Abstract
Land cover classification is often accomplished with a single multispectral image. However,
the informational utility of a multispectral image is limited by the spectral and spatial
resolution of the imaging system. Current imaging systems somehow offer a trade off
between high spatial and high spectral resolution - no existing single system can offer both of
these characteristics. In order to achieve both high spatial and spectral resolution in a single
image, image fusion may be employed. This paper examines the influences of image fusion
on thematic accuracy of land cover classification through an example using SPOT
panchromatic and multispectral images. Three fused images were generated using intensity-hue-
saturation (IHS), principal component analysis (PCA) and high pass filter (HPF) fusion
methods. All the images were then classified under the supervised classification approaches
of maximum likelihood classifier (MLC) and neural network classifier (NNC). Using the
classified result of the parent (original multispectral) image as a benchmark, the integrative
analysis of the overall accuracy indicated a certain degree of improvement in the
classification from using the fused images. The validity and limitations of image fusion for
land cover classification are finally drawn.
1. Introduction
To achieve accurate classification, a spectral image with a larger number of narrow
bandwidths spectral bands (high spectral resolution), is necessary. To gather image data with
high spectral resolution, a sensor with large-sized instantaneous field of view (IFOV) is
required to allow the necessary quantity of radiant energy to enter the detector elements. This
results in a sensor with low spatial resolution. The opposite case occurs for high spatial
resolution sensors. For example, an image coming from the spectrally oriented (seven-band)
LandSat TM sensor has 30-meter resolution and the image from the spatially oriented (1-metre
resolution) IKONOS-2 sensor is panchromatic. These fundamental sensor
characteristics directly affect the thematic and spatial accuracy of classification of a single
image. Multisensor fusion is a technique whereby single image that has the characteristics of
both high spectral and spatial resolution sensors can be generated.
Previous researches in this arena concentrated on developing fusion algorithms and their
assessment of its advantage was limited to visual enhancement. Experiments using the
integrated information of fused images were seldom repeated. In particular, the effects of
multisensor fusion in image classification have only been reported in Munechika et al (1993).
Here, a pair of SPOT panchromatic and LandSat TM images was used in the fusion process.
In testing the classification performance of the fused images, Gaussian maximum likelihood
classification was executed for five types of land cover and evaluation consisted of
comparison of the outcomes with ground truth data. The results indicated an enhanced
accuracy of classification using the fused image over using the individual (parent) images.
However, alternative types of images, fusion methods, levels of classification and classifiers
were not evaluated.
This paper investigates the influences on classification of the multisensor fusion of
(relatively) high spatial resolution SPOT panchromatic (PAN) and (relatively) high spectral
resolution SPOT multispectral (XS) images. Three fusion approaches - intensity-hue-saturation
(IHS), principal component analysis (PCA) and high pass filter (HPF) - were used
to generate the fused images containing both high spatial and spectral resolutions for
classification. The fused (child) images and the original multispectral (parent) image were
classified under the supervised classification approaches of maximum likelihood and neural
network classifiers. The classifications for USGS level-one natural land cover and level-two
cultural land use were performed for multilevel-based analysis. Instead of visual inspection,
the classification results with respect to the overall accuracy from the confusion matrices
were assessed. Using the parent image as a benchmark, the influences of classification using
the fused images were evaluated.
2. Experimental Testing
In this study, all the image processing operations (multisensor fusion and image
classification) were carried out on a Digital Ultimate workstation model 533AU2 using PCI
ImageWorks and Xpace version 6.01.
2.1 Test Data
The experiment in this study used a pair of image: SPOT panchromatic (PAN) and
multispectral (XS) images. The PAN image, was dated 92/01/31 and the XS 91/12/20. To
reduce the image size to be manipulated and to maintain geographic consistency during
fusion, equivalent sub-scenes as shown in Figures 1a (PAN) and 1b (XS) were defined.
 |
| (a) |
(b) |
Figure 1. Original image data: (a) SPOT PAN image; (b) SPOT XS image.
