TA 
UNMIXING WETLAND VEGETATION TYPES BY SUBSPACE METHOD 
USING HYPERSPECTRAL CASI IMAGE 
Yoshiki YAMAGATA, Yoshifumi YASUOKA 
National Institute for Environmental Studies 
16-2 Onogawa Tsukuba Ibaraki JAPAN 
yamagata@nies.go.jp 
Commission V, Working Group 5 
XVIII ISPRS Congress Vienna, 9-19 July 1996 
KEY WORDS: Unmixing, Wetland Vegetation, Hyperspectral , Subspace Method 
ABSTRACT: 
A new approach to unmixing with subspace methods is proposed and an experiment using hyperspectral 
images was conducted. In subspace method, unmixing is calculated as the projection of each unknown 
pixel vector on the subspace of each class. This method is more stable than conventional methods against 
noise in the data and works effectively as a feature extraction and data reduction procedure as well. The 
performance of this method was tested by an unmixing experiment using a hyperspectral airborne CASI 
image acquired over the Kushiro wetland in NE Japan. Unmixing for the 7 wetland vegetation classes 
were calculated using a least squares, quadratic programming, orthogonal subspace projection and the 
subspace method. Finally, the results of unmixing experiment were evaluated in regard to wetland 
vegetation monitoring. 
environmental monitoring (Kramer, 19923. 
1. INTRODUCTION Hyperspectral imaging is recognized as an 
effective means for estimating vegetation 
In wetland landscape, various vegetation types parameters (Gong et.al., 1994). 
are continuously distributed. Remotely sensed To unmix the very large number of channels in 
spectral data over wetland areas are spectral ^ hyperspectral imagery, it is necessary to establish 
mixtures of several vegetation types. These an algorithm which can unmix several vegetation 
images consist of mixed pixels (Mixel) which have types in a fast and stable manner. A number of 
to be analyzed using spectral unmixing unmixing methods designed for band selection, 
procedures to estimate the state of each of the feature extraction and dimension reduction which 
constituents (Settle and Drake, 1993). incorporate modern signal processing and neural 
Conventional statistical unmixing methods such network methodologies, have been explored 
as least squares use a linear mixel model. In this recently (Harsanyi and Chang, 1994; 
model, the mixed spectral vector is assumed to be — Benediktsson et. al., 1995). 
a sum of class spectral vectors which constitute the Unmixing by the subspace method (Oja, 1984) 
mixel. By solving this linear mixel model with the ^ utilized in this paper is a new approach, based on a 
pre-determined class vector, an estimate of the fundamentally different principle from 
fractional area of each class within the pixel. conventional methods. The subspace method 
However, the computational complexity ^ assigns a different subspace to each vegetation 
increases substantially as the number of image class instead of fitting a mixel model in a pre- 
channels increases and the least squares solution determined number of spectral dimensions. 
becomes unstable due to the high auto-correlation Unmixing is then performed by measuring the 
between the channels. It is necessary to reduce the projection length of mixel vector. In addition, 
spectral dimension dimensions of the problem asa subspace method unifies the process of feature 
preprocessing of unmixing (Malinowski, 1991). extraction and unmixing, which are usually 
Hyperspectral sensors are a recent development ^ separate processes in conventional methods. 
in remote sensing and have been used for In this paper, the principle of the new unmixing 
781 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996