Full text: Fusion of sensor data, knowledge sources and algorithms for extraction and classification of topographic objects

International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
4.2. Identification of Urban Structures by Morphological 
Analysis of Panchromatic Data 
Preliminary talks with potential users made clear that especially 
regional planners are very interested in landcover maps. Up-to- 
date information of the respective planning areas in medium 
scales (1 : 100,000 - 1 : 25,000) is required, to be supplied by 
analysis of high resolution satellite imagery. So far, the 
landcover is determined on the basis of topographic maps in 
scale 1 : 100,000 as well as by ground surveys. This data 
acquisition method is very time-consuming and supplies 
insufficient results regarding geometrical accuracy and 
especially up-to-date status. 
Two needs become obvious: firstly, a high resolution merge 
product, in which the high geometrical resolution of the 
panchromatic data is combined with the spectral resolution of 
the multispectral data, for a good visual analysis, and secondly a 
(semi)automatic landcover classification. Since the class ‘built- 
up areas’ represent a substantial feature for settlement dynamics 
and expansion, especially in urban but also rural areas, the 
automatic identification of built-up areas, which are to be 
combined if necessary into residential areas, is particularly 
desirable. With these specifications, the goal of this project part 
is the creation of a "settlement mask" in medium scales 
especially for regional planning. As shown above, the purely 
multispectral classification without inclusion of 
spatial-structural image information provided insufficient 
classification accuracies, particularly for built-up areas. 
Since the software ERDAS Imagine, which was so far used for 
image processing, does not currently provide (or only to a 
limited extent) tools for spatial-structural image analysis, the 
software HALCON was procured. This software package was 
developed for various applications in machine vision and is 
used worldwide for development, research and education 
purposes. It offers various processing capabilities in different 
applications, like remote sensing and aerial photo 
interpretation, production automation, quality control, medical 
image processing or monitoring. 
After importing the panchromatic image scene, grey level 
morphological analysis was performed using an elliptic binary 
structural element. In this process, a ‘top-hat’ transformation 
and a ‘bottom-hat’ transformation were applied to the 
panchromatic data. The top-hat transformation is a "peak" 
detector, i.e. highlights bright spots in the original data, while 
the bottom-hat transformation (known also as well transform) is 
a "valley" detector, i.e. highlights dark spots. During the top-hat 
transformation, a morphological opening is performed, followed 
by a subtraction from the original image. The opening consists 
of an erosion followed by a ‘Minkowski addition’ (dilation). 
The effect of opening is that large structures remain 
predominantly intact, whereas small bright structures, e.g. lines 
and points are reduced or eliminated. The application of the 
bottom-hat transformation consists of a closing followed by 
subtraction from the original image. Closing is defined as a 
dilation and a subsequent ‘Minkowski subtraction’ (erosion). 
Closing achieves the opposite effect, i.e. small dark structures 
are reduced or eliminated. These two morphological filtering 
procedures provide complementary representations of the built- 
up areas in the panchromatic image, as verified by visual 
analysis of the results. Theoretically, since the buildings appear 
usually bright, a top-hat transformation would suffice. However, 
especially in areas shadowed by higher buildings, the bottom- 
hat filtering provides a good representation of built-up areas in 
addition to the top-hat results. 
Then, the filtering results are combined into regions. Various 
algorithms are available for region segmentation. After detailed 
investigations, the ‘hysteresis thresholding’ after Canny (1983) 
was selected to segment regions. In this method, two threshold 
values are used, a lower (‘low’) and an upper (‘high’) one, for 
the segmentation. All pixel values in the input image that are 
larger or equal to the upper threshold are transferred to the 
output image as ‘safe’ points. All pixels with grey value less 
than the lower threshold are rejected (get the value 0). ‘Potential 
points’ with grey values between the two thresholds are finally 
accepted, only if they are connected to a ‘safe point’ by 
‘potential points’, whereby the path length should be less than a 
threshold. The ‘safe points’ thus radiate in their environment, 
and "have a lasting effect" (hysteresis). The grey values of the 
input images remain unchanged, only some regions contain 
rejected pixels. 
The regions segmented with the above described procedure 
consist of more or less large single structures. For the creation 
of a settlement mask, it appeared useful to combine these 
structures to more compact spatial objects (larger building 
complexes, not an overall settlement mask). Therefore, a closing 
was applied to the data. Closing smooths edges of regions, 
merges objects separated by a thin line and closes holes smaller 
than the structural element, whereas individual regions, 
separated by a distance larger or equal to the structuring 
element, do not merge. 
4.3. Combination of Spectral and Morphologic Image 
After performing the two classifications based on the 
multispectral and morphologic image information, our 
investigations focussed on the unification of the partial results 
for the creation of an accurate settlement mask. 
The settlement areas extracted from the panchromatic data by 
morphological analysis were visually checked and showed a 
high agreement to the actual settlements. However, they are 
fragmented within "closed" settlement areas, i.e. consist of 
individual components and do not cover the whole settlement 
areas. Nevertheless, it should be considered that the term 
"closed" settlement area can be defined in different ways and 
very strongly depends on the investigation scale (scale of map 
to be produced, monitoring scale). Topics like the size of the 
non-built-up areas within settlement areas which should be 
detected, the number and type of landcovers to be classified, the 
minimum size of settlement areas to be determined, are decided 
by the concrete user and depend on the application. Regional

Note to user

Dear user,

In response to current developments in the web technology used by the Goobi viewer, the software no longer supports your browser.

Please use one of the following browsers to display this page correctly.

Thank you.