Change Detection Methods Landsat Satellite Images DNR Forestry uses Landsat Thematic Mapper (TM) images to map Minnesota land cover and to set priorities for forest inventory work. For a fuller introduction to TM images and to view all the satellite scenes in DNR's collection, you can jump from here to another ForNet application, ImageView. Landsat TM scenes are digital images, like pictures from digital cameras. Each is an array of dots called picture elements (pixels) in rows and columns. Every pixel (some 38 million in a full scene, each pixel covering about 1/4 acre) contains numerical values expressing its brightness. Whereas an ordinary digital camera records only blue, green and red brightness corresponding to the range of human vision, Landsat records an additional four sets from the near, middle and thermal infrared portions of the spectrum. Analysis Overview Digital images can be manipulated by arithmetic, as everyone who has brightened up a scanned snapshot on a PC knows. And just as time-lapse pictures can be used to illustrate natural processes, we can use Landsat images to show where changes have taken place between two images taken years apart. All we have to do is subtract a digital image taken at "Time 1" from a corresponding image taken at "Time 2." The results form a "difference image." Such image differencing is one of the simplest and most reliable change-detection techniques in remote sensing. Trouble is, all sorts of irrelevant things can cause differences between images. If there is a cloud in the Time 2 image and none in the Time 1 scene, it will create a big patch of "difference." If Time 1 is a spring scene and Time 2 a fall scene, we'll detect a lot of seasonal "differences" that don't mean much in terms of long-term landscape change. Or if the two images aren't accurately registered to each other beforehand, the misregistrations will appear as "changes." So it's necessary to set certain conditions before attempting change detection: Both images must show the same season--preferably summer, when vegetation conditions are relatively stable. The two images must be accurately registered ("matched up") to the ground and to each other. Cloud-affected areas must be removed from analysis. The two images must be "radiometrically calibrated" to minimize effects of variations in sensing instrument performance and atmospheric haze between the two dates. When these conditions are met, a difference image can reveal where significant changes have taken place in surface phenomena like vegetation. But that still may not be very informative. In Minnesota, the biggest vegetation changes between successive summers take place on the farm, as crops and cultivation methods rotate from field to field. These bold shifts--which don't mean much in terms of landscape ecology--overwhelm the subtler patterns of forest change we seek to trace. So before image differencing, we must also: Broadly differentiate between forest and nonforest areas in both the Time 1 and Time 2 images. Exclude from consideration all areas shown to be nonforest in both images. If we carry out these preparatory steps satisfactorily, the pattern of between-date differences can tell us something about what kind of change is taking place in the forest. Green plants tend to absorb red light, for example, so a substantial increase in red brightness may indicate less vegetation at a location in the image. Near-infrared light is reflected by plants, so an increase in its brightness may mean more vegetation. A typical look at the ChangeView interface, displaying data from an area of severe windstorm damage, is shown below. We present the Time 1 image on the left and the Time 2 image on the right, with the difference image between them.