Using textural information for classification of remote sensing imagery

KAPLAN Věroslav
Institute of Geography, Faculty of Science, Masaryk University in Brno, Kotlářská; 2, 611 37 Brno
2935@mail.muni.cz

Most of current software used for classification of remote sensing imagery still uses per-pixel approach. This approach is applicable only to multispectral imagery and is not suitable for processing of single band imagery. We can use other characteristics than only pixel values for single band data to achieve acurrate procesing, e.g. texture measures like homegenity, entropy, energy.

Texture measures ar simple yet powerfull concept which allows us to increase accuracy of per-pixel classification and even makes it possible to classify single band imagery.

The paper is focused on computing texture measures and their usage for classification of single band data. New “bands” are derived from original data by applying textural measures and both original imagery and derived one are classified by conventional per-pixel classifiers.

Introduction

Data and imagery acquired by remote sensing techniques play an important role in current cartographic practice. As spatial resolutions of todays sensors improves, geographers can use more and more advanced methods of imagery processing, especially for classification.

The most simple imagery classiffication approach is supervised per-pixel classification. Given pixel from n-band imagery with DN values <x1x2, ..., xn> from Rn (where xi is DN value of pixel in ith band), then pixels values (signatures) are equal to items in n-dimensional subset of vector space and classes of classification can be described as small disjunct subset of this vector space. Pixel is classified into class if it belongs to the class (in terms of vector spaces).

Typical classification scenario: User (teacher, supervisor) selects training areas on the processed imagery, the processing software then extracts spectral signatures for each class and then uses these signatures for classification. The processed image is classified per-pixel – each pixel is classified separately with no relation to any other pixel.

Per-pixel classification has lot of advantages:

Nevertheless, there are still some noticeable problems related to per-pixel approach:

Texture measures

There were some attempts to solve problems of simple per-pixel approach. One simple but still powerfull concept – texture measures was introduced in [Haralick, 1973]. While this concept wasn't designed primarily to remote sensing area but to computer graphics, it's still definitely usable to process remote sensing data. The usage of texture measures in remote sensing area was developed in many later papers [Haralick, 1979] and others.

We use many textural measures in "common life" – eg. by saying "silky" we mean "there are no large differences between high and low points of the surface". Textural measures allow us to characterize properties of surface more precisely.

Each measure can be defined as function f(X, I)R, X being pixel from original image I. Therefore, measure is computed for each pixel separately creating a new "pseudo-band". This "band" can be later used for per-pixel classification as a additional source of information. Using textural measures, we can improve accuracy of multispectral imagery classification or even make possible to classify panchromatic data.

Texture measures are computed from pixel's neighbourhood. This technique uses sliding window: for each processed pixel, program creates a "window", computes measures for center pixel of the window, saves computed measure to the new band and advances the window to next pixel. (Fig.1).

Fig.1: sliding window (3 × 3 in size)
central (red) pixel acquires computed measure

The size of the window, ie. number of pixel involved in measure computing is very important. The window has to be square and odd-sized with the processed pixel in the middle of the window. The bigger the window is the more pixels are involved in texture measure. In case of too large window, measure computed for the pixel is not well related to the central pixel. So choosing the right size of window can be a bit non-trivial. (Fig 2.) Usually is used window 3×3.

Contrast
Window size 3×3
Contrast
Window size 5×5
Contrast
Window size 7×7
Contrast
Window size 9×9
Fig 2: impact of window size on computed metric (Contrast)
images will enlarge after clicking.

Simple (1st order) measures

Simple measures are calculated only from DN values of the pixels involved without analyzing inter-pixel relations . They are the same as statistical measures (and related to bitmap operations from area of computer graphics area).

Source image
Source image
mean
Mean
Standard deviation
Standard deviation
Variance
Variance
Fig 3: 1st order measures

GLCM texture measures (2nd order)

To compute "real texture measures" we have to take into account relations between processed pixels.

The most common is definition by Gray Level Correlation Matrix. (GLCM). GLCM is defined for each pixel's neighbourhood as a square matrix, which contains counts of neighbourding DN values. Size of matrix is defined by number of DN values (radiometric resolution) of the whole image (ie. the better radiometric resolution, the larger matrix). In order to achieve faster computations it is recommended to reduce radiometric resolution of acquired images.

The GLCM approach is not the only way how to characterize textures, but it is the most implemented one (eg, in GRASS, PCI Geomatica, Envi and other).

Main GLCM measures are:

Source image
Source image
contrast
Contrast
Dissimilarity
Dissimilarity
Homogenity
Homogenity
ASM
ASM
entropy
Entropy
GLCM mean
GLCM mean
GLCM
variance
GLCM variance
GLCM
variance
GLCM correlation
Fig 3: 2nd order measures

Not all measures are suitable for classification, as some of them are related (and correlated) to others. It is recommended to use only one measure from each group [Hall-Beyer].

Advantages of GLCM textures:

Problems:

Application to the panchromatic imagery

I have tried to use texture measures to classify SPOT-3 panchromatic image. I used picture of Brno Dam Lake with following target classification classes: fields with and without vegetation, woods, urban and water areas. (Fig.4)

Source image
Source image
Source image with marked classes
Training data
Result of classification
Result of classification
Fig 4: classification data and result

Overall accuracy of classification is low – 85.7%. Taking into account the fact that the single band image is not classifiable by conventional methods, it can be considered as an interesting result. In case of multispectral images it would be possible to increase overall accuracy of classification.

There is still possibility to achieve better results - eg. by computing more measures and using principal components.

Conclusion

Textural measures were mentioned as simple, but still powerfull concept which can improve accuracy of remote sensing imagery classification. It can even make possible to classify single-band data which are unclassifiable with conventional per-pixel classifiers. Increased memory and time requirements are costs of textural measures. While there are also other techniques for classification of panchromatic images, textural measures are still a valuable concept.


References:

  1. [Dobrovolný, 1998] – Dobrovolný, P. Dálkový průzkum Země - digitální zpracování obrazu. MU Brno
  2. [Hall-Beyer, 2000] – Hall-Beyer, M. GLCM Texture a tutorial. (available online, October 2005)
  3. [Haralick, 1973] – Haralick, R.M.,Shanmugan, K., and Dinstein, I. Textural Features in Image Classification. In IEEE Transaction on Systems, Man and Cybernetics
  4. [Haralick, 1979] – Haralick, R.M., Statistical and structural approaches to texture, Proceedings of the IEEE