Urban thematic information extraction and dynamic extension detection

Urban thematic information extraction and dynamic extension detection

Tian Lianghhu, Zhao Yuanhong, Zhang Fuxing
Dept.of earth Science, Zhejiang University
Hangzhou, China

Abstract
Urban remote sensing is an important direction of remote sensing. How to improve its application effectiveness is a problem of scientists. Most of previous classification was under the hypothesis of homogeneity in one landuse, which is much suitable to agricultural application. This assumption is not correct for urban thematic information extraction. Some more approaches available to classification are needed to be developed in order to meet the needs of urban remote sensing. This paper developed a contextual method to improve the accuracy of urban thematic information extraction and determination of city boundaries. Result analysis shows it accuracy is over 91.7%. This is significantly important to highly developing urban area for the administration and decision-making.

Introduction
In the urban remote sensing the problem how is to improve the classification of urban land use and its change detecting. Most of Classification approaches are primarily on the assumption that it is homogeneous in the landuse. This possesses obvious limits in the urban thematic information extraction because of the heterogeneity of urban landuse.

With the further development of high-resolution sensors as SPOT and thematic mapper, the problem above affects greatly the accuracy of urban landuse classification. Karkham and Town-send (1981) and David (1984)discovered that high resolution data (such as TM data) have lower accuracy than low resolution data (such MSS data) in urban classification.

In this paper we describe a contextual method to exploit the spatial-spectral context of a pixel to achieve more accurate classification over a 15x15 square kilometer region of yueyang urban area (see fig.1) Hunan province.

This procedure provides an urban thematic map extracted accurately from 1987's TM data and two boundaries of urban area determined from 1987's TM data. 1978's MSS data, and photographs of 1984's map. The extended changes of Yueyang from 1978 to 1984 and from 1984 to 1987 are detected. It is very successful that accuracy is over 91.7%.

Urban landuse classification
This work is realized by the contextual method applied to landuse classification to improve its accuracy. Contextual information is said to be the relationship of a pixel to any other pixels in the picture. Certain classes of ground cover arelikely to occur in the context of others . One does not expect to find wheat growing in the midst of a housing subdivision for example.

Fig.1 The study area centered on the city of Yueyang, Hunan province

Fig.2 The procedures of urban thematic extraction

Fig.3 The spectral reflection line. 1. vegetation 2.mixtural class 3.urban

The contextual method just uses that context to improve accuracy in classifying remotely sensed data. It can be applied not only to original multispectral digital data but also to the classified data to reclassify its landuse. This paper uses reclassification considering large number of contextual pixels. Fig.2 demonstrates the reclassification procedures.

Dynamic Clustering

Dynamic clustered image can give various spectral clusters which reflect land covers and their environments. Becuse clustering has some correlation with the selection of orginal clustering centers, we train some typicalm samples to compute the original centers in order to reach the best clustered results. Some essential division and mergence between classes are also made in clustering (details please refer to {1}.

Considering the terrain a land cover of Yueyang urban area, wse select 13 classes as training samples. In terms of their spectral features, TM1,3,4,5 bands are used for clustering . Photo 1 is the clustered image of TM data after several split and mergence between classes.

In photo 1, blue part(code 1) is water-body, green part (code 2) is rural land, the grey (code 7) is the old urban, the yellow (code6) is the new urban and the bright yellow (code 14) corresponds to the new building storehouse and open land (BSO) . The distribution of the pink part and its spectral features (See fig.3) shows that most of the urban and non-urban land use (Such as G,C) except some noises (Such as A,B,E,F,etc.) are the mistral class.
Contextual information Extraction and urban Landuse Reclassification

Analysis indicates that there are four problems in the clustered image (1) There exists a mixtural class between the urban and nonurban landuse. (2) Separate water points is presented in the urban area. (3) there are some urban points in the non-urban area and some non-urban points in the urban area. (4) The targets in the urban area is complicated that some old urban poits exist in the new urban area while several new urban points are in the old urban area etc.

For the reason above, a new contextual method is designed to improve the accuracy of the classification in which multi-frequency vectors, one form of the contextual information (See Fig.4) are used.

Supposing P1,P2,P3 and P4 are the frequencies of old urban, new urban, BSO and non-urban land respectively, we give the following discrimination criterion to x,a point of the mixtural class. that is:

Fig.4 With four possible classes, the frequncies corresponds to the gibven window is(5 3 0 1)^T

P4-(P1+P2+P3) > B X e non-urban
P4-(P1+P2+P3) <-B X e i, for Pi = max (P1,P2,P3,P4)
(P4-(P1+P2+P3) < -B don't be discriminated

The threshold B is determined according to the histogram of the value P4-(P1+P2+P3) of the training samples.

It is known that contextual reclassification is affected by window size as well as component frequencies change with the window size.

Fig.5 Correlation between reclassification accuracy and window size

Wharton tested this problem seriously. He used two subsences, 100x100 and 500x500 pixels. The acquired correlation curve line between reclassification accuracy and window size are shown in Fig. 5 . When window size increases, the reclassification accuracy also increases, then reachs the top at one point. as the window size increases further more, the accuracy decreases inversely. This is because the large window introduces some non-contextual even wrong contextual pixels.

It is difficult to determine the best window size in the practical classification. This paper, therefore, presents a multi-frequency vector algorithm to solve the problem. several windows which can express context most approximately are selected. These windows are applied to the pixels of mixtural class, in which the pixels that can not be discriminated by small window size are reclassified by larger windows frequency vector. Three windows, 7x7, 11x11 and 15x15 as in Fig. 6 in which weight coefficients decrease while the distance increases, are selected according to Fig.5

Fig.6 Windows with changed coeffecients

Fig.7 Flowchart of reclassification using multi-frequency vectors

Fig. 7 is the flow chart of the reclassification of the mixtural class. The procedures are also used to other three problems (2), (30 and (4) as stated above Photo 2 is the improved urban the metic map.

Dynamic detection of extension of Yueyang urban area
The 1978's MSS and remotely sensed data were processed by similar procedure. It is overlaid by the water bodies from 1987's TM data to remove exposed beaches in the Lake Dongting which is misclassified as urban landuse class due to the low water lever. Another urban area map is laso obtained from 1984's aero photographs. Comparing these three boundaries, an extension or change map is made as photo 3 in which red part and yellow part demonstrate the expansion from 1978 to 1984 and from 1984 to 1987 respectively.

Between 1978 and 1984, 4.8 square kilometers were expanded. North, east and south are its. Main extension directions. One larger extended area was the textile factory built in 1983 to the south of Dongting nitrogenous fertilizer factory. Qijialing district (on right lower part) was developed in this period.

From 1984 to 1987 ,3.0 square kilometres was rapidly extended in the city, The main extension is on the east of Yueyang near the Nanhu road. We expect that Yueyang will develop more rapidly in the future and the main extending districts will be Qijialing,Chenlingji and eastern Yueyang near the Nanhu road.

Accuracy analysis and conclusion
In order to know whether the application of contextual method has led to a significant improvement in classification, it is essential to test the classification accuracy after the procedure. For the reason of lack of essential landuse map, we use aerophotographs in 1984 as basic map and select two sub-regions for testing where there are more mixtural class pixels and less extension between 1984 and 1987. Comparing two maps with one by one pixel in A,B sub-regions, we obtain the accuracy table asTable 1. Overall accuracy is 91.7% consi-

Table1 Accuracy of reclassification

Sub-region	urban pixels of areaphoto	Urban pixels of TM data	Accuracy
A	714	751	94.9%
B	1124	1240	90.0%
Over all	1838	1991	91.7%

it is considered that the determination of the boundary of urban area and the detection of urban extension are correct. Results of applying the contextual procedure developed here indicate that the use of frequency vector can lead to significant improvements in the accuracy of classification.

This work first realizes the extension detecting and classfying over a small urban area. The procedures have only been tested on one type of classification , but can readily be extended to other types.

References

Zhao Yuanhong and Chen Lan, A Collection of Statistical Analysis and machine Processing Techniques of Multi-Remotely Sensed Data (Internal Materials), 1993
Gupta, D.M. and M.K. Munshi, Urban Change Detection and Land-use Mapping of Delhi. Int. J. remote sensing. Vol.6 No.3 1985, pp.529-594.3. Philip, J.H. and B.Emil, Land sat Digital Enhancements for Change detection in Urban Environments. Remote sens. Environment, vol 13, 1983, pp. 149-160
Swain, Ph. S.B. Vardeman, and J.C. Tilton, Contextual Classification of Multispectral Image Data. Pattern recognition, Vol.13 1981, pp.429-441.
Wharton, S.W. A Context-Based Land -use classification Algorithm for High-Resolution remotely sensed Data. J. Appl. Photographic Engineering, Vol.8, 1982, pp. 46-50. Photo 1 clustered.