Different Change Detection Techniques

Table of Contents

Introduction...................................................3
Digital Change Detection Process...............................4
Description of the most commonly used change detection methods.5 I. Post-Classification Comparison..........................5 II. Direct Classification...................................6 III. Principal Component Analysis (PCA)......................6 IV. Image Differencing......................................8 V. Change Vector Analysis (CVA)............................9

Relative accuracy of the most commonly used change detection methods........................................................9

I. Post-Classification Comparison.........................10 II. Direct Classification..................................11 III. Principal Component Analysis (PCA).....................11 IV. Image Differencing.....................................12 V. Change Vector Analysis (CVA)

Conclusion....................................................14

References....................................................15

Introduction
Remote sensing change detection has been defined as the process of identifying change in the state of an object or phenomena through the detection of differences between two or more sets of images taken of the same area on different dates (Wang, 1993). The underlying assumption is that changes on the ground cause significant changes in image pixel values (Zhang et al., 2002). Change detection is a vital technique in remote sensing because it plays a role in monitoring and managing natural resources and urban development providing quantitative analysis of the spatial distribution of the population of interest.
Change detection is useful in such diverse applications as land use change analysis, monitoring shifting cultivation, assessment of deforestation, study of changes in vegetation, seasonal changes in pasture production, damage assessment, crop stress detection, disaster monitoring, day/night analysis of thermal characteristics as well as other environmental changes. When monitoring natural resources there are four basic aspects that are relevant; 1) Detecting that changes have occurred, 2) Identifying the nature of the change, 3) Measuring the areal extent of the change, 4) Assessing the spatial pattern of the change (Macleod and Congalton 1998) and by 1995 there were eleven different change detection techniques documented in the literature; 1) mono-temporal change delineation, 2) delta or post-classification comparison, 3) multidimensional temporal feature space analysis, 4) composite analysis, 5) image differencing, 6) image ratioing, 7) multitemporal linear data transformation, 8) change vector analysis, 9) image regression, 10) multitemporal biomass index, 11) background subtraction. The purpose of this paper is to outline the change detection process, discuss the five most common change detection techniques; post-classification comparison, direct classification, principal component analysis, image differencing, and change vector analysis (CVA) and discuss the literature on the relative accuracy of each method.
Digital Change Detection Process Listed below are the general steps used to conduct digital change detection using remote sensing data (Ridd, M., and J. Hipple, 2006). 1. State the Change Detection Problem a. Define the study area b. Specify frequency of change detection (e.g. seasonal, yearly) c. Identify classes from an appropriate land cover classification system 2. Consideration of Significance When Performing Change Detection d. Remote Sensing System Considerations i. Temporal resolution ii. Spatial resolution iii. Spectral resolution iv. Radiometric resolution e. Environmental Considerations v. Atmospheric conditions vi. Soil moisture conditions vii. Phenological cycle characteristics viii. Tidal stage 3. Image Processing of Remote Sensing Data to Extract Change Information f. Acquire Appropriate Change Detection Data ix. In situ and collateral data x. Remotely sensed data 1. Base year (Time n) 2. Subsequent Year(s) (Time n-1 or n+1) g. Preprocess the Multiple Date Remotely Sensed Data xi. Geometric registration xii. Radiometric correction (or normalization) h. Select Appropriate Change Detection Algorithm i. Apply Appropriate Image Classification Logic If Necessary xiii. Supervised, unsupervised, hybrid j. Perform Change Detection using GIS Algorithms xiv. Highlight selected classes using change detection matrix xv. Generate change map products xvi. Compute change statistics 4. Quality Assurance and Control Program k. Assess Statistical Accuracy of: xvii. Individual date classifications xviii. Change detection products 5. Distribute Results l. Digital products m. Analog (hardcopy) products
Description of the most commonly used change detection methods I. Post-classification comparison
The most intuitive method of change detection is post-classification comparison, an overlay of two independently produced classified images, which results in a complete descriptive matrix of the changes between the two dates. The individual classifications can either be supervised or unsupervised. Unfortunately, every error in the individual data classification maps will also be present in the final change detection map. Therefore, it is imperative that the individual classification maps used in the post-classification change detection methods be as accurate as possible (Jensen, 1996).
Post-classification comparison is different from all other change detection methods, because it alone involves a comparison of classified data. All other methods involve a comparison of continuous data, and are therefore sometimes grouped as spectral change analysis. In many cases, however, it is necessary to apply classification to the output of spectral change analysis, in order to provide simple, summary information on changes.
II. Direct classification
In contrast to post-classification comparison, which involves two separate classifications, it is possible to combine the multitemporal images into one large data set, and classify the spectral changes as a single, direct classification. As with post classification comparison, either a supervised or an unsupervised approach can be used (Singh, 1989). The combined multitemporal spectral data set is also a starting point for principal component analysis. While this technique requires only a single classification, it is a very complex one, in part because of the added dimensionality of two dates of data (Coppin et al. 2004). The potential disadvantages of the direct multitemporal method include redundancy in spectral information present in some of the bands, and difficulty in labeling change classes if an unsupervised approach is used.
III. Principal component analysis (PCA)
PCA is a statistical technique used extensively in remote sensing image analysis in applications well beyond that of change detection (Warner, 1999). PCA is a method that produces new images, known as principal components, which are each linear combinations of the original bands. The principal components are uncorrelated with each other, and are arranged such that the first component contains the most variance possible from the original data, with each succeeding component containing the maximum amount of the remaining variance (Young and Wang, 2001). PCA has been used for determining the intrinsic dimensionality of multispectral imagery (Swain and Davis, 1978), data enhancement for geological applications (Prakash and Gupta, 1998) and land cover change detection (Fung and LeDrew, 1987).
PCA is often used to reduce the dimensionality of the data without reducing the overall information content. This dimensionality reduction is based on the assumption that high order principal component images contain most of the information present in the original image (Warner, 1999). Principal components can be produced from the eigenvectors of the covariance matrix (unstandardized) or correlation matrix (standardized principal components) (Li and Yeh,
1998). These matrices may be extracted from a subset area, or the total study area.
PCA for change detection requires first stacking two or more single date images to form a combined multitemporal image. The spectral-temporal transformation of the multi-date TM data usually creates some components correlated with change (Guild et al., 2004). The highest order principal components ideally represent the spectral distribution of the land cover that is consistent with time, while land cover changes are represented in the next highest order components (Byrne et al., 1980). Image noise should be concentrated in the lowest order bands. In practice, however, land cover changes can be mixed with the spectral variation of unchanged classes, making interpretation of the PCA images complex.

IV. Image differencing
Image differencing is one of the most widely used techniques for change detection. Image differencing is simple to implement, and the output image is relatively easy to interpret. The procedure involves the subtraction of one or more bands of an image of one date from the equivalent bands of another date. The subtraction results in large positive or negative values in areas of changes in spectral radiance, and values close to zero in areas of no change (Sohl, 1999). A critical element of the image differencing method is deciding where to place the thresholds between change and no change pixels, because noise and natural variability in the spectral classes will result in a certain minimum amount of spectral change, even if the classes have not changed. Singh (1989) and Toll (1982) suggested selecting thresholds on the basis of the number of standard deviations from the mean value of the differenced image.
Studies of image differencing with Landsat Multi Spectral Scanner (MSS) data have shown that the red band is useful in detecting urban development, while the near infrared (NIR) band is good for detecting change within agricultural land (Fung, 1990). In an early study, it was reported that the detection of residential land use development at the urban fringe showed a significant improvement when image differencing included both Thematic Mapper (TM) band 5 and texture information (Toll 1982).

IV. Change vector analysis (CVA)
CVA yields information about the degree and type of spectral changes by calculating a vector magnitude and direction in multispectral change space for each pixel (Michalek et al., 1993). The vector magnitude, or the intensity of the change, is computed by determining the Euclidean distance between end points through n-dimensional change space. The direction of the change, measured as the angle of the vector, indicates the nature of the land-cover change process (Lambin and Strahler, 1994a). A particular advantage of CVA is the potential ability to process any number of spectral bands. This is important because not all changes are easily identified in any single band or spectral feature.
The change vector direction is useful in discriminating different types of changes, while the vector magnitude is useful for comparison of the intensity of change within and among change types. Also, CVA can be useful in long-term, multi-interval monitoring studies. By extracting consistent change vector information for multiple time intervals, and merging this information in one or more change images, the change trajectory of a given site may be observed over time, and spectral, spatial and temporal variations may be computed and contrasted (Johnson and Kasischke, 1998).
Relative accuracy of the most commonly used change detection methods
Much of the change detection literatures consist of experimental studies comparing the accuracy of different change detection methods. In general, the results obtained from these experimental comparisons are somewhat contradictory. For example, in studies that compared PCA to other change detection methods, Li and Yeh (1998) and Muchoney and
Haack (1994) found that PCA produced the highest accuracy, whereas Macleod and Congalton (1998) and Singh (1989) found that PCA produced the lowest accuracy. Furthermore, many studies simply conclude that there is no single best method, and suggest that each approach may have its own strengths (Ridd and Liu 1998, Sohl, 1999).
This lack of consistency may be attributed to the range of localities where the various studies have been carried out, each with different physical geography and land use patterns. It is noticeable that the research design of studies that compare change detection methods has almost always focused on a single location, and a single set of imagery. I. Post-classification comparison
While post-classification comparison for change detection has been criticized for relying upon the accuracy of the initial classification of the two individual images classification, Toll et al. (1980), and Singh (1989) argue that post-classification can be considered the superior change detection method, and therefore should be the standard for evaluating the results of other methods. Many change detection studies do not have independent ground data, and therefore the assumption that post-classification can be used to evaluate the other methods is attractive. Others have both pointed out that post-classification comparison is likely to be the best approach when anniversary images (i.e.images acquired in the same month and day of different years) are not used in the change analysis (Sohl, 1999, Luque, 2000). In addition, the post-classification comparison technique has the advantage of compensating for varied atmospheric conditions between dates, because each classification is independently produced. On the other hand, the difficulty of producing comparable classifications on different images may give poor results and the use of pre-defined land cover classes in post-classification comparison by definition implies a loss of the information present in the continuous range of image values provided by the original data (Toll, 1982).
II. Direct classification
The success of a multitemporal composite analysis depends upon the extent to which change classes are different spectrally from the no change classes. As with all spectral methods of change detection, errors are likely if there are spectral differences between two dates for pixels that have not changed in land cover. For example, vegetation spectral reflectance may vary due to seasonal or weather conditions, without any notable land cover change. In addition, the complexity and dimensionality of the classification can be quite great, and if all bands from each date are used, there may be substantial redundancy in their information content (Lillesand et al., 2004, Michener and Houhoulis 1997).
III. Principal component analysis
PCA has been found to provide accurate and effective change analysis. For example, Byrne et al. (1980) have shown that PCA can be effective for monitoring of land cover change using multitemporal Landsat MSS data. In a study by Li and Yeh (1998) obtained a relatively high overall accuracy with a principal component analysis of urban expansion in the Pearl River Delta. Despite these positive results, PCA is known to be sensitive to noise (Muchoney and Haack, 1994). A difficulty with implementing PCA-based change detection is that decisions regarding the implementation can have a large effect on the outcome of the analysis. For example, in a study of land cover change in the Kitchener-Waterloo-Guelph area, Fung and LeDrew (1987) found that the eigenvectors used for rotation could vary greatly depending upon the choice of using standardized or non-standardized data, and whether the whole data set, or a subset, was used for generating the image statistics. They suggested that, in general, the eigenstructure derived from the entire data set is more valid for land cover change detection.
IV. Image differencing
Image differencing has also been proposed as the method likely to have the highest accuracy (Macleod and Congalton 1998, Sohl 1999). Singh (1989) concluded that simple techniques, such as image differencing, tend to perform better than complex methods, such as principal component analysis. This view is supported by Macleod and Congalton (1998), who found image differencing performed significantly better than post-classification comparison and principal component analysis for mapping changes in eelgrass populations in Great Bay, New Hampshire.
Image differencing does however have some difficulties, including its sensitivity to mis-registration and mixed pixels. Also, simple image differencing fails to consider the starting and ending location of a pixel in the multi-dimensional feature space. Moreover, image differencing does not provide direct information on the nature of the change (Singh, 1989).
V. Change vector analysis (CVA)
The strength of the CVA approach is that it alone provides information separately on the type and degree of change (Sohl, 1999). CVA potentially allows processing of the full dimensionality of multispectral data, thus providing the potential to detect all the changes present in the data (Lambin and Strahler 1994a). However, most CVA analyses collapse the dimensionality down to just two or three dimensions (Sohl, 1999). The CVA method can be a valuable tool for coastal resource surveys and monitoring, especially for identifying suspected areas of change prior to more detailed on-site observation (Michalek et al. 1993). Lambin and Strahler (1994b) showed that CVA combined with a principal component analysis is effective in detecting and categorizing inter-annual change. Sohl’s (1999) study compared four change detection methods in the Abu Dhabi Emirate using Landsat TM data found that, although quantitative values of changes were most accurately provided by image differencing, CVA provides rich qualitative detail about the nature of changes. A study performed by Dhakal et al. (2002) found that among the different change detection methods used to map areas affected by flooding and erosion, CVA was the most accurate technique.
However, it was pointed out that CVA is very sensitive to the quality of the image radiometric normalization. An additional issue is that CVA lacks an automatic method to determine the threshold of change magnitude between change and no-change pixels. An iterative approach was developed using the classification of training data to select optimal change threshold Chen et al. (2003).
Conclusion
A large number of change detection techniques have been developed. The most common techniques for change detection are post-classification comparison, direct classification, principal component analysis, image differencing, and change vector analysis (CVA). These techniques are extremely useful and should be used in accordance with the sought after specific results. Along with ancillary data, multi-source data, Previous research shown that a combination of two techniques can often improve the results and implementation of several change detection techniques followed by an integration of the results could be the most effective way to detect change in a wide range of applications. The application of multi-source data and ground truth when possible should be incorporated as well.

References
Byrne, G., P. Crapper, and K. Mayo, 1980. Monitoring Land-Cover Change by Principal Components Analysis of Multitemporal Landsat Data, Remote Sensing of Environment, 10: 175-184

Chen, J., P. Gong, C. He, R. Pu, and P. Shi, 2003. Land-Use/Land-Cover Change Detection using Improved Change-Vector Analysis. Photogrammetric Engineering & Remote Sensing, 69: 369-379.

Coppin, P., I. Jonckheere, K. Nackaerts, B. Muys and E. Lambin, 2004. Digital Change Detection Methods in Ecosystem Monitoring: A Review, International Journal of Remote Sensing, 25: 1565-1596

Dhakal, A., T. Amada, M. Aniya, and R. Sharma, 2002. Detection of Areas Associated With Flood and Erosion Caused by a Heavy Rainfall Using Multitemporal Landsat TM Data, Photogrammetric Engineering & Remote Sensing, 68: 233- 239.

Fung, T., 1990. An assessment of TM Imagery for Land-Cover Change Detection, IEEE Transactions on Geoscience and Remote Sensing, 28: 681-684.

Fung, T., and E. LeDrew, 1987. Application of Principal Component Analysis to Change Detection, Photogrammetric Engineering & Remote Sensing, 53: 1649-1658

Guild, L., W. Cohen., and J. Kauffman, 2004. Detection of Deforestation and Land Conversion in Rondônia, Brazil Using Change Detection Techniques, International Journal of Remote Sensing, 25: 730-750.

Jensen, J. R., 1996. Introductory Digital Image Processing: A Remote Sensing Perspective, Prentice-Hall, Inc., Saddle River.

Johnson, R., and E. Kasischke, 1998. Change Vector Analysis: A Technique for the Multispectral Monitoring of Land Cover and Condition, International Journal of Remote Sensing, 19: 411-426

Lambin, E., and A. Strahler, 1994a. Change Vector Analysis in Multitemporal Space: A Tool to Detect and Categorize Land-Cover Change Processes Using High Temporal-Resolution Satellite Data, Remote Sensing of Environment, 48: 231-244.

Lambin, E., and A. Strahler, 1994b. Indicators of Land-Cover Change for Change-Vector Analysis in Multitemporal Space at Coarse Spatial Scale, International Journal of Remote Sensing, 15: 2099-2119

Lillesand, T. M., R. W. Kiefer, and J. W. Chipman, 2004. Remote Sensing and Image Interpretation, 5th edition., Wiley, New York

Li, X., and A. Yeh, 1998. Principal Component Analysis of Stacked Multi-Temporal Images for the Monitoring of Rapid Urban Expansion in the Pearl River Delta,International Journal of Remote Sensing, 19: 1501-1518

Luque, S., 2000. Evaluating Temporal Changes using Multi-Spectral Scanner and Thematic Mapper Data on the Landscape of a Natural Reserve: The New Jersey Pine Barrens, a Case Study, International Journal of Remote Sensing, 21: 2589 – 2610.

Macleod, R., and R. Congalton, 1998. A Quantitative Comparison of Change Detection Algorithms for Monitoring Eelgrass from Remotely Sensed Data, Photogrammetric Engineering & Remote Sensing, 64: 207-216

Michalek, J. L., T. W. Wagner, J. L. Luczkovich, and R. W. Stoffle, 1993. Multispectral Change Vector analysis for Monitoring Coastal Marine Environments, Photogrammetric Engineering & Remote Sensing, 59: 381-384.

Michener, W., and P. Houhoulis, 1997. Detection of Vegetation Changes Associated with Extensive Flooding in a Forested Ecosystem, Photogrammetric Engineering & Remote Sensing, 63: 1363-1374.

Muchoney, D. M., and B. Haack, 1994. Change Detection for Monitoring Forest Defoliation. Photogrammetric Engineering & Remote Sensing, 60: 1243-1251

Prakash A., and R. P. Gupta , 1998. Land-Use Mapping and Change Detection in a Coal Mining Area - A Case Study in the Jharia Coalfield, India, 1998. International Journal of Remote Sensing, 19: 391 – 410
Ridd, M., and J. Hipple, 2006, Remote Sensing of Human Settlements American Society of Photogrammetry and Remote Sensing: Bethesda, Maryland, 2006
Ridd, M., and J. Liu, 1998. A Comparison of Four Algorithms for Change Detection in an Urban Environment, Remote Sensing of Environment, 63: 95-100.

Singh, A., 1989. Digital Change Detection Techniques Using Remotely Sensed Data, International Journal of Remote Sensing, 10: 989-1003

Swain, P. H. and S. M. Davis, 1978. Remote Sensing: The Quantitative Approach. McGraw Hill, New York.

Sohl, T., L, 1999. Change Analysis in the United Arab Emirates: An Investigation of Two Techniques, Photgrammetric Engineering & Remote Sensing, 65: 475-484.

Toll, D., 1982. Detecting Residential Land-Use Development at the Urban Fringe, Photogrammetric Engineering & Remote Sensing, 48: 629-643

Wang, F., 1993. A Knowledge-Based Vision System for Detecting Land Changes at Urban Fringes, IEEE Transactions on Geoscience & Remote Sensing, 31: 136-145

Warner, T., 1999. Analysis of Spatial Patterns in Remotely Sensed Data Using Multivariate Spatial Correlation, Geocarto International, 14 (1): 59-65.

Young, S. S., and C. Y. Wang, 2001. Land-Cover Change Analysis of China Using Global-Scale Pathfinder AVHRR Landcover (PAL) Data, 1982–92, International Journal of Remote Sensing, 22: 1457 – 1477

Zhang. Q., J. Wang, X. Peng, P. Gong., and P. Shi, 2002. Urban Built-Up Land Change Detection with Road Density and Spectral Information from Multi-Temporal Landsat TM Data, International Journal of Remote Sensing, 23: 3057-3078