A Flexible Tool to Extract Algorithmic Knowledge From Image Interpretation Experts

Paul Crowther and Jacky Hartnett
School of Computing, University of Tasmania, PO Box 1214, Launceston, Tasmania 7250, Australia
Email: P.Crowther@utas.edu.au


Interpretation of multi-band satellite images often requires the combination of various bands to highlight particular features. Hard coding of the algorithms required to carry out these combinations into an image analysis system reduces flexibility and increases complexity. A tool which allows a user to easily enter, name, store and reuse algorithms is required.

KAGES (Knowledge Acquisition for Geographic Expert Systems) has such a tool called the Band Calculator. This allows a user to combine image bands from SPOT HRV, Landsat or NOAA AVHRR images and immediately see the result of applying a band combination algorithm which has been interactively entered as an equation. The algorithm can then be named and stored for reuse or can be discarded without having to hard code it into the system. Once the algorithm has been stored, it can be retrieved and used in other classification systems and the results of applying it can viewed at any time.

Keywords: Remote Sensing, Expert Systems, Spatial Analysis,

1. Introduction

Expert system methods for image classification involve the use of heuristics, ancillary data and facts (Wilkinson, 1996). One problem with the use of expert systems is the acquisition of knowledge from human domain experts. This has been termed the knowledge acquisition bottleneck by many authors (for example Cullen and Bryman, 1988). The problem is further compounded when the knowledge involved is visual rather than semantic. KAGES was developed primarily to acquire visual knowledge (Crowther and Hartnett, 1997). Not all spatial knowledge is visual however and it has been found that tools to acquire non visual knowledge are also required.

To interpret multi-band satellite images, interpreters often combine bands in various ways to highlight particular features. Sometimes bands from several image sets are also combined. Common algorithms used are those for calculating Normalized Differential Vegetation Index (NDVI) and temperature, but the particular algorithm(s) used depends on the subject of the classification. The number of available algorithms is such that hard coding each of them into an automated system restricts the range available and increases complexity. Each result has to be stored as an extra dimension in an image processing package or layer in a GIS. If several of these band combination algorithms are required the data structure involved in storing all the layers becomes quite large and unwieldy.

A solution to this problem is a tool which allows an image interpreter to combine raw bands as they work and interactively define the algorithm which they want to apply. When a suitable band combination is found, the characteristics of the feature identified along with the algorithm used to combine bands is stored. This knowledge can be retrieved and reapplied when necessary. Hence only the raw bands of the image need to be permanently stored at any time, along with result layer representing classified features.

2. Geographic Knowledge

Geographic knowledge differs from knowledge used in non-spatial expert systems in that domain experts primarily use knowledge that is visually oriented. Often multiple experts with different expertise are required to interpret images to make up a final composite image (Tranowski, 1990). There have been several papers which suggest how to identify and classify knowledge in geographic and spatial systems. For example McKeown et al (1989) identifies five levels of spatial knowledge.

Armstrong (1991), working on generalization, defines three level knowledge classification scheme which includes a non visual class, Procedural Knowledge. This is knowledge of individual generalization operators and algorithms.

The classifications of McKeown et al and Armstrong are not mutually exclusive, rather they are complimentary. The McKeown et al classification appears to expand on Armstrong’s geometrical and structural knowledge classifications. Procedural knowledge is an essential component when image processing systems are considered and is not covered in the McKeown et al scheme.

The following classification is derived from and expands on those of McKeown et al and Armstrong. It is more rigorous and incorporates non visual knowledge:

 KAGES implements facilities for extracting geographic knowledge according to this classification. This classification is compatible with the KADS (Knowledge Acquisition and Development System) methodology (Wielinga et al, 1992) which provides a more general framework for knowledge acquisition for expert systems.

The primary concern of this paper is algorithmic non visual knowledge. Discussion of the visual tools can be found in Crowther et al (1997) and of the heuristic tool in Crowther and Hartnett (1996).

Algorithmic knowledge in this context is the way satellite bands can be combined to produce a composite image. It differs fundamentally from the visual knowledge types in that it is knowledge which does not directly work with scene primitives.

3. The Band Calculator

The KAGES system, written in Research Systems Inc.’s IDL, uses a graphical user interface and captures expert user actions and classification methods. One of the tools is the Band Calculator which can be activated once a image data set is loaded.

The KAGES Band Calculator (Figure 1) is a customized pop up window which has the appearance of a hand held calculator. It differs in having extra buttons representing satellite image bands and a window for entering the algorithm name. The number of these buttons is variable and depends on the number of bands of the image being loaded. For example a 5 band NOAA AVHRR image set would cause five buttons to be displayed, while loading a SPOT HRV image would cause three to be displayed.

Figure 1: The KAGES Band Calculator showing NDVI entered for a SPOT HRV satellite image.

Any combination of bands can be entered via the calculator. Once an algorithm is entered, the user can then name the algorithm. The algorithm is then applied to the image set and a resultant composite image and histogram are displayed (Figure 2). The algorithm is stored along with its name.

Figure 2: Results of applying NDVI algorithm to a December 1996 SPOT HRV image set of the North West Coast of Tasmania

Algorithms can be activated when the visual tools in the toolkit are selected. A list of satellite bands and available algorithms are displayed and the user can select the desired one. If a scene primitive is identified using a combination of bands the algorithm for creating the combination is stored as part of the rule. When other tools such as that for determining Relationship Knowledge is activated, the algorithm is automatically applied.

The band calculator is therefore used to build a list of algorithms for use during knowledge acquisition. The knowledge is general in that the file containing it can be used in several different projects. Hence once NDVI has been defined for SPOT images, it does not need to be done again, rather the file containing the SPOT algorithms can be reused.

4. Discussion

Most interpreters of remotely sensed satellite images work with image band combinations to highlight particular features. If a GIS is not being used these are often included directly in image processing code or are placed in a fixed menu system. New algorithms therefore require code updates and associated re-testing of the overall system. A second problem with algorithms hard coded into a program is that they may lack documentation. Finally it may be impossible to share an algorithm over several applications without re-coding each of them.

The Band Calculator approach can overcome these problems as the algorithm development and testing is independent from the application in which it is to be used. Since the algorithms are stored as a separate file and not as new image layers, they can be applied to other images.

KAGES currently only allows the loading of data from one satellite image set. A planned enhancement is to allow more than one set to be active at one time. When this is done, the Band Calculator could combine information from different sources. This will involve very little modification to the calculator, since new buttons can be added very easily. This would add further flexibility to the tool by allowing algorithms such as combining NDVI from both SPOT HRV and Landsat.

The technique has efficiency benefits. The resultant composite image created by the band calculator is not stored permanently as an extra layer with the image raw bands, rather it is recalculated when required. This has the effect of saving disk storage space, saving memory and simplifying data structures. Hence if several band combinations are required in a study, only those which are immediately needed are stored in memory.

5. Conclusion

Most applications using remote sensing satellite images require combinations of bands. The algorithms for these combinations should be easy to enter and once entered, independent of specific applications.

A tool such as the Band Calculator which operates using a graphical user interface is a solution. The algorithm can be entered, the results viewed, and the algorithm stored. Once stored the algorithm can the be applied to other images from that satellite.


The work reported in this paper has been supported by an Antarctic Science Advisory Committee Research Grant and a Horticultural Research and Development Grant.


Armstrong, M. P, 1991, Knowledge Classification and Organization in Map Generalization: Making Rules for Knowledge Representation, Buttenfield B.P. and McMaster, R.B., Longman Ch 5 pp86-102

Crowther, P. and Hartnett, J. 1996, ‘Using Repertory Grids for Knowledge Acquisition for Spatial Expert Systems’. Proceedings of the 4th Australian and New Zealand Conference on Intelligent Information Systems. Adelaide

P. Crowther, J. Hartnett and R. N. Williams 1997, ‘A Visual Tool For Capturing The Knowledge Of Expert Image Interpreters: A Picture Is Worth More Than A Thousand Words’. IGARSS’97, Singapore

Crowther, P. and Hartnett, J. 1997, ‘Eliciting Knowledge With Visualization - Instant Gratification For The Expert Image Classifier Who Wants To Show Rather Than Tell’. Proceedings of GeoComputation 1997, Otago, New Zealand

Cullen, J. and Bryman, A., 1988, The Knowledge Acquisition Bottleneck: Time for Reassessment, Expert Systems, Vol 5, No 3, pp 216-225.

McKeown, D. M., Harvey, W.A. and Wixson, L.E., 1989, Automating Knowledge Acquisition for Aerial Image Interpretation, Computer Vision and Graphics, Vol. 46, pp. 37-81.

Tranowski, D. 1990, A Knowledge Acquisition Environment for Scene Analysis, in The Foundations of Knowledge Acquisition - Knowledge Based Systems Vol. 4, Boose, J. and Gaines, B. (eds.) Academic Press, pp. 339 - 355.

Wielinga, B. J., Schreiber, A. T. and Breuker, J. A., 1992, KADS: A Modelling Approach to Knowledge Engineering, Knowledge Acquisition, Vol 4, No 1.

Wilkinson, G. G., 1996, A review of current issues in the integration of GIS and Remote Sensing Data International Journal of GIS, V10 No 1 pp 85-101.