bul 2004 
mole CLASSIFICATION OF VEGETATION AND SOIL 
USING IMAGING SPECTROMETER DATA 
neral 
rkey, 
J. H. Lumme 
efl Institute of Photogrammetry and Remote Sensing, Helsinki University of Technology, P.O.Box 1200, FIN-02015 HUT, Finland 
‚Key, - juho.lumme@hut.fi 
Commission VII, WG VIII 
neral 
irkey 5 : ; ; ; : 
7 KEY WORDS: Remote Sensing, Hyper Spectral, Land Cover, Classification, Pushbroom 
eas ABSTRACT: 
nsed 
dings ou Es : : s : 
= Monitoring the Earth using imaging spectrometers has necessitated more accurate analyses and new applications to remote sensing. 
New algorithms have been developed for hyperspectral data classification lately, but also traditional classification algorithms have 
in often been used. This study compares different classification algorithms for classification of vegetation using imaging spectrometer 
Seed data. The test area located in southern Finland was imaged by an AISA airborne imaging spectrometer using 17 visual and near 
Chan infrared bands. The area included lakes, rural areas, cultivated fields and forests. The area was classified into seven different 
of vegetation and soil types. The effects of various classification algorithms and different training areas were investigated. Besides, the 
reflectance spectra of different plants were examined and compared under varying illumination. Spectral Angle Mapper (SAM), 
Spectral Correlation Mapper (SCM) and Spectral Unmixing algorithms developed for hyperspectral data were used in the 
near classification. Besides, the data was classified using conventional algorithms as Minimum Distance and Maximum Likelihood 
prole classifiers that have often been used for multispectral data classification in the past. The dimension of the data was decreased by 
principal component analysis before using conventional classifiers. Reference spectra for SAM, SCM and Spectral Unmixing were 
collected from the training sites of the data. Two methods were used in gathering the reference spectra. The reference spectra were 
and chosen from the reflectance of individual image pixels or they were calculated from pixels of the training sites. If individual pixel 
orne was chosen accurately, it led to better classification results. Maximum Likelihood classifier led to good results as well, but it 
ER). requires more computation time. The overall accuracy of the Maximum Likelihood classification was 91 percent, but the results 
sing, deteriorated under varying illumination. SAM and SCM were faster and they led to better classification results in poor illumination. 
The hardest part in Spectral Unmixing classification was finding suitable reference spectra from mixed pixels. When the essential 
spectra were found, the classification led to good results, although the results varied between different classes. 
's for 
| the : 
). pp. 1. INTRODUCTION methods are needed to reduce hyperspectral feature space 
dimension. 
Imaging spectrometers have been developed very rapidly over 
ical past decades. They have more channels with better spatial and Feature extraction methods may remove small differences in the 
Asian spectral resolution. Individual bands are only a few nm wide spectra between various materials. Sometimes these differences 
while the spatial resolution is good as well. Besides, computer characterize different materials and play an important role in 
number-crunching power, data-transfer rate and storage the classification. When using a hyperspectral classification 
capacity have increased considerably in recent years. This has algorithm like Spectral Angle Mapper the feature extraction is 
made it possible to handle and analyse larger data sets acquired not always essential. This way the whole reflectance spectrum 
by an imaging spectrometer. There are some new algorithms of the material is taken into account. For instance, the effect of 
being developed to hyperspectral data classification at the different soil types on the reflectance spectrum of vegetation is 
moment. slight and it comes up only with some specified wavelength 
bands. 
Most hands-on applications of imaging spectrometry relate to a 
sensor mounted on aeroplane. Satellite sensors do not have as Different institutions use imaging spectrometer data for various 
good spectral and spatial resolution as aeroplane sensors, which research projects in Finland. For example, workers in the 
may have hundreds of channels with one meter spatial Finnish Forest Research Institute are researching imaging 
resolution. Nowadays, satellite sensors tend to have tens of spectrometer data for inventory of forest resource (Mäkisara 
channels with spatial resolution of tens or hundreds meters. and Tomppo, 1996). Hyperspectral data is also used for 
monitoring the quality of water (Kallio et al, 2001) and 
Traditional algorithms like Maximum Likelihood or Minimum atmosphere at the Finnish Environment Institute. 
Distance have also been used in hyperspectral data 
classification. Unfortunately, they tend not to work properly The goals of this study were to investigate the suitability of 
with hyperdimensional data. Hyperspectral image data consists AISA imaging spectrometer data for vegetation and soil 
of hundreds of channels and it leads to extended run time. classification. Different classification algorithms were 
When channel number doubles, run time squares. Highly compared using various reference data and illumination. 
correlated channels may even crash the classification —— Besides, reference spectra of different materials were analysed 
programme. For these reasons effective feature extraction and material identification by its spectrum was investigated. 
83