Printer Friendly

Advanced land use/ land cover classification using high resolution IRS imagery for the analysis of urban structure.


Human land use decisions on the environment are influenced by socioeconomic factors which can be represented by spatially distributed data. Up-to-date and area-wide information management in highly dynamic urban settings is a critical endeavor for their future development. Especially in growing and altering cities lack of up-to-date data is apparent. The challenge of classifying urban land cover from high resolution remote sensing data arises from the spectral and spatial heterogeneity of such imagery. There to the high dissimilarity of functions like industrial or residential areas as well as parks or agricultural regions causes problems in terms of an indirect inferring of land use.

Study Area And Methodology

Study Area

The study site, Vijayawada city, known as the political capital of the State, located in the south-east of India is the third largest city of Andhra Pradesh state. Vijayawada is located on the banks of the sacred Krishna River and is bounded by the Indrakiladri Hills on the West and the Budemeru River on the North.

The other details of Vijayawada city are:

District: Krishna

Coordinates: 16.30[degrees]N 80.37[degrees]E

Area: 58 [km.sup.2]

Elevation: 125 m

Time zone: IST (UTC+5:30)

Population (2006): 1,025,436

Density: 17,679/[km.sup.2]


High resolution multispectral IRS P-6 LISS-3(Band 2,3,4 &5) images were taken. This satellite carries three sensors (LISS-III, AWiFS & LISS-IV) with 5.8m, 23.5m & 56m resolutions and fore-aft stereo capability. The payload is designed to cater to applications in cartography, terrain modeling, cadastral mapping etc., These images were supplied by NRSA, Hyderabad, India.(

Global Positioning System (GPS) receiver has been used for ground truth data that records the coordinates for the polygons of homogeneous areas, and also it records the coordinate that will be used for geometric correction. The GPS is in existence since the launch of the first satellite in the US Navigation System with Time and Ranging (NAVSTER) system on February 22, 1978, and the availability of a full constellation of satellites since 1994. The US NAVSTAR GPS consists of a constellation of 24 satellites orbiting the Earth, broadcasting data that allows a GPS receiver to calculate its spatial position (Erdas imagine, 2001).

Ground truth data is used for use in image classification and validation. The user in the field identifies a homogeneous area of identifiable land cover or use on the ground and records its location using the GPS receiver. These locations can then be plotted over an image to either train a supervised classifier or to test the validity of a classification.


Here, the description about the land cover types and their distributions of the study area is given. Except this, the remote sensing images, ground truth used in this study are described in detail and also the data preprocessing before conducting the classification is described. Methodology to perform this research is given in figure 1.



Unsupervised and Supervised classification

The basic premise for unsupervised classification is that spectral values within a given land cover type should be close together in the measurement space, whereas spectral data in different classes should be comparatively well separated (Lillesand, 2001).Unsupervised classification is fast and has the ability to analyze the image spectral statistics completely and systematically, thus unsupervised classification can give useful indication of detectable classes for supervised classification (Mather, 1987).

Supervised classification results of the study area (Vijayawada city) with different land cover types is presented in figure 2.


Accuracy assessment

Accuracy assessment for Unsupervised classification and Supervised classification of Study area using IRS P-6 LISS-3 (Band 2) image taken in the month of Jan 2007 have been evaluated from the error matrices that are generated using training set data, random sampling data and test area data.

Due to poor producer's and User's accuracy of Urban and Vegetation classes in the unsupervised classification result (Though over all accuracy is 87.67%) of IRS P6 LISS-3 image, it does not attain the aim of land cover mapping and also does not attain the purpose of mapping the locations of the Urban and vegetation properly.

Accuracy assessment for supervised classification from error matrices has been evaluated using Random sampling data and Test area data. From the error matrix by random sampling data the producer's and user's accuracy of urban class is not good. The error matrix by test areas is as given in table 1.

* U.A-Urban Area, W-Water, V-Vegetation, R-Rocky

Producer's accuracy can be calculated using the formula:

PA (class I) = [a.sub.ii] / [[summation].sup.n.sub.t=1] []

Producer's accuracy(%):

Urban area=840/844=99.5




User's accuracy can be calculated using the formula:

UA (class I) = [a.sub.ii] / [[summation].sup.n.sub.t=1] [a.sub.ik]

User's accuracy(%):

Urban area=840/871=96.4




Over all accuracy can be calculated using the formula:

OA = [[summation].sup.n.sub.k=1] [a.sub.kk] / [[summation].sup.n.sub.t,k=1] [a.sub.ik] = 1/n [[summation].sup.n.sub.k=1] [a.sub.kk]

Over all accuracy= (840+546+1313+214)/2985=97.5%

Kappa Statistics can be computed as:

K = N [[summation].sup.n.sub.t=1] [a.sub.ii-] [[summation].sup.n.sub.t=1] ([a.sub.i+] * [[??].sub.+i]) / [N.sup.2] - [[summation].sup.n.sub.t=1] ([a.sub.i+] * [[??].sub.+i])


n=no. of the rows in the matrix

[[??].sub.ii]=the no. of observations in row i and column i (on the major diagonal)

[[??].sub.i+]=total of observations in row i

[a.sub.i]= total of observations in column i

N=total no. of observations included in matrix


Kappa Statistics:


Table 1 is the error matrix by test areas. The overall accuracy is 97.5%. Information classes of "Urban", "Water" and "Vegetation" have both high or relatively high producer's accuracy and user's accuracy. It can be seen from the table 1 that the producer's accuracy and user's accuracy of "Rock" is 91.45% and 87.7% respectively. Because of the poor producer's and user's accuracy of "Urban" and "Water" by random sampling, it can be concluded that using the test areas as the reference data to generate the error matrix more likely give high accuracy value than using random sampling.

The reason mentioned earlier is that this is because test areas are areas that have typical and representative spectral characteristics of the classes thus easier to be classified correctly while random sampling is using the pixels which is randomly distributed by the software and those pixels may locate on the boundary between classes and are not easily classified correctly and thus the error matrix created by using test areas as the reference data has relatively high values of producer's and user's accuracy than those by random sampling.

Object Oriented Image Analysis

Using the object oriented image analysis approach to classify the image is performed in eCognition. Object oriented processing of image information is the main feature of eCognition. The first step in eCognition is always to extract image object primitives by grouping pixels. The image objects will become building blocks for subsequent classifications and each object will be treated as a whole in the classification. The segmentation rule is to create image objects as large as possible and at the same time as small as necessary. After segmentation, a great variety of information can be derived from each object for classifying the image. In comparison to a single pixel, an image object offers substantially more information.

Segmentation Procedure in Ecognition

Multi-resolution segmentation is a basic procedure in eCognition for object oriented image analysis. It is used here to produce image object primitives as a first step for a further classification and other processing procedures. Multi-resolution is a bottom up region-merging technique starting with one-pixel objects. In numerous subsequent steps, smaller image objects are merged into bigger ones. Throughout this pair-wise clustering process, the underlying optimization procedure minimizes the weighted heterogeneity of resulting image objects. In each step, that pair of adjacent image objects is merged which stands for the smallest growth of the defined heterogeneity. If the smallest growth exceeds the threshold defined by the scale parameter, the process stops. Throughout the segmentation procedure, the whole image is segmented and image objects are generated based upon several adjustable criteria of homogeneity in color and shape.

Comparison of segmentation results with different scale parameters in the study area

Figure 3 is the original image of the study area. Figures 4, 5 and 6 show the effect of segmentation results using different segmentation parameters. Except scale difference, the other parameters that influence the segmentation result are color, shape, smoothness and compactness but these are kept constant.





Figure 4 is the segmentation result with a scale parameter 5. Comparing this segmentation result with the original image, it is found that neighbor pixels are grouped into pixel clusters-objects, and because of the low value of scale parameter, there are too many small objects. Figure 5 is the segmentation result with scale parameter 10. It is found by comparing it with figure 4 that higher scale parameter value generates larger objects. Figure 6 is the segmentation result with scale parameter 20.

By visual comparison, a scale parameter of 10 is selected because the segmentation result fits the information class extraction best. Based on these parameters, segmentation process is performed.

Image classification

Classification is the process of connecting the land cover classes with the image objects. After the process of classification, each image object is assigned to a certain (or no) class. In eCognition, the classification process is an iterative process. The classification result can be improved by editing the result: defining unclassified objects with the correct classes, correcting wrongly classified objects with the correct classes, etc.

Accuracy assessment

Accuracy assessment values were generated in eCognition by creating a test area and training mask (TTA) as shown in table 2. The TTA mask contained 52 "Urban," "Vegetation," and "rocky" objects and 25 "water" objects. These objects, representing actual land cover were compared against the classified identity of these objects. The "water" class was very accurately classified, and was therefore limited to 25 testing objects in order to reduce its inflationary effect on the accuracy statistics.

Producer's accuracy can be calculated using the formula:

PA (class I) = [a.sub.ii] / [[summation].sup.n.sub.t=1] []

Producer's accuracy(%):

Urban area=12766/17934=71.18




User's accuracy can be calculated using the formula:

UA (class I) = [a.sub.ii] / [[summation].sup.n.sub.t=1] [a.sub.ik]

User's accuracy(%):

Urban area=12766/13717=93




Over all accuracy can be calculated using the formula:

OA = [[summation].sup.n.sub.k=1] [a.sub.kk] / [[summation].sup.n.sub.t,k=1] [a.sub.ik] = 1/n [[summation].sup.n.sub.k=1] [a.sub.kk]

Over all accuracy= (12766+59897+17600+30336)/127898=94.2%

Kappa Statistics can be computed as:

K = N [[summation].sup.n.sub.t=1] [a.sub.ii-] [[summation].sup.n.sub.t=1] (a.sub.i+] * [[??].sub.+i]) / [N.sup.2] - [[summation].sup.n.sub.t=1] ([a.sub.i+] * [[??].sub.+i])


n=no. of the rows in the matrix

[[??].sub.ii]=the no. of observations in row i and column i (on the major diagonal)

[[??].sub.i+]=total of observations in row i [a.sub.+i]= total of observations in column i N=total no. of observations included in matrix


Kappa Statistics:


An overall accuracy of 0.942 and a Kappa Index of Agreement (KIA) of 0.91 are fairly reasonable and good accuracy levels. However, it is felt that there is still much misclassification that can be improved upon. It is hoped that this can be improved by exploiting some class related features and topological relationships.

Histogram for this classification is given in figure 7.


Statistics of the classified image are given in table 5.4

From the Histogram of this classification it is clear that out of the 58 sq.kms of the study area the urban area covers 34.6 which includes residential, commercial, industrial, traffic and transportation, public utility etc., the vegetation (trees, plants, shrubs etc.,) covers 11.4 and water, rocky area etc., covers 10.6

Planning Efforts

The way with which the city is growing and developing due to the migration of population from rural areas for employment and other opportunities, it has been proposed that the ultimate land use structure of the Vijayawada urban area in the coming 20 years should be around 130 The residential area is proposed to cover about 48% followed by transport and recreation uses. The land use pattern for the coming 20 years should definitely be far more balanced compared to the prevailing situation if the authorities concerned look in to the following proposals.

Recommendations and Proposals

* The proposals aim at municipal performance improvement of environmental infrastructure and aims at socio-economic development.

* The proposals for municipal reforms are aimed at enhancing the efficiency, effectiveness and service delivery with accountability.

* The reform proposals should include privatization of advertisement tax collection, revenue improvement, town development, operation and maintenance of critical infrastructure investment.

* The environmental infrastructure proposals aim at improvement of infrastructure in the prioritized poor settlements as per poverty and infrastructure deficiency matrices and linked infrastructure for poor settlements.

* These include rehabilitation of existing infrastructure provision of water supply, roads, drains, sanitation and street lighting based on community prioritization and construction of drains to improve the living environment.

* The social development proposals aim at addressing the socio-economic needs identified and prioritized through participatory micro planning process.

* These proposals cover areas of health, education, livelihood, vulnerability and strengthening of SHGs (Self help groups), with focus on gender issues.

* This leads to the reduction of poverty and improvement in living conditions of the people in the poor settlements.


[1] Erdas Imagine, 8.5, 2001, User Guide.

[2] Lillesand, T.M., and Kiefer, R.W., 2001, "Remote Sensing and Image Interpretation," 4th ed, John Wiley and Sons, inc. USA, 2001, ISBN: 0471255157.

[3] Mather, Paul, M., 1987, "Computer Processing of Remotely-Sensed Images," St Edmundsbury Press Ltd., Bury St Edmunds, Suffolk, Wiley and Sons, ISBN: 0471-90648-4.

[4] ECognition user guide, 2001, Concept and Methods.

[5] Benz, U.C., Hofmann, P., Willhauck, G., Lingenfelder, I., Heynen, M., 2004, "Multi-resolution, object-oriented fuzzy analysis of remote sensing data for GIS --ready information," ISPRS Journal of Photogrammetry and Remote Sensing., 58, pp. 239-258.

[6] Goward, N., Huemmrich, F., and Waring, H., 1994, "Visible-near infrared Spectral reflectance of landscape components in western Oregon," Remote Sensing environment., 47, pp 190-203.

[7] Klimesova, D., Ocelikova, E., 2003, "Spatial data in land management and local government," Proceedings of the 4th Conference of the European Federation for Information Technology in Agriculture., Debreceen, 363-368.

[8] Kushwaha, S.P.S., Subramanian, S.K., Chennaiah, G.CH., Ramana Murthy, J., Kameswara Rao, S.V.C., Perumal, A., and behera, G., 1996, "International Remote Sensing and GIS methods for sustainable rural development," International Journal of Remote Sensing., 17, 3055-3069.

[9] Price, J.C., 1994, "How unique are spectral signatures? Remote Sensing of Environment," 49, 181-186.

Afroz Shaik Mohammed (1) * and Shaik Rusthum (2)

(1) Deccan college of Engg. and Technology, Dar us Salam, Near Nampally, Hyderabad-500 001,(A.P), India

(2) Professor & Principal, VIF College of Engg. & Technology, Gandipet, Hyderabad., (A.P), India
Table 1: Error Matrix By Test Areas For Supervised Irs P-6 Liss-3

                        Reference data *

 Classification       U.A        W      Total
      Data             V         R

      U.A.            840        13      871

       W               6         12      555

       V               0        546     1315

       R               3         6       244

                       0         0
                     1313        2

                       4         2

                      24        214

     Total            844       561     2985
                     1346       234

Table 2: Error matrix and accuracy statistics for this Classification
is given below

                               Reference data *

Classification     U.A         W         V        R      Total

      U.A         12766        0         0       951     13717

       W           5168      59897       0        0      65065

       V            0          0       17600      0      17600

       R            0          0        1180    30336    31516

     Total        17934      59897     18780    31287    127898

* U.A-Urban Area,  W-Water,  V-Vegetation,  R-Rocky

Table 3 Statistics of classification result

Land cover lasses           Pixel    Pixel no.     Area
                           number       P(%)      (Sq.Km)

1) Urban                   152746      10.54       34.6

2) Vegetation(Forestry)     50326       3.47       11.4

3) Others(Water, Rocky      46795       3.23       10.6
area etc..)
COPYRIGHT 2008 Research India Publications
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 2008 Gale, Cengage Learning. All rights reserved.

Article Details
Printer friendly Cite/link Email Feedback
Author:Mohammed, Afroz Shaik; Rusthum, Shaik
Publication:International Journal of Applied Engineering Research
Date:Nov 1, 2008
Previous Article:Studies on the chloride permeability of fly ash and silica fume based ternary blended concrete.
Next Article:Admissible set of robust PID controller values using Hurwitz Criterion for a Pure Integrating Process with dead time.

Related Articles
Removing redundant rules from a neuro-fuzzy classification system for remotely sensed imagery.
Modeling urban land cover from TM imagery.
Matrix to mosaic: habitat fragmentation from 1982-1999 in Sabah, Malaysian Borneo.
Mapping land-use/land-cover change in the Olomouc region, Czech Republic.

Terms of use | Privacy policy | Copyright © 2022 Farlex, Inc. | Feedback | For webmasters |