Written by Jason Karl and Anne Axel
Other Names:
None known
Description
In the context of remote sensing, change detection refers to the process of identifying differences in the state of land features by observing them at different times. This process can be accomplished either manually (i.e., by hand) or with the aid of remote sensing software. Manual interpretation of change from satellite images or aerial photos involves an observer or analyst defining areas of interest and comparing them between images from two dates. This may be accomplished either on-screen (such as in a GIS) or on paper. When analyzing aerial photographs, a steroscope which allows for two spatially-overlapping photos to be displayed in 3D, can aid photo interpretation. Manual image interpretation works well when assessing change between discrete classes (e.g., forest openings, land use and land cover maps) or when changes are large (e.g., heavy mechanized maneuver damage, engineering training impacts). Manual image interpretation is also an option when trying to determine change using images or photos from different sources (e.g., comparing historic aerial photographs to current satellite imagery).
Automated methods of remote sensing change detection usually are of two forms: post-classification change detection and image differencing using band ratios. In post-classification change detection, the images from each time period are classified using the same classification scheme into a number of discrete categories (i.e., land cover types). The two (or more) classifications are compared and the area that is classified the same or different is tallied. With image differencing, a band ratio such as NDVI is constructed from each input image, and the difference is taken between the band ratios of different times. In the case of differencing NDVI images, positive output values may indicate an increase in vegetation, negative values a decrease in vegetation, and values near zero no change (see Rangeview example below). With either post-classification or image differencing change detection, it is necessary to specify a threshold below which differences between the two images is considered to be non-significant. The specification of thresholds is critical to the results of change detection analysis and usually must be found through an iterative process.
Similar Methods
Change detection can be applied to most other remote sensing methods. Below are some methods commonly used for change detection.
NDVI, Supervised Classification, MSAVI
Output
The goal of change detection is generally a layer or image that highlights areas that have changed between two (or more) time periods and the direction and magnitude of change.
Successful Rangeland Uses
Technical References
- University of Idaho’s College of Natural Resources has a good set of slides on remote-sensing change-detection methods: http://www.webpages.uidaho.edu/veg_measure/Modules/Lessons/Module%204%28Range%20Health%29/Forest_Health_Monitoring_overview_11-29-05.pdf
- Eastman, J.R. and M. Fulk. 1993. Time Series Analysis of Remotely Sensed Data Using Standardized Principal Components. In Proceedings: 25th International Symposium Remote Sensing and Global Environmental Change 1:485-496
Limitations
While change detection from remotely sensed images is helpful for assessing large landscapes, results are typically not as accurate or precise as those obtained from field monitoring. Vegetation data collected in the field may be as spatially accurate as the mapping product (GPS, topographic map) used to position the field samples, while spatial accuracy of vegetation data derived from satellite images is limited to the resolution of the pixel. Furthermore, while image pixels may be comprised of a mix of vegetation types, there will be only a single DN value for each pixel.
Using field collected data, one can clearly identify the nature of the change that has taken place between two sampling dates, however, when implementing the image differencing algorithm, this level of information is not available. While there are additional change detection algorithms which allow for identifying the nature of the change over time, these techniques require classification of each of the images.
Finally, change detection works best with easily distinguishable landscape features (e.g., tree encroachment); some aspects of land condition, such as cheatgrass invasion in western installations, are difficult to assess via remote sensing. Despite these limitations, automatic change detection of remotely sensed images provides an efficient, cost-effective method of assessing land cover change, especially of large landscapes. Additionally, availability of archived imagery allows for retrospective analyses¬ an option not available when there is no field data.
Implementing Change Detection
Implementing change detection using image-processing software requires a number of steps (Figure 3). First, you must obtain a time-series of images each having a cloud-free view of the study area. Ideally, the images should be on or near the anniversary date of the earliest image (or at least from the same phonological season) and also be obtained at approximately the same time of day. When possible, obtain the time series of images from the same sensor to maintain constancy in spatial resolution and image characteristics (e.g., look angle, spectral bands, and radiometric properties of the bands).
The image processing protocol is data-specific; while some images will arrive already orthorectified, others will need to be orthorectified using a digital elevation model to remove geometric distortions. In addition, most satellite images will arrive georectified to a standard map projection (e.g., Universal Transverse Mercator). Ensure that all images are in the same projection; if needed, re-project the images to a common map projection. If images have different a spatial resolution, resampling must be performed to achieve a uniform pixel size. In most cases the choice would be to resample all the images to the coarsest resolution of the input images. Next, the images must be co-registered so that points on one image match up with the same point on the other images; co-registration is also facilitated when using multiple images from the same sensor. Humidity, pollution, and other atmospheric conditions (e.g., smoke from fires) can degrade the quality of an image and influence the radiance values recorded by the imaging sensor. Although the use of band ratio techniques can lessen the impact of differences in conditions between images, atmospheric correction is still an important step in digital change detection. A discussion of atmospheric correction is beyond the scope of this whitepaper, and readers are referred to an introductory remote sensing text for more information (see section on additional information below). Once the images have been standardized and coregistered using the steps above, the same band ratio is applied to each image. A “difference” image is calculated by subtracting the values of the younger image from those of the older image on a pixel-by-pixel basis. Thresholds of change are determined, and “areas of interest” defined in order to mask out areas that are not of interest to the study. These are applied to the difference image to arrive at a final change detection product which can then be interpreted for your objective. It is important to note that change detection is often an iterative process and may involve a lot of back-and-forth to achieve a result that captures real landscape change and minimizes error. |
Software/Hardware Requirements
Implementing remote-sensing-based change detection can be done with most commercially available GIS/RS software. Specialized image-processing applications can make the job much easier.
- ENVI (http://www.exelisvis.com/ProductsServices/ENVI.aspx) – Perform change detection using standard ENVI tools. SPEAR Tools provide automated workflows for many common image processing tasks, including change detection.
- ERDAS (http://www.erdas.com/Homepage.aspx) – Perform change detection using standard Imagine tools or use their DeltaCue module.
- IDRISI Land Change Modeler (http://www.clarklabs.org/products/Land-Change-Modeler-Overview.cfm) – automated, easy-to-use toolset for implementing change detection methods
- Feature Analyst (http://www.vls-inc.com) – feature detection and extraction software includes pattern, shape, and proximity in identifying landscape features. Works better than traditional classification methods for mapping discrete features (e.g., roads, trails, disturbed areas).
- ArcGIS (http://www.esri.com) – The most ubiquitous GIS software package – can be used for manual change detection and photo interpretation.
Additional Information
- Links to other info sources
Who Is Using This Method?
- Organization information and/or project info
- Contact information (i.e., email, phone):
Existing datasets
- The Rangeview program at University of Arizona maintains a website where frequently-updated MODIS images can be accessed. Their online viewer can be used to look at two images side-by-side, or a number of standard “difference-from” products can be viewed. http://rangeview.arizona.edu
Implementation help
- Contact information for groups/people who can help implement this method either for free or for a fee.