Satellite Data Processing and Image Analysis

D. A. QUADIR
Chief Scientific Officer
Bangladesh Space Research and Remote
Sensing Organization (SPARRSO),
Agargaon, Sher-e-Banglanagar,
Dhaka 1207, Bangladesh.

1. INTRODUCTION

Digital image processing is a subject that deals with the manipulation of images by computers. This technique has widely been used for processing the remotely sensed data in the digital form. The digital image processing requires special hardware and software, the collection of which is called the ‘Image Processing System’.

The sensors in the space platforms scan the earth in different bands of the electromagnetic spectrum and sense the irradiated energy of the earth’s surface. These data are subsequently transmitted to the satellite ground station in the digital form.

Certain preprocessing is needed to convert these data into - workable imagery. For practical purposes, it is necessary to manipulate the data in various ways as needed by the individual applications.

In the following sections we provide an overview of the image processing techniques and analysis of satellite data for environmental mapping and monitoring.

2. DIGITAL IMAGE

A digital image is a two dimensional array of integers (matrix). Each element of the matrix has a discrete value and refers to a small area with a horizontal distance (x) and vertical distance (y) which is called a picture element or pixel. The size of the pixel determines the spatial resolution of the image ie, a smaller pixel means higher spatial resolution. The position of a pixel is given by the line and column of the matrix and the integer value of the pixel represents its illumination or brightness. For most of the digital image the pixel values are represented by 8 bit number which produces 256 levels of gray scale.

The digital image stored in the computer is displayed on the display device, printing device or photographic films for human perception. The brightness distribution of the visual image is dependent on the gray values. The value zero means no illumination (black) and the value 255 means the maximum illumination (white).

In case of the satellite imagery, the pixel corresponds to the small area on the earth’s surface which is covered by the Instantaneous Field Of View (IFOV) of the sensor. The satellite sensors collect information from every such small area to make a whole image.

3. DIGITAL IMAGE PROCESSING

The image processing system is a collection of hardware and software having the facilities for input, processing and output of the digital images.

The multispectral satellite image processing and analysis highly depend on the nature of the application and also on the data to be used. The following image processing techniques are generally used for application of satellite imagery to various environmental mapping and monitoring.

The image processing functions which are commonly used for analyzing the satellite imagery are described in the following subsections.

3.1 Radiometric processing

The radiometric processing involves the following:

- Calibration of the digital counts produced at the sensor to an understable geophysical unit like irradiation (watt/m²-sr), percent reflectance or albedo and in the case of thermal radiation to the black-body temperature (°K)

- Correction for atmospheric absorption and scattering

- Sun-angle correction

- etc.

3.2 Geometric processing

Because of the earth’s rotation, the satellite orbit makes an inclination with the true north. Moreover, the image captured by the satellite sensor is the projection of the curved surface on the two dimensional plane.

Thus the geometric correction of the satellite images is to be done to convert the image geometry to a preferred cartographic projection. For the data of environmental satellites like NOAA - AVHRR, the geometric correction is commonly done by the orbital information of the satellite and the characteristics of the sensor. For the high resolution data from Landsat MSS, TM and SPOT - HRV and panchromatic, the geometric correction is done by referencing the image to the ground through a set of ground control points. The ground control points are obtained from the maps using a digitizer or from the measurements of Global Positioning System (GPS) at the ground.

The geometric correction is done by transformation of the coordinates of the original image to the new coordinates which is defined by the ground control points. The new points are expressed as functions of old points:

X = f₁ (x, y)
Y = f₂ (x, y)

The other functions that include the geographic processing are image rotation, displacement and image enlargement.

3.3 Image Enhancement

A digital image from the satellites may be enhanced to increase the contrast among the various features. In the initial step, the histogram of the intensity distribution is calculated. For a low contrast image, the histogram is concentrated within a small region of the gray scale. The principle of the contrast enhancement is to distribute the gray values of the original image to the full range of the gray scale ie, from 0 to 255. The following enhancement schemes are generally used for increasing the level of contrast.

Linear contrast stretching

The linear contrast stretching technique involves the mapping of the pixel values from the observed range to a output range specified by a linear transfer function.

Scaling of the histogram

In this technique the observed histogram is scaled over the whole range of the gray scale, where the overall shape of the histogram remains unchanged.

Histogram equalization

In this sophisticated technique, the histogram is distributed over the whole range of the gray scale (0 - 255) in such a way that each class in the output image has approximately the same number of pixels. In this process, the lower populated gray values at the beginning and at the end of the histogram of the input image are amalgamated to a fewer number of classes and the highly populated classes are placed away from each other for increasing the contrast. The histogram of the output image have more or less uniform distribution of pixel population.

Gaussian stretch

The gaussian stretch technique involves the fitting of the observed histogram to a normal gaussian histogram following the distribution defined by

f (x) = C e^{-ax 2}
where C = (a /p)^0.5 and a = 0.5 d²
d is the standard deviation of gray values of the image.

Piece - wise linear transformation

For some applications, it is needed to enhance the image by breaking the histogram into pieces and then applying the linear transformation on each part of the observed histogram, separately. Some of the image processing systems provide the facilities to enhance the image by the user interactively using trackball or mouse-button.

Stepwise transformation

The input image histogram may be transformed stepwise using a user - defined lookup table to a fewer number of classes. Sometimes a binary transformation of the input image is made at a user specified break point. In the transformed image there are only two highly contrasting gray levels.

Colour Enhancement

There are many techniques for colour enhancement of the imagery of which pseudo color transformation, density slicing and user defined colour transformation are common. The user interactively determines which of the colour enhancement is suitable for his work. Each provides a method for mapping from one dimensional gray scale to a three dimensional colour space defined by Red, Green and Blue axes. There are two colour models used for colour transformation, one is RGB colour cube model and the other is HSI model where RGB stands for Red, Green and Blue and HSI stands for Hue, Saturation and Intensity.

Density slicing

In density slicing, the histogram is divided into a number of classes or slices and the range of contiguous gray levels in each class is mapped to a point in the RGB colour cube.

Pseudocolour Transformation

A pseudocolour transformation is carried out by setting the three look up tables (LUT) (red, green and blue) from the observed one dimensional gray scale. Red - LUT corresponds to certain range of the input histogram along the one dimensional gray scale, Green - LUT corresponds to next range of gray scale in the histogram having some overlap with earlier range and Blue - LUT corresponds to the third range covering the rest of the gray scale having an overlap with its earlier one. After the transformation using these lookup tables a three band colour image is produced. Unlike the density sliced image, each pixel of the pseudocolour image has a discrete colour, although difference between 80% red and 85% red may not be physically discriminated in reality.

User - specified colour transform

In density slicing the user have the provisions to choose the classes and the colours for the respective classes. Say, for example, for mapping a warm surface, naturally the user would like to use red for the hottest class and yellow for the intermediate class and so on. Pseudocolour transform similarly provides the facilities within the image processing system to set the LUTs to convert the one dimensional grayscale to a colour representation of an image. The user can choose its own transform by changing the shape of the LUTs. Then the colour scheme is saved in the disk which may be repeatedly used when necessary. The user may produce the colour schemes for individual applications once for all and use the same any time it is needed.

4. MULTISPECTRAL AND MULTITEMPORAL IMAGE PROCESSING

Previous section has dealt with operations such as enhancement which are applied to single band images or separately to single bands of a multiband image set. In this section, we deal with operations on multiband images which may consist of a single multispectral image of a particular area or a number of images of the area taken at different times.

The multispectral image processing mainly consists of the following:

- False colour composite (FCC)

- Arithmetic operations between the bands or between the images of the same area.

- Multispectral classification

- Binary operations between the classified imagery of different times

- Calculation of image statistics

- etc.

The multispectral image processing depends on the kind of application and type of data used.

4.1 False Colour Composite

When 3 bands of a multi-band image are displayed in the monitor in the red, green and blue colour planes a False Colour Composite (FCC) image is prepared. Then the individual bands may be enhanced to increase the contrast of the image and bring up the features on which the user is interested. In a standard FCC, the near-infrared (NIR) band is displayed as red, the red band as green and green band as blue. As the vegetation has high reflectance in NIR band and low reflectance in the visible bands, it appears red in the standard FCC product. Bare soil has nearly same reflectance in all the bands and thus it appears dark to bright gray depending on the type of soil and moisture content. The drier soil appears brighter. The water absorbs almost all the radiations in NIR, while the visible light is partly reflected by the suspended sediment (or from the bottom of the shallow coast in case of clear water). Thus the turbid water seems blue in the enhanced FCC. The deep water with low suspended sediment appears dark.

For the Landsat MSS which has 4 bands, namely bands 4, 5, 6 and 7, the standard FCC is composed of bands 7 (red), 5 (green) and 4 which correspond to the false colours red, green and blue respectively. For Landsat TM, the same is done using the bands 4, 3 and 2. But there are other applications which might require the band composites with different other combinations. The LUT of the FCC enhancement schemes are saved and may be used if and when needed. The high quality photographic negatives or positives films and hardcopy paper products are produced in a definite scale. These imagery are then interpreted visually for landuse mapping/environment monitoring purposes.

The FCC of AVHRR bands 2, 1 and 4 represents the green vegetation as red, bare soil as light gray, coastal turbid water as green and clear water as dark.

4.2 Arithmetic operations

Arithmetic operations among the individual images or of one image with other images or constant numbers can be performed. Addition, subtraction, multiplication, division and logical operations or any other arithmetic operation may be performed as per user’s need. The arithmetic operations are carried on pixel to pixel. For example:

C (x, y) = A (x, y) + B (x, y)

This means, the values of the pixels of the resultant image is obtained by adding the values of the corresponding pixels of images A and B. More clearly:

C (1, 1) = A (1, 1) + B (1, 1)
C (1, 2) = A (1, 2) + B (1, 2)
.
.
.

C (m, n) = A (m, n) + B (m, n)

The mean of two imagery is given by

P (x, y) = {A (x, y) + B (x, y)}/2

The image operations of the type,

S (x, y) = a x A (x, y) + b x B (x, y) + C

can also be performed where a, b and C are constants. Some more examples are given below:

Image multiplication: M (x, y) = A(x, y) x B(x, y)

An example of algebraic expression of image manipulation:

where a, b, d, e and f are constant.

4.3 Vegetation indices

Besides the visual interpretations using FCC products as discussed in section 4.1, the digital data can be analyzed and interpreted using suitable arithmetic manipulations. Such digital analysis and interpretation begin with examination of the digitally enhanced FCC and the image statistics corresponding to the prominent features. Let us pick some samples of vegetation, bare soil (dry, medium, wet), water, etc. on the FCC image and calculate the mean, standard deviation and histogram of the individual classes. We draw the spectral information of these classes in a two-dimensional scatter-diagram with Red band as X-axis and NIR as Y-axis. It may be seen that the pixels are distributed around two lines: the soil line and the vegetation line. Approximately these two lines are orthogonal.

From this scatter-diagram, it is evident that the vegetation have high reflectance in NIR compared to soil and water and low reflectance in RED band. From such a diagram it may be envisaged that the ratio of NIR and RED bands would separate land from water and enhance the vegetation information. Such a ratio is called Ratio Vegetation Index (RVI) and is given by:

The interpretation of RVI is as follows:

RVI < 1	Water
2 > RVI > 1	Bare soil or soil with low coverage of leaf area
RVI > 2	Green vegetation; the higher the RVI, the more is the green vegetation coverage.

Higher the value of RVI, higher is the leaf-area coverage. RVI saturates nearly at the value 5.

Another vegetation index is called Normalized Vegetation Index (NVI). The NVI is expressed as:

The NVI is explained as the following:

NVI < 0	Water
NVI > 0	Land
.2> NVI > 0	Bare soil or soil with lower coverage of vegetation
NVI >.2	Green vegetation. Higher the NVI value higher is the vegetation coverage

The percent coverage of vegetation is dependent on the NVI value. The higher the NVI the higher is the vegetation coverage. The NVI is limited between -1 to +1. However, the NVI saturates at around 0.6.

The IR band corresponds to band 7 of MSS, band 4 of TM and band 3 of SPOT-HRV, and Red band corresponds to band 5 of MSS, 3 of TM and band 2 of HRV. The RVI and NVI may be calculated for each of the setypesof satellite data following equations 1 and 2. For better recognition of land water boundaries bare soil and different levels of vegetation coverage, the vegetation index imagery are colour coded using a predefined colour scheme. Such images of vegetation indices are then used for crop monitoring, preparation of vegetation maps, determination of green biomass, crop modelling, water-bodies survey, etc.

4.4 Image classification

Multispectral classification is a technique for automatic classification of the spectrally different features. The classified image is digitally interpreted using the classification statistics, scatter diagrams and the ground truth information. There are two general techniques for image classification: Unsupervised and Supervised classifications.

The unsupervised classification is the one which determines the characteristics of non-overlapping groups of pixels in terms of their spectral band values. These groups are therefore known as spectral classes, and their relationship with features must be worked out through field work. The overall feature types may however be interpreted from the position of the classes in the spectral domain.

The supervised classification is carried out by taking a training set, the feature relationship of the training set is already known. Then the statistical properties of each of the classes are studied and the pixels of the image are then allocated to one of those classes using a classification algorithm. It is likely that a number of pixels will not find their similarities with any of those classes or groups and will remain as unclassified.

There are several classifier algorithms of which the maximum likely hood and minimum distance classifiers are most common. The software packages provide many classifier algorithms, however, it is up to the user to decide which one of those is to be used.

The digital classification is a standard practice of the users of remote sensing techniques for analysis and interpretation of imagery and the interpreted results are produced as landuse/resource maps and also as tabular information. A properly geo-referenced and classified image with appropriate heading, legends and other required annotations can be transferred as the hard copy photographic product in a chosen scale.

The classified imagery at different seasons are used to remove the effects of interference between the various features, Say, for example, we have classified the Landsat TM image of February 27, 1991. Our objective was to identify the winter rice. But some of the rice areas were confused due to the presence of wheat, shrubs, forests and home-stead vegetations. For solving this problem, we chose the December 26, 1990 image when most of the fields were vacant. In this image, the wheat, mustard, pulses etc. were identified quite successfully. The home-stead vegetations and the forest vegetations and the shrubs were also identified. A detailed ground truth data was available for this winter. The December image was classified successfully using the supervised classification technique. Then the February image was also classified. These two classified images were then used to remove the confusions of rice with other classes of vegetation.

5. CONCLUSION

An overview of the satellite data processing and image analysis has been given in this paper. For the purpose of image processing, a digital computer, image processing hardware and software and the input and output devices are necessary. Without one or the other, it is not possible to use this technology satisfactory.

In SPARRSO, there are a number of powerful image processing systems (IIs system 575 and 600 and PC ERDAS) which allows lots of image processing activities. SPARRSO’s image processing capabilities can be used to conduct the country-wide mapping and monitoring of the environment and resources using remote sensing techniques.

The joint use of digital image processing and GIS enhances the user capabilities to conduct remote sensing studies reference to other spatial information.

SOME BIBLIOGRAPHIC REFERENCES

1. Castleman K. 1979	Digital Image Processing Prentice-Hall Signal processing Series, Prentice - Hall, INc, Englewood cliffs, New Jersey.
2. Colwell, R.N. 1983	Manual of Remote sensing, American Society of Photogrametry, Falls Church, Virginia.
3. Gonzalez R.C. and Wintz P. 1977	Digital Image Processing, Addison Wiley Publishing Company.
4.	Image Processing System, IIS System 600 (reference manuals), International Imaging Syste, Milpitas, California.
5. Mather P.M. 1987	Computer Processing of Remotely - Sensed Images - An Introduction, John Wiley and Sons, 352 p.
6. Quadir D.A, Ali, A. and Hue, O.K. 1989	A study of vegetation pattern in Bangladesh with AVHRR data, Asian-Pacific Remote - Sensing Journal vol., number 2 pp. 37-57.