04 
cal 
illy 
ital 
SD 
age 
pill 
Dy. 
tant 
EFFECTS OF UNSUPERVISED FULLY CONSTRAINED LEAST SQUARES LINEAR 
SPECTRAL MIXTURE ANALYSIS METHOD ON AUTOMATIC CLASSIFICATION OF 
TM IMAGE 
Hong-xia LUO * **, Jianya GONG", Jianping PAN? 
a State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing ,Wuhan University, 
129 Luoyu Road, wuhan. Hubei, China-geogjy@163.net 
b School of Resources and Environment, Southwest Normal University, 2 Tiansheng Road, Chongqing, China - 
tam 7236(a)swnu.edu.cn 
KEY WORDS: classification of remote sensing image, linear spectral mixture model, fully constrained least squares (FCLS) 
algorithm, unsupervised FCLS (UFCLS) method, constrained optimization problem 
ABSTRACT: 
Statistical analysis is a widely used traditional technique for classification and discrimination in remotely sensed images, which 
supposes that there is only one endmember in an image pixel. However, the fact is that the ground sampling distance is generally 
larger than the size of targets of interest. So the statistical analysis technique is not suitable. In this case, classification and 
discrimination must be carried out at subpixel level. In this paper the abundance fractions of endmembers in an image pixel are 
estimated by UFCLS (unsupervised fully constrained least squares) method based on the inversion of linear spectral mixture model. 
This method allows us to extract necessary endmember information from an unknown and no prior knowledge image scene so that 
the endmembers present in the image can be quantified. The pixel classification generates a gray scale image, whose gray level 
values are determined by the estimated abundance fractions of endmembers. The band expansion technique is used to create 
additional bands from existing multispectral bands using band-to-band nonlinear correlation. These expanded bands ease the 
problem of insufficient bands in TM imagery. In the two experiments, the results of the pixel classification show that the effects are 
good. The pixel classification image of vegetation agrees significantly with the NDVI image, but the contrast of the former is a little 
larger than that of the latter, so there was lack of the information of details and edges. However, compared to color composite image 
of raw bands 4, 3 and 2 in red, green and blue respectively, especially in the second experiment, the results of vegetation 
classification are excellent. The shade areas in the first experiment are not classified correctly. Compared to CSMA (constrained 
spectral mixture analysis) method, UFCLS method is better in both the effects of classification and the consumption of computation 
time. 
1. INTRODUCTION Currently, the approach mostly used to decompose the mixed 
pixels by linear model, also be called model inversion, is the 
Statistical analysis such as maximum likelihood, minimum least squares method. If considering it from the point of 
distance and mahalanobis distance, is a widely used traditional minimizing the merit function, this is a constrained 
technique for classification and discrimination in remotely optimization problem because the parameters of the linear 
sensed images, which differentiates pixel one by one according model generally have explicit physical meaning and constraints. 
to the average spectral signature of materials, is a classification The optimization algorithms include the regularly searching 
method of pure pixel discrimination. These traditional statistical algorithms such as downhill simplex, conjugate direction set 
classification and discrimination methods suppose that there is and passive set, and the randomly searching algorithms 
only one endmember in an image pixel. However, the fact is proposed in the recent years such as GA (genetic algorithm), ES 
that a pixel is generally mixed by a number of endmembers in and EP. However, the latter is not better than the former for the 
the image given spatial resolution. Consequently, mixed pixel inversion of the linear spectral mixture model, and cost longer 
analysis techniques such as spectral mixture model have been computation time (Tang Shihao et al., 2002). The regularly 
proposed to describe such mixing activities. The mixed pixel searching algorithms are frequently used in the inversion of 
classification generally generates a gray scale image whose remote sensing models currently. Unsupervised fully 
gray level values are determined by the estimated abundance constrained least squares (UFCLS) recently proposed by Daniel 
fractions of the endmembers resident in the image pixels. et al. (2001) improved the computation speed, which applied 
Although there are many spectral mixture models, they all the method based on the least squares and resembling the 
belong to two classes, linear and nonlinear models respectively. passive set of the regularly searching algorithms. Under the 
To this day, the linear spectral mixture model is favorably circumstance without prior knowledge, this method is valid for 
received and most widely used. Its prominent characteristic is decomposing mixed pixels in remote sensing images. In this 
very simple. It was reported by Zhao Ying-shi (2001) that the paper, we analyze the effects of UFCLS on classification of TM 
linear spectral mixture model was better than the tasseled cap image. 
transformation when used to extract sands information. 
  
" Correspondence to: Hong-xia LUO 
1181