Roeland de Kok
of which terrain knowledge is available, is correctly classified as a coniferous-middle stand. The critical band here is
mean value of band 4, SPOT —4 , infra red, giving a 92.3% class membership. The second class option, the object
belongs 38.6 % to the class coniferous-open (figure 5) . Also in this case, mean value of Band 4 from SPOT-4 is the
critical feature. Gaussian distribution can be assumed, explaining the shape of the fuzzy logic curves in figure 3. For
other features, such as standard deviation of the panchromatic band, the Gaussian assumption is not quite correct. In this
case a box classifier is better at place (see figure 2). Gaussian distribution can be checked in a 2-D plot of independent
object features (Figure 6). An explanation for the shape of the graph in figure 7; Gaussian distribution for panchromatic
data is corresponding to expectation for object mean value and standard deviation (range X=15-210).Mean value
depends on photo count, standard deviation is a measurement of surface roughness. There is no factor that allows to
assume that brighter objects are more smooth or vise versa. In very dark objects (shadow !), however variance is
minimal. In very bright objects, the sensor saturates, therefore the variance turns to zero as well (Manakos, oral
remarks). This curve is therefore a better indication for sensor sensitivity. The terrain knowledge allows a proper
selection of typical representative objects and analysis of their critical features. After this phase, an overall evaluation
can take place. For more automatic procedures, the difference between first and second membership function is an
interesting one (Baatz, 1999). In the case of the object in figure 4 and 5 this would mean 92,28-38.6=53.68. If this
object would be 99% member of coniferous--middle and 9796 member of coniferous- open the difference is only 2, Still
the class membership is very high. This illustrates the difficulties of object evaluation and the deviation from class
evaluation Until communis opinio on automatic procedures are not fully developed, visual check remains the basic
evaluation procedure, but automatic alternatives are available and waiting for acceptance among the user community.
6 CONCLUDING REMARKS
The analysis on a per-pixel basis is not very useful in image data with high internal variance. To overcome this
problem, an object based analysis allows a good solution. Object oriented analysis alone is not enough. In software like
eCognition, a few bottlenecks are solved at the same time. First, a single database allows a full integration of GIS and
remote sensing data. Using a hierarchical semantic network and a fuzzy logic ‘query’ facility on the database, the expert
knowledge can be incorporated and the implicit information richness in the database can be fully exploited. Image and
data fusion are a by-product from the visualization of the central database. The landscape model as defined in object
layer construction and flexible database query through adjustable decision curves, requires much more terrain
knowledge. This offers a extended role for the field expert. The software tools are very flexible and strong but the
landscape model is very much depending upon user specifications. The landscape object model therefore is the crucial
part of the analysis. For a specific application, exchange of modeling and object-sensor relationships becomes necessary
between users. Spatial relationships in a hierarchical network is a rather new way of defining semantics between image
and 'geo'- objects. Still this field is in full development and the results are promising enough to keep a keen eye on the
developments. As literature showed, over the past decades there is a development in favor of object oriented analysis of
environmental imagery and GIS data. Therefore these strategies are not a fashion flaw that comes and goes. This type of
analysis is very well rooted in long term development and will dominate remote sensing research in the near future.
Automatic accuracy assessment remains a very important topic. Confusion matrices are a too simple tool to achieve
proper inside in the robustness of the classification. Only expert consensus can achieve the formulation of procedures
that allow an acceptable level of automatic accuracy analysis. Object oriented analysis is here to stay; Educational
remote sensing literature should add a chapter accordingly.
REFERENCES
Baatz, M.,Schüpe,A. 1999,Delpi2 creative technologies GmbH, Software tutorial for eCognition ‚München
Cross, A. M., Mason, D.C., 1988. Segmentation of remote-sensed images by split-and-merge process.
Int. J. Remote Sensing, 9(8):1329-1345.
Flack, J., 1995. Interpretation of remotely sensed data using guided techniques, Ph.D. Thesis, School of Computer
Science, Curtin University of Technology, Western Australia.
Gorte, B., 1996. Multi-spectral quadtree based image segmentation. Int'l Archives of Photogrammetry and
Remote Sensing, Vol. 31, Part B3, pp. 251-256.
Gorte, B. 1998A. Probabilistic segmentation of remotely sensed images. ITC, publication 63, PhD
Thesis, ITC, Enschede.
Gorte,B.,1998B. Segmentation pyramid classification. Int. Arc of Photogr.and RS, Vol32 Part B3/1,pp 225-232.
Haralick, R. M., Shanmugan, K., and Dinstein, I., 1973. Textural features for image classification.
IEEE Transactions on Systems, Man, and Cybernetics, SMC- 3(6):610-621.
Haberäcker,P.,1995,Praxis der Digitalen Bildverarbeitung und Mustererkennung,Hanser, München.
Hinton,J.,C.,1999,Image classification and analysis using integrated GIS, .In: Advances in remote sensing and GIS
analysis, Atkinson,P.M., TateN.J. Ed., Wiley & Sons, Chichester.
Janssen,L.L.F, 1994, Methodology for updating terrain object data from remote sensing data.
228 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.