International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999

• if the surrounding heights are also error affected (e.g. due to

interpolation from mismatched points), the spike will be

replaced by an erroneous value - a small ‘peak' or ‘hole'

remains.

• as the filter is not terrain selective, natural terrain variations

are misinterpreted and removed.

• also valid heights are randomly replaced by the median

value of the filter window, a problem, which increases with

the size of the window.

Honikel (1998) has introduced a different approach for terrain

filtering, where the most error-affected bands of stereo and

InSAR DEMs are filtered. It has been shown, that error

behaviour of both DEMs in the spatial frequency domain is

complementary, which means that the filtered erroneous parts

can be replaced by those of the counterpart. In addition, the low

systematic error of the optical DEM is preserved in the fused

DEM. This terrain selective method reduces drastically the

amount of gross errors and gives a smooth approximation of the

terrain. The performance in terms of root mean square error

(RMS) is comparable to that of the median filtered stereo DEM.

But, as it reduces the amount of large errors more effective than

the median filter, it is the preferred method for the proposed

fusion method.

It has been mentioned that statistic errors may also occur in

highly correlated areas. It is necessary to localise them, as far as

possible, before the fusion and give them zero weight, as they

would otherwise unrestrictedly enter the fusion process because

of their high correlation values. In the stereo case, the

presentation of errors as spikes, and in consequence the rough

terrain appearance, is utilised for error detection. Assuming a

smooth terrain, the local variance indicates variations and is

used to detect spikes by their locally high variance values,

which distinguishes them from valid terrain slopes, which

appear as linear features. Such identified high variations and

their adjacent neighbours are rejected by giving them zero

weight. Beside the spikes, single high correlation values in areas

of low correlation are rejected from data fusion. In the InSAR

case, also isolated points of high correlation have been

removed. In addition, values neighbouring areas of low

correlation are weighted with zero, as errors may leak out of

those areas, depending on the unwrapping algorithm (see

above).

The squared correlation coefficient puts stronger emphasis on

high correlation values. Heights with correlation coefficient

below 0.5 will be rejected from the fusion process. The filtered

DEM is introduced into the data fusion process with a general

weight of 0.5.

The proposed method will especially reduce statistic errors

occurring in both DEMs, due to its averaging properties.

However, it performs more selective than the simple averaging

of both DEMs would do, as it uses weights to distinguish valid

heights from less reliable for the computations. Regions, where

measurements performed well, are emphasised, while suspicious

areas are less emphasised and will be adapted to the data from

the other sources.

5. TEST DATA

5.1. Dataset

The test dataset consists of a pair of ERS-1 SLC quarter scenes

(Frame 819, Quarter 1, Orbit 829 and 872), taken in a three day

interval (9/12/ and 9/15/91), and a SPOT stereo pair, taken in a

four day interval (9/3/ and 9/7/86). Both datasets are part of a

remote sensing dataset of Catalonia used for our work within

the EU concerted action ORFEAS (Optical Radar sensor Fusion

for Environmental Applications). Two test DEMs, each

covering a region of approximately 50 km“ (grid spacing: 30m,

each test site 65536 points), have been generated from both

datasets in an area south of the city of Lleida, close to the

villages Llardecans (test site 1) and La Granadella (test site 2).

Further details are given in Table 1. The cartographic projection

is UTM zone 31 and the ellipsoid used is the Hayford ellipsoid

with datum in Potsdam (ED50).

Xmin,

Xmax

(m)

Ymin,

Ymax

(m)

Zmin,

Zmax

(m)

Test site 1

291030

4582680

194

298710

4590360

445

Test site 2

302430

4580880

289

310110

4588560

599

The filtered DEM is refined with the weighted average of the

InSAR and stereo DEM. The squared correlation coefficient of

each point is used as weight. Thus

h(X, Y) =

h f(X,Y)p f 2 +h opt {X,Y)p opt 2 +h sar (X,Y)p sar 2

2 2 2

P f + Popí + P sar

(4)

with,

h(X,Y): resulting local height estimate

h(X,Y): local height of the filtered (f), optical (opt) and SAR

DEM

p: correlation coefficient

Table 1. Test site data.

The terrain in both sites is undulating with height differences of

251m and 310m, respectively. In test site 1, the terrain changes

smoothly, leading only to sporadic layover, while in test site 2

various slopes steeper than 21°, the incidence angle of ERS,

occur. A DEM, derived from the contour lines of a 1:5,000

topographic map, served as reference for our computations. The

RMS of the reference DEM was approximately lm. Permanent

crops is the dominant landuse in both sites, while site 2 shows

additional coniferous forests, which bias the measurements in

the optical case.