real estate monitoring system based on remote sensing and image recognition technologies

System's practical application is automatic building recognition from LiDAR data using saliency based method, vector map generation and change detection in actual cadastral maps.. Keywor

Procedia Computer Science 104 ( 2017 ) 460 – 467

1877-0509 © 2017 The Authors Published by Elsevier B.V This is an open access article under the CC BY-NC-ND license

( http://creativecommons.org/licenses/by-nc-nd/4.0/ ).

Peer-review under responsibility of organizing committee of the scientific committee of the international conference; ICTE 2016

doi: 10.1016/j.procs.2017.01.160

ScienceDirect

ICTE 2016, December 2016, Riga, Latvia Real Estate Monitoring System Based on Remote Sensing and

Image Recognition Technologies Sergejs Kodorsa,*, Aldis Rausisb, Aivars Ratkevicsc, Janis Zvirgzdsd, Artis Teilansa,

Ivonna Ansonee

a

Rezekne academy of Technologies, Atbrivoshanas str.115, Rezekne, LV-4601, Latvia

b

State Land Service of Latvia, 11.novembra krastmala Str 31, Riga, LV-1050, Latvia

c

Latvia University of Agriculture, Akademijas Str 19, Jelgava, LV-3001, Latvia

d

Riga Technical University, Kalku str.1, Riga, LV-1658, Latvia

e

Manchester Metropolitan University, Cavendish Str., Manchester, M16 GB6, United Kingdom

Abstract

Geoinformation are changing fast, therefore a change detection of real estate must be processed in short time The increasing resolution of sensed geospatial data creates critically important to develop high performance computing solutions to process geospatial information

The topic of scientific work is the real estate monitoring system based on image recognition and remote sensing technologies System's practical application is automatic building recognition from LiDAR data using saliency based method, vector map generation and change detection in actual cadastral maps The scientific work describes high performance computing solution and gives its performance comparison with traditional method

© 2016 The Authors Published by Elsevier B.V

Peer-review under responsibility of organizing committee of the scientific committee of the international conference; ICTE 2016.

Keywords: Building recognition; Geospatial; High performance computing; Land administration; Laser scanning; Remote sensing; Saliency

* Corresponding author Tel.:+371-28342737

E-mail address: sergejs.kodors@ru.lv

© 2017 The Authors Published by Elsevier B.V This is an open access article under the CC BY-NC-ND license

( http://creativecommons.org/licenses/by-nc-nd/4.0/ ).

Peer-review under responsibility of organizing committee of the scientific committee of the international conference; ICTE 2016

1 Introduction

Geographical information systems play essential roles in many fields (e.g ecology, meteorology, logistic, business and land administration) for solving scientific, business and administration problems improving decision making procedures related with spatial data

If geographical information systems are tools to visualize, to analyze, to process and to store spatial information, the full life cycle of intelligence is complex process, which can be broken into two stages: data collection and data analysis; where both processes are equivalently important to provide the precise and valuable intelligence results (e.g meteorological and hydrological forecasting models, strategic plans or analytical reports)

This scientific paper encloses the sphere of the geospatial data acquisition process and the automatic interpretation of collected geodata in land administration field, where the industrial research was aimed to evaluate, verify and validate automatic recognition system from two points of view: processing speed of geospatial data and its recognition accuracy; with a goal to determine the significant improvements for the existing system and applied methodology

2 Massive cadastral data actualization and its application in land administration domain

The proposed system was developed considering the whole life cycle of geospatial intelligence, which is required

to provide the qualitative decision-making process in land administration domain (see Fig 1) The system is constructed to prepare the automated building recognition process from remote sensing data, that is the most significant step to organize the massive and operative geospatial data actualization

The geospatial data actualization plays three important roles for land administration process:

x To provide actual and quality information for land administration decisions

x To provide information for new land administration tasks

x To evaluate the previous land administration activities

The system produces the vector layer of buildings, which is intersected with actual cadastral map using the vector overlay analysis with spatial relation “exclusive or” to detect the changes of real estate

Fig.1 Cadastral data actualization and its application

3 Requirements for information in land administration processes

Effective land administration should deal with various aspects of real property, including land and buildings and constructions, to cover different needs of land management It should embrace sectors of interest like land as resource, security of ownership and tenure, support for land and property taxation, security for credit; development

and monitoring of land markets, reduction of land disputes, improvement of urban planning and infrastructure development, support for environmental management and production of statistical data Information about land and land market processes that can be derived from effective land administration systems plays a critical role in all economies1

Evolving idea of sustainable development of the society entailed the present land management paradigm in which land tenure, value, use and development are considered holistically as essential and omnipresent functions performed by organized societies2

Fig 2 A global land management perspective 3

.

All activities related to land management includes stage of decision – to start the process, to make benchmarking,

to re-arrange the process and to finalize activity Keeping in mind the general trend of acceleration of economic and social processes, there are several requirements for the information needed to make particular decision

First of all, set of information should be tailored for specific type of the decision Consequently, any changes in decision making processes of land administration installs new requirements for necessary land information Secondly, updates of information should be arranged in a way to provide optimum actual data for any particular type

of land administration decision

The most important components of typical real property are buildings and constructions From the viewpoint of land tenure, land value, land use and land development buildings and constructions are the most valuable and changeable objects The variability of the buildings induces demand for frequently updated building data in land information systems For example, building taxation and construction processes needs yearly updated information High value of buildings as components of the real property demands high accuracy and completeness of the data Therefore, all the changes in the main physical features of all the buildings should be detected

Aforementioned requirements for the land information technically can be settled by existing methods of data acquisition, using field survey of the land parcels to detect changes However, costs of such a data flow will be disproportionate to the effectiveness of the land administration processes Therefore, new technologies should be used to solve current needs for actual and complete building information

4 Data collection using remote sensing technologies

The massive data acquisition is organized using an aircraft-mounted laser altimeter A laser altimetry or Light Detection and Ranging (LiDAR) is an active remote sensing technology, which captures 3D model of terrain recording it as a point cloud, where the location of each point is calculated using an information about scanning device location, a scan direction and measured distance to object The distance to object is calculated using the travel time of light (1):

݀ ൌௗ௧

where c - the speed of light and dt - the travel time of light

5 Geospatial data processing

The milestone of data collection is only a point cloud, where each point is independent one from another, although the geospatial objects are the set of points grouped into a robust cluster If it is ordinary task for human eye

to identify the object in a point cloud, it isn't so easy for a machine The authors, C.Hug and A.Wehr, provided the next features, which describe geospatial objects in LiDAR point cloud4:

Table 1 Descriptive features of geospatial objects

have regular and smooth shapes

have constant reflectance values

are bounded by distinct edges, both in geometry and reflectivity

most artificial materials have low reflectance values, between 10% and

40% in NIR band

have random shapes with many vertical discontinuities don't have distinct segments of constant surface normal, which could

be expected vegetation has high reflectance values, between 45% and 70% in

NIR band

But these descriptive features must be converted into the mathematical form to make them usable for machine logic Such comprehensive list of mathematical features is analyzed in the scientific work5, the authors have grouped seventeen features into five groups:

x Height features

x Echo features

x Eigenvalue features

x Local plane features

x Full-waveform LiDAR features

Where the height difference was mentioned as the most important feature for geospatial object recognition

6 Building recognition method and methodology

The proposed system applies the saliency method based on the energy minimization approach, which is described

in the scientific work6 The described method uses the height difference feature in the segmentation and recognition algorithm:

Algorithm “Building Recognition”

Start

Input: LiDAR point cloud [.las]

Preprocessing: point cloud projection by the maximal single or last point

Attention points: slope detection using the logical reasoning (2)

݌ሺ௫ାଵǡ௬ሻെ  ݌ሺ௫ǡ௬ሻ൐ ͳǤͺܱܴ݉݌ሺ௫ǡ௬ାଵሻെ  ݌ሺ௫ǡ௬ሻ൐ ͳǤͺ݉. (2)

Saliency detection: min-cut/max-flow segmentation using Dinic's algorithm, where seeds are low and high

points of the detected slopes

Salient additional recognition: filter by area

Postprocessing: vectorization using Theo Pavlidis' algorithm

Output: vector layer [.shp]

End

The approvement of min-cut/max-flow segmentation as fully fledged classification algorithm can be found in the source7 together with its more detailed metric definition and image transformation description for recognition needs

7 Computerized building recognition

The subject of this research is to evaluate two engineer technical solutions with a goal to to determine the optimal geospatial region for each solution and processing time and to validate and verification the developed recognition system The compared solutions are a single computer and high performance computing solution with a cluster computer 25 samples of LiDAR data with area 1km x 1km have been processed

The statistical data of the spend time to process each sample are depicted in Fig 3 Using the collected data, the total processing time has been calculated for different data amount (see Table 2) The total time for single computer

is calculated as the sum of sequent calculations, but the time for cluster computer is calculated by the equation (3)

ݐ ൌ ሾܽݎ݁ܽ ܿ݋ݎ݁ݏሿΤ כ ݀ݐ, (3)

where dt – time to process one sample,

[area/cores] – numbers are rounded up to whole number

Fig 3 The statistical data of the spend time to process each sample using single computer and cluster computer core

Table 2 Processing time for different data amount

Levels Area

Median Pessimistic Median Pessimistic

8 Validation and verification of recognition system

The most important and significant part of the research is the evaluation process of the collected data The goal of validation and verification process is to determine the quality of obtained data and to express this quality using quantitative and qualitative parameters The results of quality assessment are the basis to determine the further development of technology, the technology possibilities, its practical applications and efficiency potential There must be the clear understanding of the real numbers The clear understanding of the current stage of product and its effectiveness allows to take the correct decisions about the further system development and its implementation into production as the practically working model It looks very simple from the first review:

x The obtained result is depicted in Fig 4, which can be evaluated visually

x It is possible to see and identify the underlying problems - incorrectly recognized objects

x Formed idea about how to recognized the number of objects is generally to the dominant expressed in relation to incorrectly set of objects number

The obtained result can be considered good by the first visual review, but it is only subjective judgment

Fig.4 The main errors: bridge borders and robust vegetation

The strong scientific assessment is systematic and complex evaluation process with the uniform descriptive criteria, when the results are compiled, analyzed and compared using numerical and statistical data Taking into account that the result is associated with the possibility to verify the cadastral registration of buildings, the essential part of the evaluation process must use the available cadastral information about these buildings

In the first stage of system validation, the recognition results are compared with the available cadastral information (see Fig 5), when the comparison results are divided into three groups:

x Recognized buildings, which overlap with the similar cadastral data

x Recognized objects, which are not registered in cadastre

x Buildings, which are registered in the cadastre, but they are not detected by the automatic system

The more qualitative assessment of the technical solution requires additional information about the situation of the region in the period of time, when the input data were collected for the computerized recognition system The additional information can be taken from the same scanning results, which were used in the computerized

recognition system, as well as the similar time aerial photography (which were used in this study) or ortophoto from these aerial photography, or other specific information obtained by specialists in a manual or semi-instrumental assessment process

Fig 5 Comparison of building cadastral data with automatic recognition results

This part of the assessment process is completed in camera working conditions and it draws nearer the actual situation - the initial estimate of results From uncertain cases about 60-80% of information total amount are specified and verified in the evaluation process with a high confidence level In particular, it consists from unregistered cadastral objects as well as from objects, which are registered in cadastre and are not detected by the system, as well as from the area, where layers are with significant differences The evaluation phase concludes with the appropriate specified performance analysis and the preparation of references for field surveying

Qualitative evaluation of the final phase of information gathering is already packed in the field survey conditions

- the revision of specific object in the nature, according to the prepared information in the previous assessment phase All the objects are fully inspected, if the previous evaluation stages failed to obtain some conclusive information about these objects or the information is considered to be under developed Around 10% of clearly recognized objects are checked on a random basis to gain the confidence in the automated recognition efficiency and safety as foundation for undisputed cases At the end of the examination procedures, carried out all the information obtained testing, systematization, creating the administration tables and as the results of final analysis are included

in the evaluation report

Evaluation criteria positions and their number can be significantly expanded - in accordance with possible changes the direction and objectives of the evaluation determination or sometimes as of the results assessments influences founded changes In such cases, the entire evaluation process can be repeated and in addition for

fine-tuning in the interests of obtain some information In turn, to ensure that the obtained results in future for different places are similar, the evaluation process is repeated many times to covered different, but equivalent areas

The survey has showed the next results: 92% of detected and recognized objects and 8% are misclassified objects

9 Conclusion

If it is decided, that the service must be completed for one week, a single computer is the appropriate solution for the municipality level, but a cluster computer – for the region and country level

According to the scientific work6, the used method is working with the precision Cohen’s kappa 0.76 and total precision 98%, the survey has showed the result 92%, that is greater than Cohen’s kappa coefficient and smaller than the total precision So, the mean value can be used to set the more real precision of recognition using automatic statistic methods to measure the accuracy of recognition

The used recognition method is complex system, which can be improved in multiple directions It is possible to create the agent solution, which connects the group of computers in the network to create the parallel processing This solution can be intermediate solution between a single computer and a cluster computer The recognition method uses distance metric, which can be improved by combining different features The precision can be improved in the post processing level using the better filter

The sample processing time can be decreased, if a faster min-cut/max-flow algorithm is used The other possible case is to find the more appropriate size of sample, because the processing time of min-cut/max-flow algorithms depends on the number of vertices, but search objects (buildings) have compact geospatial shapes, which are significantly smaller than the processed region

Acknowledgements

The authors express their gratitude to the State Land Service of Latvia and to Latvian Geospatial Information Agency for providing samples for the research purposes The researchers thank Riga Technical University for providing the cluster computer for experiments

References

1 Land Administration Guidelines: With Special Reference to Countries in Transition, European Commission for Europe Geneva Available: http://www.unece.org/fileadmin/DAM/hlm/documents/Publications/land.administration.guidelines.e.pdf

2 Williamson I, Enemark S, Wallace J, Rajabifard A Land Administration for sustainable development Available: http://www.csdila.unimelb.edu.au/publication/conferences/Fig2010/LandAdministrationforSustainable_Development.pdf

3 Enemark S, Bell KC, Lemmen C, McLaren R Fit-For-Purpose Land Administration FIG Publication No 60, 2nd edition Available: http://www.fig.net/resources/publications/figpub/pub60/figpub60.asp

4. Hug C, Wehr A Detecting and identifying topographic objects in imaging laser altimeter data IAPRS 32; 1997 p 19-26

5. Chehata N, Guo L, Mallet C Airborne LiDAR Feature Selection for Urban Classification using Random Forests IAPRS 38; 2009 p

207-212

6. Kodors S, Ratkevics A, Rausis A, Buls J Building Recognition Using LiDAR and Energy Minimization Approach Procedia

Computer Science Vol 43; 2015 p 109-117

7. Kodors S Land Cover Recognition using Min-Cut/Max-Flow Segmentation and Orthoimages Environment Technology Resources.

3; 2015 p 127-133

Sergejs Kodors is the researcher, who is working in Rezekne Academy of Technologies His main research field is the intersection of artificial intelligence and image processing with geospatial information systems and remote sensing Contact him at sergejs.kodors@ru.lv