CHAPTER 11
An MRF Model Based Approach
for Sub-pixel Mapping from Hyperspectral Data
Teerasit Kasetkasem, Manoj K. Arora, Pramod K. Varshney
11.1
Introduction
Image classification is a key task in many remote sensing applications. As
discussed in Sect. 2.6 of Chap. 2, the objective of classification is to allocate
each pixel of a remote sensing image into only one class (i. e. hard or per-pixel
classification) or to associate the pixel with many classes (i. e. soft, sub-pixel
or fuzzy classification). A number of hard classifiers are in vogue based on
approaches such as statistical (Mather 1999), neural networks (Foody 2000a)
and decision tree (Hansen et al. 2001).
However, in general and particularly in coarse spatial resolution images
such as those obtained from hyperspectral MODIS and multispectral AVHRR
sensors that provide data at spatial resolutions ranging from 250 m to 1.1 km,
a majority of pixels are mixed at the scale of measurement. Even where the
spatial resolution is medium (e. g. 30 m Landsat ETM + sensor) or fine (e. g.
4 m IKONOS multi-spectral sensor), the high spatial frequency of some classes
may result in a large number of mixed pixels (Aplin and Atkinson 200!).
The causes for the occurrence of mixed pixels are well known (Foody and
Cox 1994; Foody and Arora 1996) and their presence is a recurring problem
for extracting accurate information from remote sensing images. Therefore,
sub-pixel classification may be more appropriate than per-pixel classification.
In sub-pixel classification, a pixel is decomposed into a number of classes to
reflect their proportions contained within the mixed pixel. Some of the prevalent techniques used for sub-pixel classification are fuzzy c-means clustering
(see Sect. 2.6.4 of Chap. 2) (Bezdek et al. 1984), linear mixture modeling (Settle
and Drake 1993) and artificial neural networks (Foody 1996; Binaghi et al.
1999). Recently, support vector machines (Brown et al. 2000) and possibilistic
c-means clustering (Foody 2000b; Ibrahim et al. 2003) have also been used to
unmix the class proportions in a pixel. The result after sub-pixel classification
is a set of fraction images equal to the number of classes; a fraction image
describes the proportion of a particular class within each pixel. While, under
most circumstances, classification at the sub-pixel level is meaningful and informative, it fails to account for the spatial distribution of class proportions
within the pixel (Verhoeye and Wulf2002). A limited amount of work has been
reported in the literature (e.g. Atkinson 1997; Foody 1998; Tatem et al. 200!)
P. K. Varshney et al., Advanced Image Processing Techniques for Remotely Sensed Hyperspectral Data
© Springer-Verlag Berlin Heidelberg 2004
An MRF Model Based Approach
for Sub-pixel Mapping from Hyperspectral Data
Teerasit Kasetkasem, Manoj K. Arora, Pramod K. Varshney
11.1
Introduction
Image classification is a key task in many remote sensing applications. As
discussed in Sect. 2.6 of Chap. 2, the objective of classification is to allocate
each pixel of a remote sensing image into only one class (i. e. hard or per-pixel
classification) or to associate the pixel with many classes (i. e. soft, sub-pixel
or fuzzy classification). A number of hard classifiers are in vogue based on
approaches such as statistical (Mather 1999), neural networks (Foody 2000a)
and decision tree (Hansen et al. 2001).
However, in general and particularly in coarse spatial resolution images
such as those obtained from hyperspectral MODIS and multispectral AVHRR
sensors that provide data at spatial resolutions ranging from 250 m to 1.1 km,
a majority of pixels are mixed at the scale of measurement. Even where the
spatial resolution is medium (e. g. 30 m Landsat ETM + sensor) or fine (e. g.
4 m IKONOS multi-spectral sensor), the high spatial frequency of some classes
may result in a large number of mixed pixels (Aplin and Atkinson 200!).
The causes for the occurrence of mixed pixels are well known (Foody and
Cox 1994; Foody and Arora 1996) and their presence is a recurring problem
for extracting accurate information from remote sensing images. Therefore,
sub-pixel classification may be more appropriate than per-pixel classification.
In sub-pixel classification, a pixel is decomposed into a number of classes to
reflect their proportions contained within the mixed pixel. Some of the prevalent techniques used for sub-pixel classification are fuzzy c-means clustering
(see Sect. 2.6.4 of Chap. 2) (Bezdek et al. 1984), linear mixture modeling (Settle
and Drake 1993) and artificial neural networks (Foody 1996; Binaghi et al.
1999). Recently, support vector machines (Brown et al. 2000) and possibilistic
c-means clustering (Foody 2000b; Ibrahim et al. 2003) have also been used to
unmix the class proportions in a pixel. The result after sub-pixel classification
is a set of fraction images equal to the number of classes; a fraction image
describes the proportion of a particular class within each pixel. While, under
most circumstances, classification at the sub-pixel level is meaningful and informative, it fails to account for the spatial distribution of class proportions
within the pixel (Verhoeye and Wulf2002). A limited amount of work has been
reported in the literature (e.g. Atkinson 1997; Foody 1998; Tatem et al. 200!)
P. K. Varshney et al., Advanced Image Processing Techniques for Remotely Sensed Hyperspectral Data
© Springer-Verlag Berlin Heidelberg 2004