhance the classification
with low resolution MODIS data.
2. TEST AREA AND DATA
Test area used in this study is situated at Hyytiülà in middle
Finland (Lat. 61d50'N, Long: 24d22'E). Hyytiülà has a forestry
research station which belong to the Faculty of Forestry and
Agronomics at Helsinki University. The area is covered by
standwise forest inventory.
2.1 Satellite images
The SAR dataset consists of 8 ERS-2 SAR intensity images
taken during 1999. The image resolution is 25 meters, and the
raw images have a pixel spacing of 12.5 meters. The
characteristics of of ERS-images are illustrated in table |
(Kramer, 1996). The DEM produced by National Land Survey
used in the image processing has a pixel size of 25 meters. The
image processing was conducted using Gamma Ltd. Software. It
performs topographic correction with a very high positional
accuracy (Wegmüller et al. 1998). The images were averaged
from 12.5 meter to 25 meter pixel spacing after the image
processing.
The optical dataset consists of 176 MODIS-spectrometer
images (table 1) taken during 2000. MODIS-spectrometer has
36 channels (Masuoka et.al., 1998) with three different spatial
resolution levels. Channels 1 (red: 0.62 — 0.67 um) and 2 (nir:
0.841 — 0.876 um) were used due to their relatively high spatial
resolution (250 meters) compared to other channels.
Data ERS (microwave) MODIS (optical)
Wavelenght 5.7 cm (C-band, 1: 0.62 — 0.67 um
5.3 GHz) 2: 0.841 — 0.876 um
Number of 176: 1.4. — 30.9.2000
images and
8:31.3,164.,55.,
9.6. 14.7, 8.10.,
acquisition 27.10. and
times 12.11.1999
Spatial res. (m) 25-30 250
Table 1: Overview of the acquired satellite images.
