Automatic digital terrain modelling

Farhad Samadzadegan

A ROBUST AUTOMATIC DIGITAL TERRAIN MODELLING METHOD
BASED ON FUZZY LOGIC

Farhad SAMADZADEGAN*, Mehdi REZAEIAN*, Michael HAHN**
*
University of Tehran, IRAN
Faculty of Engineering, Department of Surveying Engineering
Paradeyes@kanoon.net
**
University of Applied Sciences, Stuttgart, GERMANY
m.hahn.fbv@fht-stuttgart.de

Working Group III/2

KEY WORDS: Fuzzy Logic, Fuzzy Reasoning, Fuzzy Inference System, DTM, Matching, Robust Estimation, Finite
Elements

ABSTRACT

Most of the proposed methods for automatic DTM reconstruction are based on parametric estimation processes with
very little capabilities for reasoning and decision making. Human operators which measure DTMs solve interpretation
related problems while they carry out the geometric measurements. When working with terrain often decisions have to
be made which are of imprecise nature. For those problems fuzzy logic is a perfect basis to design a fuzzy inference
process.

In this paper we start to develop a fuzzy system for DTM acquisition. Fuzzy logic can be blended with proven concepts
and modules of DTM generation and therefore may supplement rather than substitute classical procedures for DTM
reconstruction. In a first step we design and implement algorithms for fuzzy feature detection and fuzzy matching in
image space. The transfer from image space to object space is then carried out by classical spatial intersection combined
with robust finite element modelling for the reconstruction of the terrain surface. Experiments with a Iranian test site are
carried out and show the applicability of the procedure. In the next step of future work imprecise hypothetical
information extracted from image space will be combined in a fuzzy manner with surface modelling to complete the
fuzzy inference system.

1 INTRODUCTION

It is more than ten years ago that success in image matching has initiated the development of numerous concepts and
routines for automatic acquisition of Digital Terrain Models (DTMs). First experiments with those routines have
indicated that the automatic procedures promise high accuracy of the reconstructed surface. Today, DTM modules are
mostly included in the software packages of digital photogrammetric systems and are widely applied in production. By
listening to critical voices it has to be noticed that neither commercial nor academic DTM modules are doing the job in
such a way that production highly benefits from automatically acquired DTMs. Many reports state that there is a fairly
high demand for interactive DTM editing and post processing in standard applications.

A closer look to these methods shows that in general DTM collection procedures are based on parametric estimation
processes combined with some robustification to eliminate errors. Obviously this classical concepts can not solve DTM
data collection properly. There is, for example, no procedure in such processes which may help to find out whether
measured points are on the terrain or on top of some topographic or manmade objects. Thus a lack of reasoning about
possible causalities has to be noticed in these approaches.

To cope with unsatisfactory DTM quality other sensors like laser scanners and radar are taken into account. The
discrimination between first and last response of the reflected laser signal contributes to a simplification of the
interpretation related aspects of DTM acquisition. But this does not mean that the basic interpretation problem is solved
or bypassed with range measurement scanners.

Instead of working with complementary data sources like range data and images more sophisticated algorithms may be
developed which is basically the background of the presented work. In the following we propose a method which
employs established concepts of hierarchical feature based matching and robust finite element modelling and introduce

International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 799

Farhad Samadzadegan

new ideas based on fuzzy reasoning for feature detection and matching. To deal with imprecise knowledge in the
reasoning system the fuzzy logic concept is used. The advantage is that no yes/no decisions are necessary in an early
stage of reasoning. The long-term objective of our development is to improve the interpretation related capability of the
algorithms, for example, by combining given uncertain prior information about smoothness of a surface with hints
about discontinuities or other textural information extracted from the images.

2 OUTLINE OF THE MAJOR PROCESSING MODULES

Coarse-to-fine procession using image pyramids and multi-resolution modelling of DTMs are well-known conceptual
aspects of DTM generation procedures. By introducing a fuzzy inference system for DTM acquisition the established
modules are taken over and used as a basis of the system. As mentioned before fuzzy logic can be beneficially blended
with proven concepts and modules of DTM generation.

The major components of the proposed approach are the following. Detection and matching procedures are started on
the top level of the image pyramids. A corresponding rough approximation of the terrain surface might be given or
derived using existing ground control points. This surface defines the coarse layer of the DTM pyramid. Projection of
the DTM nodes into image space using the given collinearity equations is carried out for the left and the right feature
pyramid. This information is used to initiate the matching process and delimit the location and search space for
establishing conjugate features. The detection of features as well as feature correspondence are taken into account by
employing fuzzy rules. The aspects of fuzzy inference are discussed in detail in the next section.

Transfer of conjugate feature points into object space is carried out by spatial intersection. The calculated object co-
ordinates of feature points are used to interpolate a regular grid terrain model based on robust finite element (FE)
modelling. Matching and DTM interpolation are executed using coarse-to-fine processing. This results in a progressive
densification of the DTM which finally is generated with its highest grid point density.

A refinement of the DTM interpolation process can be developed by taking fuzzy knowledge about the relation of
image features and object space modelling into account. A fuzzy rule which relates to all edge points may state that
’’corresponding edges of the images might be spatial discontinuities in 3D’’. Obviously this is not true in all cases
because there might be more edges which originate from texture, land use, etc. than 3D discontinuity lines. Therefore a
further rule stating ’’if the parallax on one side of an edge is similar to parallax on the other side then this edge will not
correspond to a 3D surface discontinuity’’ may contribute for clarification. These rules which are obviously of fuzzy
nature have to be processed by applying fuzzy operators, proper implication methods and aggregation of the outputs.
The task of fuzzy reasoning is to work out the different possibilities in an integrative manner. Reasoning and matching
have to be re-iterated with 3D FE modelling of the surface. If fuzzy reasoning indicates 3D discontinuities this can
directly be taken into account within the FE model by relaxing the continuity constraints for the corresponding
locations. Because fuzzy reasoning between image space and object space as indicated above is not fully worked out so
far we focus in the following on image space processes based on fuzzy logic and only shortly outline the object space
processes based on robust FE modelling. This reflects the current development state of the implemented software.

3 IMAGE SPACE PROCESSES BASED ON FUZZY LOGIC

Image space processing can be assigned to two separate stages: feature extraction and feature matching. In the first
stage key points have to be detected and in the second stage key points are matched to obtain conjugate points in two or
more images. Because the focus is laid on fuzzy inference processes for extraction and matching some basic ideas on
fuzzy reasoning will be reviewed first. Examples of fuzzy inference relating to the implemented detection and matching
processes will be presented afterwards.

3.1 Fuzzy Reasoning

Fuzzy inference is the actual process of mapping from a given input to an output using fuzzy logic. The process
involves membership functions, fuzzy logic operators, and if-then rules (Zimmermann, 1993). Since the terms used to
describe the various parts of a fuzzy inference process are far from standard and the idea is not to explain them all in
detail here we try to be as clear as possible about the different terms. A fuzzy reasoning processes may be composed by
five parts:

1- Fuzzification of the input variables. For each input variable the degree is determined to which it belongs to
each of the appropriate fuzzy sets via membership functions. The fuzzy set is a ’container’ of elements

800 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.

Farhad Samadzadegan

without a clearly defined boundary. The membership function defines how each point in the input is
mapped to a membership value.

2- Applying fuzzy operators. If–then rules are those conditional statements which make fuzzy logic useful.
The if-part of the rule is called the antecedent (premise), while the then-part is called the consequent
(conclusion). If the antecedent of a given rule has more than one part, the fuzzy operator is applied to
obtain one number that represents the result of the antecedent for that rule. This number will then be
applied to the output function, The output is a single truth value.

3- Implication. Before applying the implication method, one must take care of the rule's weight, which is
applied to the number given by the antecedent. The implication method is defined as the shaping of the
consequent (a fuzzy set) based on the antecedent. It occurs for each rule.

4- Aggregation of outputs. Aggregation is a matter of taking all the fuzzy sets that represent the output of
each rule and combining them into a single fuzzy set. The unified output of the aggregation process is one
fuzzy set for each output variable.

5- Deffuzification. The final step is defuzzification. The input of the defuzzification process is the aggregate
output fuzzy set and the output is a single number called the crispness.

The illustration of this 5 steps is given in Figure 1. It shows fuzzy inference for the example of matching small
patches or points using fuzzy conditions for correlation based similarity and conditions for expected and
tolerated parallaxes in image space.

Fuzzy Input Fuzzy Operation Implication

IF correlation is good AND Y_parallax is excellent THEN certainly conjugate pt.

Aggregation

IF correlation is good OR X_parallax is good THEN probably conjugate pt

Crispness
Defuzzification

Figure 1: Fuzzy reasoning diagram with an example based on fuzzy knowledge about
image correlation and conditions for the X and Y parallaxes

3.2 Detection of Key Points

Key points are points in the images which may take a key role within the overall processing. In particular these are
prominent points, edge points or other points which may represent significant texture. In the experiments interest points
extracted by Förstner's interest operator (Förstner and Gülch, 1987) are used. Other points in particular edge points have
to be integrated into processing at a later stage.


Farhad Samadzadegan

This interest operator can be developed starting with elementary statistical texture parameters like local gradient
variances in x- and y-direction. From this parameters a roundness measure and a weight is derived which defines (non)
isotropy of texture in the window and the quality of point localisation, respectively.

The selection process involves thresholding of both parameters, the roundness measures and the weights. If, in addition,
classification between circular shaped features and junction points is taken into account a further sharp decision has to
be made by carrying out a statistical test. In consequence, the results of the detection and classification process of the
interest operator might be different sets defined by all isotropic textured windows, by all highly weighted windows, by
circular features, by junctions or by general isotropic textures, respectively. In fuzzy set theory these are considered
classical sets with a crisp (well-defined or sharp) boundary rather than fuzzy sets.

Theoretically the interest point detection process my be fuzzified by introducing fuzzy sets instead of classical sets and
fuzzy rules instead of rules with hard thresholds. But instead of working out exactly such an inference process we want
to use additional basic texture measures, in particular, the grey value mean and the grey value variance. Those measures
shall be merged with the Förstner operator in a fuzzy reasoning system. In the following some of these criteria and rules
are discussed as an example.

Fuzzy Rules in the Detection Process

The following fuzzy rules are applied to interest windows, i.e. only those windows which are already detected and
selected by the interest operator by using low thresholds only.

Rule_1:

To avoid interest point extraction in predominantly dark or bright 1
areas (which are also avoided if point selection is carried out by 0.5
human operators) the mean grey scale in an interest window 0
should be higher than 30 and lower than 230 (Figure 2). 0 25 50 75 100 125 150 175 200 225 250

Rule 1. IF Mean Is Central THEN Probably Key Point ELSE Figure 2: Membership function of mean
Probably No Key Point

Rule_2:
1.5
For interest windows the grey scale standard deviation must be 1
0.5
above a certain limit (Figure 3). The corresponding rule reads as 0
0 10 20 30 Higher
Rule 2. IF Variance Is High THEN Probably Key Point ELSE
Probably No Key Point Figure 3: Membership function of variance

Some further rules are formulated which relate to the roundness and the weight measures of the interest operator
basically stating that the higher this measure are the higher is the probability for being an key point.

For all these fuzzy if-then rules the linguistic output is ''key point''. Applying the implication and aggregation steps leads
to the following membership function of key points (Figure 4). After defuzzification the resulting key point
probabilities are determined. All key points with probabilities higher than 50% are considered as detected points and are
plotted in Figure 5. Displayed are the detected points in three different levels of the image pyramid.

1
0.8
0.6
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1

Figure 4: Membership function of key points Figure 5: Detected points based on implemented fuzzy
reasoning


Farhad Samadzadegan

3.3 Matching of Corresponding Key Points Based on Fuzzy Logic

Another important part of automatic feature based DTM generation is matching of key points. Fuzzy inference can be
related to geometrical constraints and to radiometric constraints. Fusion of the constraints of both types leads to a fuzzy
reasoning process for finding establishing correspondence.

3.3.1 Geometric Constraints

As outlined previously the matching process takes advantage of multilevel DTM generation and coarse-to-fine
processing. The result of this is that an approximate location is predictable for establishing corresponding points. The
fuzzy knowledge about the search space and the given orientation of the images can be use to formulate fuzzy matching
rules for constraining the search space geometry.

Rule_1:

Given the predicted location of a certain feature point in the
1
second image of an image pair the distance between the predicted
0.5
locations and the location of extracted point in the search window
0
in X-parallax direction should be small. For convenience this
0 2 4 6 8 10 Higher
distance is abbreviated as X-distance. A sample for a membership
function is plotted in Figure 6. Figure 6: Membership function of X_Distance
Rule 1. IF X_Distance Is Short THEN Match ELSE Not Match

Rule_2:

Similar to rule 1 a second rule may apply for the Y-parallax of 1.5
two points. But given the image orientation or epipolar line 1
0.5
geometry the deviation orthogonal to the epipolar line (Y- 0
distance) must be near to zero. 0 1 2 3 Higher

Rule 2. IF Y_Distance Is Small THEN Match ELSE Not Match Figure 7: Membership function of Y_Distance

3.3.2 Radiometric Constraints

These constraints are based on the similarity of the grey values in the neighbourhood of extracted points in different
images. Cross correlation values (CC) of the grey scale images and furthermore of local texture descriptions can be
taken into account. But other features like local estimates of noise of the SNR are also of interest in this context.

Rule_1:

By correlating grey values, corresponding points should show up 1

with high cross correlation values. A corresponding membership 0.5

function is plotted in Figure 8. 0
0 0.2 0.4 0.6 0.8 1 Higher

Rule 1. IF CC Is High THEN Match ELSE Probably Not Match
Figure 8: Membership function of CC

Rule_2:

Local estimates of the SNR should be similar to for corresponding 1
points. For convenience the differences are abbreviated by S_diff 0.5
(Figure 9).
0
0 2 4 6 8 10 Higher
Rule 2. IF S_Diff Is Small THEN Probably Match ELSE Probably
Not Match
Figure 9: Membership function of S_Diff


Farhad Samadzadegan

Altogether currently 10 fuzzy rules are formulated and used for constraining matching. For all these fuzzy if-then rules
of matching the linguistic output is ’’corresponding point’’. Taking all membership functions into account fuzzy
inference will lead to the membership function of corresponding points plotted in Figure 10.

1
0.8
0.6
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1

Figure 10: Membership function of corresponding points

After defuzzification the resulting matched point probabilities are determined. All corresponding points with
probabilities higher than 50% may be introduced into further processes.

4 OBJECT SPACE PROCESS

Processing in object space basically starts with the co-ordinates of 3D points calculated by spatial forward intersection.
From the generally huge number of discrete mass points a continuous surface is approximated by FE modelling.
Continuity and smooth transition between neighbouring finite elements is introduced by regularisation. In the
implemented procedure a spline based FE model is used.

To deal properly with outliers the least squares estimation process of estimating the terrain surface is robustified.
Robust estimation using weight functions and iterative solutions are powerful for big data sets with high redundancy.
Because the number of matched points for each DTM grid is generally high, robust estimation is quite successful in
carrying out DTM interpolation. Details on those concepts are eg. given in Hahn, 1989 and Ackermann and Krzystek,
1991.

To develop a refinement of this procedure of DTM interpolation with a higher degree of robustness the points should
possess a certain semantic meaning. For this purpose the key points should be assigned to fuzzy sets containing, for
example, general topographic points, edge points with 3D discontinuities of different order, points which are likely to
belong to surfaces of manmade objects or other objects like trees. Rules like those outlined in section 2 have to be
developed and integrated in fuzzy inference processes to make the information more explicit which is only incompletely
and implicitly available in the images. The assignment to fuzzy sets without the need to be semantically accurate is
certainly a benefit for a development based on fuzzy logic.

5 EXPERIMENTS

Some first experiments are carried out with the presented
fuzzy method for automatic DTM generation.

The data set used is a mountainous area in the North-
Western part of IRAN. The scale of the photographs is
1:10.000, the focal length 150 mm and the length overlap
of two stereo images is about 60%. Thus the terrain area
imaged in both images covers an area of 1.5km × 3km.
The images were scanned by 20µ pixel size by the high
precision photogrammetric scanner of Intergraph. Figure
11 shows an overview image of this area.

Obviously this is a very meagre, rocky area. But it is a
typical image for the mountainous regions in IRAN
showing the challenge for manual DTM data capture as
well as for automatic procedures.

Manual measurements are carried out by a very
experienced operator using the ParadEyes station, which is
a digital photogrammetric workstation. The contour lines
Figure 11: Image used in the experiments


Farhad Samadzadegan

measured manually are shown in Figure 12. This measurements have been carried out for other purposes but can be
used in our experiments for visual comparison with the results of the automatic DTM generation.

In addition, a very dense grid is measured manually with a
20m sampling interval. These measurements are taken as a

Figure 13: Visualization of the automatically
reconstructed DTM
reliable reference for the automatically generated DTM
based on the fuzzy reasoning concept presented in chapter 3.
A shaded visualisation of the reconstructed DTM is plotted
in Figure 13. Even though the scale of both figures does not Figure 12: Manually measured contour lines
allow for any detailed comparison the similarity of the
morphologic structure of the terrain surface can be well recognised by comparing Figures 12 and 13.

Some statistical data about point extraction, matching and DTM reconstruction are collected in Table 1. In addition to
the 20 µ pixel size (layer 0) five additional layers have been used within coarse-to-fine processing. The respective
numbers of points detected by the fuzzy point extraction procedure if given for each layer together with the number of
matched points. Typically about 50 to 70% of the points are eliminated by the fuzzy conditions applied in matching.

An indicator for the accuracy of the automatically generated DTM is the rmse calculated by taking the check points into
account. The rmse values of the differences of the height of the check points and the DTM heights are indicating an
accuracy of 1:2 000 of the flying height which is considered to be a good results facing the low texture available in the
images (Figure 11).

Image Space Object Space

Detected Detected Match Number Mean RMSE
Points Points Points of Check Deviation to Check
(left) (right) Points to Check Pts(m)
Points (m)
Layer 5 489 422 226 11543 -5.916 12.558
Layer 4 1927 1759 900 11543 -1.886 6.003
Layer 3 8178 7250 3486 11543 -0.708 3.495
Layer 2 33817 30025 13509 11543 -0.386 1.966
Layer 1 134925 122252 51212 11543 -0.208 1.381
Layer 0 533648 490245 170267 11543 -0.111 0.763

Table 1: Statistics on point extraction , matching and DTM reconstruction in the test area


Farhad Samadzadegan

Altogether we do not want to overrate the results of this first test with our new fuzzy reasoning approach for DTM
reconstruction. But with this experiments we could show that the developed method has the capability to acquire DTMs
even in difficult terrain.

6 CONCLUSIONS

The goal of the presented work was to develop a new approach for DTM reconstruction based on fuzzy reasoning.
Fuzzy logic offers the theoretical framework on which processes of fuzzy nature can be build. Fuzzy inference
processes allow to utilise a variety of constraints by formulating proper fuzzy rules. Corresponding algorithms for point
feature extraction and for matching have been developed and presented. These algorithms are embedded into an overall
process for DTM reconstruction

First experiments are carried out which show the applicability of the developed and implemented procedure. This first
success encourages to extend the fuzzy framework by including the object based reconstruction modules in fuzzy
inference processes. Some ideas about refining DTM generation by relating image feature extraction and surface
modelling are outlined in the paper. A detailed elaboration of the corresponding fuzzy rules are left for future work.

7 REFERENCES

Ackermann, F., Krzystek, P., 1991. MATCH-T: Automatic Mensuration of Digital Elevation Models, Proceedings of
Technical Seminar of the Sociedad Espanola de Catografia Fotogrametria y Teledetection. Barcelona, pp.67-73.

Ebner H., B. Hofmann-Wellenhof, P. ReiB, and F. Steidler, 1980. HIFI - A minicomputer program package for height
interpolation by finite elements, International Archive for Photogrammetry and Remote Sensing, Congress Hamburg,
Vol. 27(3), p. 205-215

Förstner, W., Gülch, E., 1987. A Fast Operator for Detection and Precise Location of Distinct Points, Corners and
center of Circular Features, Proceeding of ISPRS Intercommission Conference on Fast Processing of Photogrammetric
Data, Interlaken.

Hahn, M., 1989. Automatic Measurement of Digital Terrain Models by Means of Image Matching Techniques.
Proceedings of the 42nd Phoogrammetric Week at Stuttgart University, p.141 – 151.

Huber, J.P., 1981. Robust Statistics, John Wiley & Sons.

Terzopoulos, D., 1986. Regularization of Inverse Visual Problems Involving Discontinuities, IEEE, Transaction on
Pattern Analysis and Machine Inteligence, Vol 4, p. 413-424.

Zadeh, L. A., 1965. Fuzzy Sets, Information and Control, Vol. 8, p. 338-353.

Zimmermann, H. –J., 1993. Fuzzy Set Theory, and Its Applications., Kluwer Academic Publishers.


Automatic digital terrain modelling

More Related Content

What's hot (17)

Viewers also liked (9)

Similar to Automatic digital terrain modelling (20)

More from Sumant Diwakar (20)

Recently uploaded (20)

Automatic digital terrain modelling