---> Application ---> Environment

Change Detection Using Various Image Processing Algorithms on Dehradoon City

Mr.Anshu Gupta
Mtech Student
Indian Institute of Remote Sensing,

Frequent acquisition of remotely sensed data makes it possible to use the satellite images to determine type and extent of changes in the environment. Many digital change detection algorithms have been developed since the launch of ERTS-1 in 1972 to reveal changes. With the launch of new satellite generations with different sensor characteristics and advancement in mathematical data processing algorithms, the use of new techniques to compare multi-resolution/ multi-temporal image data in change detection procedure is required. The purpose of this study was to reveal urban/agriculture changes using multi-scale analysis in Dehradoon city in India. Many authors have experienced employment of satellite imageries for land use mapping as well as change detection studies. The authors of this paper have compared the results of five different techniques of band combination, subtraction, band division, principal component analysis and classification to find the change detection of Dehradoon city,India.

Environmental protection is faced a critical problems due to several factors as the increasing population, demolishing natural resources, environmental pollution, land use planning as well as others. Presently unplanned changes of land use have become a major problem. Most land use changes occur without a clear and logical planning with any attention to their environmental impacts. Major Flooding, air pollution in large cities as well as deforestation, urban growth, soil erosion, desertification, are all consequences of a mismanaged planning without considering environmental impacts of development planes. Desertification is a common consequence of improper land use change.

During recent years Dehradoon city,India has undergone the land-use changes, usually without any land evaluation for specified purposes. This leads to intricate, and serious problems such as: potential agricultural failures, soil erosion, deforestation, etc. Due to increasing changes of land-use, mainly by human activities, detection of such changes, assessment of their trends and environmental effects are necessary for future planning and resource management. Among these, change in agricultural area is one of the important cases that occur in land-use. This happens in account of various factors (e.g., urban expansion, climate changes, etc). In India, cities have been usually surrounded by agricultural area, and changes are in particular harmful when urban expansion occurs against surrounding agricultural areas. The launch of the first satellite was the major advancement in monitoring the environment from space. Since that time remotely sensed data have been broadly applied to assess land cover changes. With the launch of new satellite generations with different sensor characteristics, and advancement in mathematical data processing algorithms, there is a need to use new techniques to compare multi-resolution/ multi-temporal image data.

Hoffer (1978) defined temporal effects as Variation in spectral response involves situations where the spectral characteristics of the vegetation or other cover type in a given location change over time.. Singh (1989) described change detection as a process that observes the differences of an object or phenomenon at different times. Since the launch of -ERTS-1- in 1972, there has been a range of attempts to reveal changes using remotely sensed data, resulting in various methods. Lunetta (1999) categorized change detection analysis approaches into either post classification change methods or pre-classification spectral change detection.

1.Spectral Change Detection Technique
In spectral change detection, images of two dates are transformed into a new single-band or multi-band image, which contains the spectral changes. The resultant image must be further processed to assign the changes to specific land cover types (Yuan et al., 1999). Since these methods are based on pixel-wise or scene-wise operations, they are sensitive to image registration and co-registration accuracy. Discrimination of change and no-change pixels is of the greatest importance in successful performance of these methods. A common method for discrimination is use of statistical threshold (Yuan et al., 1999). In this method a careful decision is required to place threshold boundaries to separate the area of change from no-change (Singh, 1989).

spectral change detection methods are as follows:

1.1 Image Differencing: In this method, two co-registered image dates are subtracted pixel by pixel in each band to produce a new change image between two dates (Singh, 1989; Jensen, 1996; Yuan et al.,)

1.2 Image Ratioing: Like the previous method, two co-registered image dates are ratioed pixel by pixel in each band. The no-change area is characterized by ratio values close to 1. Depending on the nature of changes between two dates changed areas will have higher or lower values (Singh, 1989).

1.3Image Regression: This method assumes that pixels from time t1 to be a linear feature of time t2 pixels. It considers differences in mean and variance between pixel values from two dates (Singh, 1989).

1.4Change Vector Analysis .:Yuan et al. (1999) defined a change vector of a pixel as . the vector difference between the multi-band digital vector of the pixel on two different dates.. A spectral change vector describes direction and magnitude of change from date one to date two. The output encompasses two images, one containing the magnitude of the change vector, the other its direction. Comparison of magnitude of changes with a specified threshold determines the change (if exceeds) and direction of change vector represents the type of change (Singh 1989).

1.5 Vegetation Index Differencing :In vegetation studies the ratio (known as vegetation indices) is used to enhance the spectral differences between strong reflectance of vegetation in the near-infrared part of spectrum and chlorophyll-absorption band (red part) of the spectrum (Singh, 1989). Typical vegetation indices include: Ratio Vegetation Index, Normalized Vegetation Index, and Transformed Vegetation Index.

1.6Multi-date Principal Component Analysis :In multi-date principal component, two image dates of the same area are superimposed and analysed as a single image. Yuan et al. (1999) reported that the major component images show the albedo (reflectance) and radiometric differences, and that minor component images reveal the local changes (minor changes). Li and Yeh (1998) reported the usefulness of this method to monitor rapid land-use changes and urban expansion in comparison with post classification method. They carried out a principal component analysis on two images from different dates, and an interactive supervised classification of land-use change was done on the compressed PCA image. In order to compare the proposed method with the conventional post-classification approach, images from two

1.7 Post-classification Technique :In the post-classification approach, images belonging to different dates are classified and labeled individually. Later, the classification results are compared directly and the area of changes extracted (Singh, 1989; Jensen, 1996; Yuan et al., 1999). Supervised and unsupervised classifications are used in this approach. Individual classification of two image dates minimizes the problem of normalizing for atmospheric and sensor differences between two dates (Singh, 1989).

Accuracy dependency of the classification’s results is the main disadvantage of this method. Poor classification accuracy of individual classification leads to propagation of uncertainties in the change map, which results in inaccurate information of land-use changes. Shi and Ehlers (1996) described the uncertainty sources in change detection as error of the source image, classification methods, and determination of changes. They explained three main error sources in Maximum Likelihood (ML) classification-based change detection: (1) The process of training data collection is subjective; (2) the ML classifier assumes that the probability distribution of each class is normal, and (3) method used to determine changes (based on amount of uncertainties). Confusion matrix (error) is the common method to describe the uncertainty in a classified remote sensing image. Several error indicators can be derived from confusion matrix, such as error of commission.

1.8 Manual On-screen Digitization of Change : This method is usually used for high-resolution remote sensor data and scanned aerial photographs. This method can be used for updating erroneous government urban infrastructure databases using on-screen photo interpretation of high-resolution imageries (Jensen, 1996).


Two kinds of satellite data with different resolution and acquisition dates (ASTER and TM) were used in this study.The Advanced Space-borne Thermal Emission and Reflection radiometer (ASTER), has high spatial resolution , multi-spectral channels in the Visible and Near Infra-red Radiometer (VNIR),the Short Wavelength Infra-red Radiometer (SWIR), and the Thermal Infra-red Radiometer (TIR).Some ASTER characteristics are as follows:

Observation Spectral Coverage VNIR 3 Bands 0.52 ~ 0.86 µm
SWIR 6 Bands 1.60 ~ 2.43 µm
TIR 5 Bands 8.125 ~ 11.65
Spatial Resolution VNIR 15 m
SWIR 30 m
TIR 90 m
Swath Width 60 km

The Visible and Near Infra-red Radiometer (VNIR) measures solar radiation in three visible and near infra-red bands. The stereoscopic image sensor views 29.7 degrees forward of the band 3sensor in the same orbit. Stereoscopic observation capability is useful for geomorphologic studies and the creation of digital elevation models. Some of the VNIR functional parameters are as follows:

Spectral Bands
Band 1 0.52 ~ 0.60 µm
Band 2 0.63 ~ 0.69 µm Nadir looking
Band 3 0.76 ~ 0.86 µm
Stereoscopic Band
Band 0.76 ~ 0.86 µm Forward looking
Geometric Resolution 15 m
The data applied in this study were the first three bands of ASTER data (VNIR) with 15 meter resolution acquired on date of 11th March 2002, and LANDSAT 5-TM data acquired on 17th September 1990.

Reference Data
1:100,000 scale land-use map resulting from visual interpretation of TM data 1990
ASTER land-use map of the study area resulting from on-screen digitizing of the image based on field checking.

The main goal of this study is to reveal urban/agriculture changes using multi-resolution/multi-temporal satellite data. In order to extract changes, many above mentioned techniques were used.

Data Preparation
Geometric Correction
Multi-resolution/multi-temporal ASTER and TM data with acquisition dates of 11 March 2002, and 17 September 1990, respectively, were used in this study. Correction of ASTER data with 15 meter resolution was done by 1:25,000 scale topographic map using 19 control points with a RMSE as 0.37. The TM image with 30 meter resolution was registered by corrected ASTER image using18 control points with RMSE as 0.36.

Frequent acquisition of remotely sensed data makes it possible to use satellite imagery to determine type and extent of changes in the environment. Many digital change detection algorithms have been developed to reveal changes since the launch of ERTS-1 in 1972. The complexity of change detection procedure depends on the characteristics of data sets. Selection of a single sensor series, low cloud cover and matching dates of two image data can restrict this intricacy. The difference between spatial resolution and spectral band pass of two image dates acquired with two sensors complicates direct comparison of data to detect changes (Yuan et al., 1999). The most common method that can be used to detect the changes of multiresolution data set is post-classification approach. Singh (1989) explained that the problem of normalizing for atmospheric, and sensor difference between two image dates could be minimized using post-classification comparison, because two image dates are classified separately. Different classification algorithms (mainly using the spectral domain) have been developed, which have their own advantage or disadvantages. Shi and Ehlers (1996) reported about uncertainty propagation in classification based change detection. In this approach, the accuracy of the change product is the multiplication of the accuracies of the two theme maps. Hence a large number of erroneous change indications can be produced based on the errors in each theme map (Singh,1989). The following are shortcomings that limit the accuracy of classification based changedetection, described by Castelli et al., (1999):

  • Limited spectral separation of classes.
  • The “statistical independence” assumption- Pixel-wised classifications involve the DNvalues individually without considering the neighboring pixels.
  • The separation of the classification, and change detection steps- Change detection step uses the processed information from classification, not the original data.
  • The construction of training sets- Collection of ground truth is expensive, time-consuming, and sometimes it is impossible to assign the class label for spectral clusters. A less expensive alternative is obtaining training sets from photo interpretation that limits the resulting accuracy to the skills of human expert who constructs the training sets.
  • The classification taxonomies- Classification taxonomies are often based on land-use that cannot be identified using remotely sensed data in some cases, rather than land cover type(e.g., discrimination between reservoir and lake). Also, different types of land cover are often aggregated under a single class.
  • The intrinsic limitation of classifiers- Comparison of multi-resolution/multi-temporal image data in a change detection procedure requires more research that would involve new approaches of multi-scale analysis. In this study an unsupervised multi-level texture segmentation (combination of neural network and wavelet transformation) was used to detect urban/agriculture changes using multi-resolution/multitemporal image data.
Following are the outcomes of the study:

  1. In the case of multi-resolution data, direct comparison of two multi-level image dates is restricted because various spectral and texture phenomena exist at different scales and resolutions.
  2. The procedure should be used in the area with the significant changes.
  3. Second scale is more appropriate to reveal urban/agriculture changes due to smoothness and details of the images. A method is required to allocate the unique value to each smoothed area (representing one class) resulting from segmentation. Selection of the method depends on the number of classes, the spectral variability of the classes (urban has noisy appearance and complex nature), and the situation of the study area..
  4. The multi-level segmentation (the applied algorithm) is an unsupervised process so there is no need to select training data, which cause to save time rather than using classification method.
Some recommendations are as follows:

Investigation of either using the result of change detection algorithms (e.g., image differencing…) on initial image in multi-level segmentation or applying the change detection algorithm to the result of multi-level segmentation.?Further investigation in the role of texture maps in the result of multi-level segmentation in change detection of urban/agriculture, which requires study in different areas. ?Examination of multi-level segmentation (using both spectral and texture information) to recognize species and forest types. ?Using multi-level segmentation in other land cover change detection, which covers large area with low variability in DN values (e.g., recognition of clear-cut in forested area). ?Investigation of segmentation accuracies with the accuracy of conventional method (e.g., classification) by applying two methods in the same area.


  1. Acton, S. T. and J. Landis. Multi-spectral anisotropic diffusion. Int. J. Remote Sensing,18(13), 2877-2886, 1997.
  2. Biging, G. S., D. R. Colby and R. G. Congalton. Sampling systems for change detectionaccuracy assessment. In Remote sensing change detection: Environmental monitoring methodsand applications, edited by R.S. Lunetta and C. D. Elvidge. London, UK: Taylor & Francis,281-308, 1999.
  3. Bruzzone, L. and D.F. Prieto. An adaptive parcel-based technique for unsupervised changedetection. Int. J. Remote Sensin., 21(4), 817-822, 2000.
  4. Carvalho, L. M. T., L. M. G. Fonseca, F. Murtagh, and J. G. P. W. Clevers. Digital changedetection with the aid of multi-resolution wavelet analysis. Int. J. Remote Sensin.,22(18),3871-3876, 2001.
  5. Castelli, V., C. D. Elvidge, C. S. Li and J. J. Turek. Classification-based change detection :Theory and applications to the NALC dataset. In Remote sensing change detection:Environmental monitoring methods and applications, edited by R.S. Lunetta and C. D. Elvidge.London, UK: Taylor & Francis, 53-73, 1999.
  6. Chan, J.C.W., K.P. Chan, and A. G. O. Yeh. Detecting the nature of change in an urbanenvironment: A comparison of machine learning algorithms. Photogrammetric Engineering &Remote Sensing, 67(2), 213-225, 2001.
  7. Fukuda, S. and H. Hirosawa. Suppression of speckle in synthetic aperture radar images usingwavelet. Int. J. Remote Sensin., 19(3), 507-519, 1998.
  8. Fung, T. An assessment of TM imagery for land-cover change detection. IEEE Transactionson Geoscience and Remote Sensing,, 28(4), 681-684,1990.
  9. Hoffer, R.M. Biological and physical considerations in applying computer-aided analysistechniques to remote sensor data. In Remote Sensing: The quantitative approach, edited by P.H.Swain and S. M. Davis. U.S.A : McGraw-Hill, 1978.
  10. Horgan, G. Wavelets for SAR image smoothing. Photogrammetric Engineering & RemoteSensing, 64(12), 1171-1177, 1998.
  11. Howarth, P.J. and E. Boasson. Landsat digital enhancements for change detection in urbanenvironments. Remote Sens. Environ, 13, 149-160, 1983.
  12. Howarth, P. J. and G. M. Wickware. Procedures for change detection using Landsat digital48 Referencesdata. Int. J. Remote Sensing, 2(3), 277-291, 1981.
  13. Jensen, J. R. Introductory digital image processing: A remote sensing perspective. Secondedition, New Jersey, U.S.A: Prentice Hall, 257-278, 1996.
  14. Jensen, J.R. and D. L. Toll. Detecting residential land-use development at the urban fringe.Photogrammetric Engineering & Remote Sensing, 48(4), 629-643, 1982.
  15. Kwarteng, A.Y. and P.S. Chavez. Change detection study of Kuwait city and environs usingmulti-temporal Landsat Thematic Mapper data. Int. J. Remote Sensing, 19(9), 1651-1662, 1998.
  16. Lark, R. M. Geostatistical description of texture on an aerial photograph for discriminatingclasses of land cover. Int. J. Remote Sensing, 17(11), 2115-2133,1996.
  17. Li, X. and A.G.O. Yeh. Principal component analysis of stacked multi-temporal images forthe monitoring of rapid urban expansion in the Pearl River Delta. Int. J. Remote Sensing,19(8), 1501-1518, 1998.
  18. Lunetta, R. S. Applications, project formulation and analytical approach. In Remote sensingchange detection: Environmental monitoring methods and applications, edited by R.S. Lunettaand C. D. Elvidge. London, UK: Taylor & Francis, 1-19, 1999
  19. Lyon, J.G., D. Yuan, R.S. Lunetta, and C. D. Elvidge. A change detection experiment usingvegetation indices. Photogrammetric Engineering & Remote Sensing., 64(2), 143-150, 1998.
  20. Mallat, S. G. A theory for multi-resolution signal decomposition: The wavelet representation.IEEE Transactions On Pattern Analysis And Machine Intelligence., 2 (7), 674-693, 1989.
  21. Morisette, J. T., S. Khorram, and T. Mace. Land-cover change detection enhanced withgeneralized linear models. Int. J. Remote Sensing, 20(14), 2703- 2721, 1999.
  22. Myint, S. W. A robust texture analysis and classification approach for urban land-use and land-cover feature discrimination. Geocarto international,16(4), 27-37, 2001.
  23. PCI Geomatics User’s Guide, Richmond Hill, Ontario, Canada, 2001.
  24. Perona, P. and J. Malik,. Scale-space and edge detection using anisotropic diffusion. IEEETransactions On Pattern Analysis And Machine Intelligenc., 12(7), 629-639, 1990.
  25. Quarmby, N. A. and J. L. Cushnie. Monitoring urban land-cover changes at the urban fringefrom SPOT HRV imagery in South-east England. Int. J. Remote Sensing, 10(6), 953-963,1989.
  26. Ridd, M. K. and J. Liu. A comparison of four algorithms for change detection in an urbanenvironment. Remote Sens. Environ, 63, 95-100, 1998
  27. Shi, W.Z. and M. Ehlers. Determining uncertainties and their propagation in dynamic changedetection based on classified remotely-sensed images. Int. J. Remote Sensing, 17(14), 2729-2741, 1996.References 49
  28. Singh, A. Review Article: Digital change detection techniques using remotely-sensed data.Int. J. Remote Sensing, 10(6), 989-1003, 1989.
  29. Singh, V. K. Discrete wavelet transform based image compression. Int. J. Remote Sensing20(17), 3399-3405, 1999.
  30. Small, C. Multi-temporal analysis of urban reflectance. Remote Sens. Environ, 81, 427-442,2002.
  31. Stow, D. A. Reducing the effects of misregistration on pixel-level change detection. Int.J.Remote sensing, 20(12), 2477-2483, 1999.
  32. Stow, D. A. and D. M. Chen. Sensitivity of multi-temporal NOAA AVHRR data of anurbanizing region to land-use/land-cover changes and misregistration. Remote Sens. Environ,80, 297-307, 2002.
  33. Turker, M. and E. Derenyi. GIS assisted change detection using remote sensing. GeocartoInternational., 15(1), 49-54, 2000.
  34. Witkin, A. P. Scale-space filtering. Int. Joint Conf. Artificial Intelligence.,Karlsruhe, West Germany, 1019-1021, 1983.
© All rights reserved.