Urban Indicator for Database Updating

A Decision Tool to Help Stakeholders and Map Producers

Bénédicte Navaro, Zakaria Sadeq and Nicolas Saporiti

Geo212, 25 bis rue Jean Dolent, Paris, France

Keywords: Urban Areas, Object Detection, Spatial Databases, Minimum Noise Fraction, Supervised Learning,

Geodesic Dilation.

Abstract: The issue of regular spatial databases updating is partly solved by the abundance of satellite images. It is,

though, time consuming, requires qualified human resources, high financial costs and requests efficiency

(Bernard, 2007). This article presents a semi-automatic tool for urban detection, to guide the stakeholders

and the producers throughout the updating process. The industrial context of the study implies a fast,

instantaneous applicative workflow, operational on various landscapes with different sensors; it is thus

based on existing algorithms and software resources. The process is generic and adaptable, with a phase of

uncorrelation, chaining a Minimum Noise Fraction transformation with a textural analysis, a learning phase,

processed from an existing database, and an automatic modelling of the detected objects. The quantification

of the results shows the successful recreation of the existing database (90% of its surface) with a 7% rate of

potential big omissions. A specific highlight is made on the detection of disappeared buildings,

corresponding to 17.5% of the potential important omissions. This process has run in “real” updating

operations, on 1.5 and 6 meters resolution Spot6 images, a 15 meters Landsat-8 image and a 1.5 meters

resolution Pleiades image. A quantification of its results is also proposed in this study.

1 INTRODUCTION

A tremendous amount of techniques in the change

detection and database updating fields have already

been explored. Lu et al., 2003 “Change Detection

Techniques”, give an overview of the most common

techniques and qualify them in terms of

characteristics, advantages, disadvantages and key

factors. The heterogeneity of urban environments

and the large number of mixed pixels inherent

images often induced difficulty in urban land

use/cover classification based on spectral signature

(Lu and Weng, 2004). Recently, in the image change

detection field, much attention shifted to advanced

classification algorithms like neuronal network,

object-oriented and knowledge-based classification

approaches (Zhang and Wang, 2003). This study

aims at updating an existing database of urban GIS

objects with recent satellite imagery. The concept is

that, having assumed that the number of wrongly

detected GIS objects and the number of changes in

the real world are substantially less than the number

of all GIS objects of the data set, training areas can

be derived from existing GIS data (V. Walter, 2000).

As this study is realised in an industrial context and

aims at a generic and adaptable process, it is focused

on a simple chain processing, based on existing

algorithms. The easiest way to configure these

algorithms is to use the ones implemented in

software like ENVI and ArcGIS, available in the

company. These tools are not mandatory as the

image processing algorithms and statistical measures

they use, are well-known by the community and can

easily be reproduced in any computing languages.

The choice of a radiometric analysis is made due to

the simplicity of its implementation. The first phase

chains a Minimum Noise Fraction transformation

and a textural analysis resulting in two images: a

relevant component from the MNF transformation

and a grey scale image corresponding to the

previous component’s variance. The pixels’ values

in the two resulting images are combined with a

selection of relevant objects in the database to

establish a threshold. This learning phase finishes

with the morphological reconstruction of the

detected objects using a geodesic dilation with the

existing database as a mask.

Navaro, B., Sadeq, Z. and Saporiti, N.

Urban Indicator for Database Updating - A Decision Tool to Help Stakeholders and Map Producers.

DOI: 10.5220/0006327600810089

In Proceedings of the 3rd Inter national Conference on Geographical Information Systems Theory, Applications and Management (GISTAM 2017), pages 81-89

ISBN: 978-989-758-252-3

2 IMAGE PROCESSING

A multispectral Pleiades image (4 bands: blue,

green, red and PIR) from 2015/09/02, with a

2 meters resolution, is processed. The method aims

at extracting a specific thematic (urban objects) from

spectral information within a complex landscape

mixing vegetation, water, anthropogenic features,

soil ... To reduce the dimensionality of the image

and obtain a signal uncorrelated from the noise, a

Minimum Noise Fraction forward transform is

processed. The MNF transformation as modified

from Green et al. (1988) and implemented in ENVI,

is a linear transformation that consists of the

following separate principal components analysis

rotations (ENVI 4.2, 2005, Vermillion and Sader,

1999):

- The first rotation uses the principal

components of the noise covariance matrix to un-

correlate and rescale the noise in the data (a process

known as noise whitening), resulting in transformed

data in which the noise has unit variance and no

band-to-band correlation.

- The second rotation uses the principal

components derived from the original image data

after they have been noise-whitened by the first

rotation and rescaled by the noise standard

deviation.

The result of the MNF first rotation is a two

part data set, one part associated with large

eigenvalues and coherent eigenimages, and a

complementary part with near unity eigenvalues and

noise dominated images. The information is

compressed in different bands in which the

redundancy of the information is eliminated. In our

case, the majority of the information is contained in

the first 3 bands.

Then, the analysis of the spatial variation of a

component’s grey scale levels is processed: the

variance measures the dispersion of the values

around the mean. The solution implemented in

ENVI uses a co-occurrence matrix to calculate

texture values. This matrix is a function of both the

angular relationship and distance between two

neighbouring pixels. It shows the number of

occurrences of the relationship between a pixel and

its specified neighbour. Haralick et al. 1973 refer to

this as a “gray-tone spatial-dependence matrix”. The

texture analysis is done, in our case, on the second

band of MNF components (Figure 1).

(a)

(b)

(c)

Figure 1: Image processing results: (a) Pleiades 2m image

of Dire Dawa (Ethiopia), (b) MNF band 2, (c) Variance of

MNF band 2.

The chosen MNF component (band 2) and the

variance image are considered to be the resulting

images of the image processing phase. The learning

phase is applied on these two images.

3 LEARNING PROCESS

In order to calculate the threshold that will

discriminate urban objects in the resulting images

(Figure 1), the mean of the pixels’ values that are

located in areas labelled as “urban” amongst the

available database, is calculated.

GISTAM 2017 - 3rd International Conference on Geographical Information Systems Theory, Applications and Management

3.1 Labelling Objects as “Urban

Areas”

The selection of anthropogenic objects in the

database can be different regarding the type of the

available database. The positive learning in an

outdated database is justified by the rare

disappearance of urban objects. In the case of this

study, the buildings are represented individually so

the selection of a sample can easily be done in the

whole existing “building” layer. In other cases a

simple SQL request can sort the data by attributes

and randomly create fragments within the selection.

The choice of urban areas results from a human

consideration as it depends on the attributes the user

considers as “urban”. For example, urban places

with a lot of vegetation should not be inventoried

(like cemetery), nor punctual objects that are too thin

to be representative of the local area (like pylon or

water tanks). The choice of representative objects in

the database is, then, flexible, and can be updated

when changing area, if a new type of relevant

anthropogenic object appears. Nevertheless, the list

of these attributes has to be made just once for a data

model.

3.2 Thresholding

In order to establish a threshold, the mean of the

pixels’ values located in fragments of buildings, is

calculated on both images resulting from the image

processing (Figure 2).

Figure 2: Thresholding results based on the fragments (red

areas) in Dire Dawa, Ethiopia.

The final result consists in the intersection of

the two thresholded images and is presented in

Figure 3 over the original footprint of the buildings

(grey polygons).

Figure 3: Urban detections in a Pleiades image (2m) over

Dire Dawa (Ethiopia).

Figure 3 shows the multiple detections of

anthropogenic objects inside the city and the

differences with the database. These detections are

still blobs with undetermined shapes, so for them to

get the shape of the buildings, a step of

reconstruction is necessary.

4 MORPHOLOGICAL

RECONSTRUCTION

The geodesic dilation enables the reconstruction of

the buildings’ shapes. This is the dilation of an

image constrained by another image (Figure 4). The

first image is the assembly to dilate (marker) by a

structuring element, which is a 3 pixels square (this

is a size one dilation). The second image limits the

expansion of the dilation: it is the mask.

Figure 4: Blob extraction by marking and reconstruction,

Computer Vision (2015-2016), Marc Van Droogenbroeck.

In the example of Figure 4, even if the

marking blob only represents a small part of the

mask it is, thus, reconstructed.

In this study, the anthropogenic detections are

the blobs to reshape (red elements in Figure 5).

Blobs intersecting the existing database are the

markers to be dilated (blue elements in Figure 5),

Urban Indicator for Database Updating - A Decision Tool to Help Stakeholders and Map Producers

and the database itself is the mask (yellow polygons

in Figure 5).

Figure 5: Geodesic dilation of the MNF detections.

This reconstruction allows obtaining

maximum benefit from the MNF method, able to

detect small and isolated buildings. A certain

number of iterations lead to the reconstruction of the

buildings existing in the database and detected by

the MNF method (Figure 6).

Figure 6: Evolution of the geodesic reconstruction

(iteration from 1 to 185).

As the database used as a mask is outdated,

detections corresponding to new buildings are out of

the database: not all of the detections become

markers. So, this morphological reconstruction must

be completed by another step of reshaping. The

anthropogenic detections that have not been

reconstructed are reshaped.

5 AUTOMATIC RESHAPING

The automatic reshaping method is the one

imbedded in the ArcGIS solution: Minimum

Bounding Geometry.

The first step of the reshaping is the selection

of the MNF detections that have not been

reconstructed. This step is a simple GIS combination

of merging and intersection that enables to analyse,

for each MNF detection, if it is “well represented”

by the reconstruction. If the detection is covered by

the reconstruction polygon at a minimum of a 60

percent rate, it is, thus, considered as well

represented in the reconstruction. If not, the part of

the detection that is not covered at all by the

reconstruction is isolated and reshaped, as shown in

Figure 7.

(a)

(b)

Figure 7: Result of the reshaping (green polygons) from

the MNF detections (red blobs) after the reconstruction

step (yellow polygons).

Figure 7 illustrates the reshaping of a

detection that is not correctly represented in the

reconstruction step. Indeed, the blue circle shows an

extended building visible on the image that is not in

the existing database. The detection of this building

is reshaped, as are the isolated surrounding ones.

Notice that detections on the left side of the image

are not reshaped because they are considered as

“well represented” by the step of reconstruction

(yellow polygons).

Polygons smaller than 5 m² are cleared from

the results, considered as non significant.

6 RESULTS AND

QUALIFICATION

The compilation of the 3 steps (MNF detection,

morphological reconstruction and automatic

reshaping) leads to a global detection of buildings

whether they already exist in the database, or are

GISTAM 2017 - 3rd International Conference on Geographical Information Systems Theory, Applications and Management

new buildings (respectively yellow and green

polygons in Figure 8).

Figure 8: Building detections in Dire Dawa

The process enables to delete buildings

present in the database that do not exist anymore and

to complete the database with new buildings whether

they are simple extensions of buildings or obvious

urban expansion (green polygons in Figure 8).

6.1 Quantification with Individual

Buildings Composed Database

Out of the 58,473 buildings present in the database,

52,528 have been reconstructed which represent

89% of the original database (90% in surface).

Figure 9: Result of morphological reconstruction.

Out of the 5,945 non reconstructed objects,

93 % of them are small buildings with a surface

smaller than 100 m². In the 7 % of potential

omissions of big buildings, the analysis of every and

each building reveals (Figure 10) that nearly 75 % is

a real omission, but in 17.5 % of the cases, the

buildings that have not been reconstructed have

actually been destroyed.

Figure 10: Analysis of the non reconstructed buildings

which surface is over 100 m².

In terms of percentage, 11% of the initial database

has been omitted (Table 1), but significant omissions

only represent 0.23%.

Table 1: Error Type II considering the reconstructed

objects.

Figure 11 highlights buildings present in the

database (red polygons) that are not reconstructed

because they do not exist anymore in the recent

satellite image.

Figure 11: Non reconstruction of destroyed buildings (red

polygons).

Urban Indicator for Database Updating - A Decision Tool to Help Stakeholders and Map Producers

Detecting destroyed buildings allows to

rapidly point outdated objects to producers in the

context of rapid mapping.

One can notice that a large amount of

omissions is due to a tile coating of houses. Indeed,

this type of coating has a similar signal to vegetation

and is not representative of the majority of houses.

The learning process, as it uses a sample of

buildings, is based on the major type of coating

which is not tile. This problem is inherent to a

radiometry based analysis in a complex landscape.

An optional improvement step can be performed

using additional data (cf. 0 7 Improvement

step

An updated version of the database not being

available, it is complicated to estimate type I errors

concerning false positives. This would imply to

check each of the reconstructed and reshaped

polygons to validate their existence. This work has

been done on a 100 polygons sample to estimate the

quantity of false positives. Each polygon is checked

and qualified as :

- positive: the corresponding building is

clearly seen on the image. The shape of the polygon

matches the reality or makes a relevant envelope

around the buildings.

- false positive: no corresponding building on

the image.

- not identifiable: the polygon does not really

match the outline of the corresponding building. In

this case it is possible that the building is not the

detected element, but the detection is still relevant to

point out anthropogenic features (cars and elements

of the road are not comprised in these features).

The results are presented in Table 2.

Table 2: Error type I estimated on a sample polygons.

This analysis is made separately on the

polygons resulting from the phase of reconstruction

and the ones resulting from the phase of reshaping,

in order to estimate which one of the phases

produces more errors. Table 2 highlights that the

phase of reconstruction produces few errors of

commission (5%). Moreover, the mean size or false

positives in the reshaped buildings is around 30 m²

which is a small surface. These results may be

improved by changing the threshold of significant

surfaces applied at the end of the reshaping phase

(cf. 0).

6.2 Visual Qualification on Various

Areas

The robustness of the MNF method has been tested

with 3 different types of images (Spot6, Pleiades,

Landsat), with 3 different spatial resolutions (1.5 m,

6 m, 15 m), with different landscapes (desert, dense

urban area, mangrove, mine), on different swathes

and with different levels of the database’s

obsolescence. Each one of the 11 tests were realized

with the same settings: a unique list of representative

urban objects for the learning phase (no “building”

layer available, as in the Dire Dawa case), a variance

calculated from the second band of the MNF result,

and a size of fragment proportional to the pixel size.

And each one of them led to an enriched analysis of

the database’s obsolescence. Most of all, the method

has been tested in an operational context as an input

data for the map producers.

In Saint-Louis (Senegal) the method spotted

urban extensions, in Mali it detected a whole gold

mine (Loulo) that was absent from the database

(Figure 12). In the folowing figures the red marks

correspond to the indicator and the grey areas

correspond to the objects in the database.

(a)

(b)

Figure 12: Urban extensions in Saint-Louis, Senegal (a)

and detection of the Loulo gold mine in Mali (b).

GISTAM 2017 - 3rd International Conference on Geographical Information Systems Theory, Applications and Management

At a different scale, in Dubai and Bamako

new infrastructures and non-existing ones were

highlighted by superimposing the indicator on the

initial database (Figure 13).

(a)

(b)

Figure 13: New (circles) and obsolete (square)

infrastructures in Dubaï (a) and Bamako (b).

In addition of detecting the infrastructures,

the process limits the false detections which are, in

most of change detection methods, a limiting factor

of use. In Lu and Weng (2006) the impervious

surface was overestimated in the less-developed

areas but was overestimated in the well developed

areas. In our case, if the false detections still exist,

they are so few that they don’t perturb the visual

analysis of the indicator. Figure 14 illustrates the

ability of the process to concentrate the detections in

the urban area amongst a wide desertic study area.

(a)

(b)

Figure 14: Desertic areas with very few false detections in

Abeche (a) and Forchana (b), Chad.

Note that detections in (b) Figure 14 correspond to

small settlements.

7 IMPROVEMENT STEP

If the method gives good results, an improvement is

always possible. A choice of useful band is done

after the MNF transformation. It appears that the

urban information is not represented in just one

band. The detailed process can be performed on

another band or with additional data.

Surface elevation models provide a useful

additional information as it is independent of the

radiometric aspect of the data. With such a model,

the tile coating of the buildings, in the Dire Dawa

example, and the fact that its spectral signal is very

close to vegetation, is not a problem anymore. The

introduction of a geometric primitive (and not

radiometric) allows a more invariant detection of

buildings. Nevertheless, we chose not to work

initially with a surface elevation model as we

focused on the satellite image. Derivating such a

model from a stereo or tri-stereo satellite image is

thus possible, but is not the focusing point of our

work. Champion, 2011 proposes an innovative

database updating based on the improvement of a

Digital Terrain Model (DTM) derived from a Digital

Elevation Model (DEM).

In the case of Dire Dawa, DEM and DTM

data were available. The vegetation index (NDVI) is

calculated on the image to improve the distinction

between elevated objects corresponding to buildings

and other ones corresponding to trees. The elevation

data, masked from the vegetation, is then used in the

reconstruction step. In the first reconstruction step,

156 buildings with a surface superior to 200 m² were

missing. The improvement step allows to reconstruct

73 % of the omissions and 70 % of the tile coating

buildings (Figure 15).

Urban Indicator for Database Updating - A Decision Tool to Help Stakeholders and Map Producers

Figure 15: Improvement due to elevation and NDVI data.

The major inconvenient to this step is that it adds

commission errors to the results. Indeed, Figure 15

shows the reconstruction of 33 % of destroyed

buildings.

8 CONCLUSION

This semi-automatic tool provides detections of

rapidly evolving features that does not match any

evolution scheme: urban areas. It tends to answer the

problem of efficiency in the updating databases

process answering the questions “Where? How

much? When?”. The results can be presented in a

map, to spot the necessity of the database update, it

can be presented over the original image to help the

producers to focus on important updating area

(especially destroyed buildings), or it can be used at

the end of the production process as a quality

control. The method, using radiometric primitives

improvable with geometric primitives, is adaptable

to the type of landscapes, images, scales, and has

been recognized as useful by independent producers

during a real updating process test, not only as a

detection of anthropogenic areas process, but as a

real decision support instrument.

9 DISCUSSION

If this tool has proved its efficiency in a context of

rapid mapping, it cannot be trusted as it is for an

exhaustive automatic mapping. Indeed, a building

that has changed its shape between the date of the

database and the date of the image will probably be

reconstructed as it was in the past.

If we were able to observe very few errors of

commission (error type I) we didn’t establish a real

statistics on a representative amount of elements.

Moreover, in the statistics established on type II

errors, some missing polygons were considered as

omission but the field reality (the image) was

obviously different than the representing database:

in this case the qualification of omission is not really

correct.

In this study, two types of databases were

tested. The most advanced case (Dire Dawa) used a

database in which the buildings were individually

drawn. This allowed us doing the reconstruction and

the derived statistics and declaring the method

efficient. This is not possible with a database in

which the buildings are not individually drawn. This

is why a simple visual qualification is done on the

other tested areas (§ 6.2). When using this type of

database another type of statistics can be calculated

to estimate the changes and the amount of work in

the pre-production phase, but the quality assessment

of the method by calculating the reconstruction rate

is not possible. Another option would be to compare

our detection to other urban products (mixing

different types of data) like Global Urban Footprint

or Landscan.

REFERENCES

ArcGIS for desktop Help http://desktop.arcgis.com/

en/arcmap/latest/tools/data-management-toolbox/

minimum-bounding-geometry.htm

M. Bernard, T. Rousselin, N. Saporiti and M. Chikri, 2007

“Data harmonisation and optimisation for

development of multi-scale vector databases”, ISPRS

Workshop on Updating Geo-Spatial Databases with

Imagery, Urumqi.

N. Champion, 2011 “Détection de changement 2D à

partir d’imagerie satellitaire. Application à la mise à

jour des bases de données géographiques ”, thesis

ENVI 4.2, 2005 User’s Guide Volume 2 Chapter 7.

A.A. Green, M. Berman, P. Switzer, and M.D. Craig, 1988

“A transformation for ordering multispectral data in

terms of image quality with implications for noise

removal”. IEEE Transactions on Geoscience and

Remote Sensing (GRS), Volume 26 No 1 p.65-74.

R. Haralick, K. Shanmugan and I. Dinstein, 1973

“Textural features for image classification”, IEEE

Transactions on systems, man and cybernetics Vol.

SMC-3, N°6, pp.610-621

D. Lu et al., 2003 “Change Detection Techniques”,

International Journal of Remote Sensing vol.25, no 12,

2365-2407.

D. Lu and Q. Weng, 2004 “Spectral mixture analysis of

the urban landscapes in Indianapolis with Landsat

ETM+ imagery”, Photogrammetric Engineering and

Remote Sensing, 69, 973-980.

D. Lu and Q. Weng, 2006 “Use of impervious surface in

urban land-use classification”, Remote Sensing of

Environment 102 (1), 146-160

M. Van Droogenbroeck, 2015-2016, “Computer vision”,

open Repository and Bibliography, Université de

Liège.

S. Vermillion and S. Sader, 1999, Use of the Minimum

Noise Fraction (MNF) Transform to Analyze Airborne

Visible/Infrared Imaging Spectrometer (AVIRIS) Data

GISTAM 2017 - 3rd International Conference on Geographical Information Systems Theory, Applications and Management

of Northern Forest Types, AVIRIS workshop

bibliography

V. Walter, 2000 “Automatic change detection in GIS

databases based on classification of multispectral-

data”, ISPRS Archive Vol. XXXIII Part B4,

Amsterdam.

Q. Zhang and J. Wang , 2003 “A rule-based urban land

use inferring method for fine-resolution multispectral

imagery”, Canadian Journal of Remote Sensing, 29, 1-

Urban Indicator for Database Updating - A Decision Tool to Help Stakeholders and Map Producers