Prior to conducting the change analysis, the TM images to be used for the analysis needed to be spatially and radiometrically calibrated. Spatial calibration was accomplished using a Space Imaging proprietary software called AutoWarp, which was able to achieve, in most cases, a sub-pixel co-registration accuracy.
Once the imagery was spatially calibrated, radiometric calibration was necessary, and this was done using the methods of Eckhardt et al. (1990). 15 - 25 areas were chosen throughout a scene, usually water bodies and bare sites that did not change between the dates. Data from these area were output to a regression algorithm, which calculated intercepts and first- order coefficients for each band. These values were then used to balance the band histograms.
Several techniques were explored in order to choose a suitable methodology for detecting change from the calibrated imagery: Band7/4 ratio differencing, Synthetic Color Transformation (SCT), differences in Tasseled Cap Bands 1,2, and 3, and differences in Principal Components Bands 1 and 2 (additive). The technique chosen was a combination of Tasseled Cap Band 1 and NDVI difference. The two seemed to compliment each other well, often identifying the same change, but occasionally identifying change that the other missed.
To identify change in a scene, the Tasseled Cap Bands 1 and NDVI bands from the early- and late-date images were differenced, and change thresholds were selected empirically and conservatively, erring on the side of commission, or labeling pixels of “no change” as “change” which could be eliminated in a later step. Prior to adding the two results, the images were clumped and sieved to remove any clump of pixels of less than five pixels. The result was the binary change image.
In order to classify the change previously identified, a regression technique was used in which the unchanged portions of the late-date basemap were used as training data. The basemap was subset to the scene being analyzed and to the unchanged portions as identified in the change detection phase. Using the resulting thematic layer as the dependent variable in the regression would introduce a large amount of noise into the system, due to misclassifications, speckle, and slight misregistrations between the basemap and the imagery. To correct this, the basemap was clumped and sieved, removing any clumps with fewer than 9 pixels.
The clumped and sieved basemap was input into the regression algorithm as the dependent variable, along with a Tasseled-Cap transformation of the 1995-era Landsat scene as the dependent variable. The result was subset to the changed areas, and summarized against the basemap.
In the change detection phase, one change scenario that was commonly detected was phenological or rotational changes in cultivated fields. In the analysis region of Southern Ontario, this was the predominant change type detected. Once the early-date classification had been processed to a point that the analyst was confident in the accuracy of the Cultivated class, those pixels which were classified Cultivated in both dates could be dropped out, or coded to “no change”. In order to get to that point, the Grassland and Cultivated classes needed further processing to separate them. This was done by using supervised classification on those classes previously defined as Grassland and Cultivated with training sites selected using the basemap and 1995-era orthophotos as guides.
At this stage, another clump and sieve was run to eliminate stray pixels. These generally occurred at the edges of cultivated fields and along linear features which were imaged differently in the Landsat scenes used in the change detection (mixed pixels).
Small clumps of high class variety still persisted in the map at this point. Class variety refers to the number of distinct land cover classes present in a clump of pixels. The vast majority of these high-variety small clumps were false change. A threshold of 10 pixels was decided upon to eliminate these clumps, and a high class variety was also a criteria. A variety layer was made, with the number of classes in a contiguous clump as the value. This was intersected with a layer showing the number of pixels in a clump, using the equation (classvariety * 10000 / #pixels). Clumps with a value of 3000 or greater and a size of 10 or less were eliminated. The thresholds were arrived at empirically, by inspection.
The final step in producing the change map was editing of the changed areas. A bivariate map was produced, and the changed areas were examined at a 1:40,000 scale or finer against the TM imagery, concentrating on the early-date classification of those areas. During this process, additional speckle and artifacts imposed by linear features were removed.
During the change detection process, some cultivated or grass features in fields were identified as change, where other portions of the same fields were identified as unchanged. The decision was made to classify the entire field as changed if a portion was identified as change, since in reality a field usually changes as a unit. To accomplish this, the 1995-era imagery was segmented using the Ecognition software, producing polygons of similar spectral response. The polygon size was set to the minimum size that the hardware could handle. (The smaller the polygon size, the more numerous they were, and the more memory required for processing). The polygons were then imported into GRID format, and a Zonal Majority function was run using the edited change map to label the polygons. The Cultivated and Grass classes were subset and overlaid on the change map. The overlay was run so as to not overwrite any other classes of change. This was the final step in the production of the early-date classification.
Eckhardt, D. W., J. P. Verdin, and G. R. Lyford. 1990. Automated update of an irrigated lands GIS using SPOT HRV imagery. Photogrammetric Engineering and Remote Sensing 56:1515-1522.
Includes areas dominated by single stemmed, woody vegetation unbranched 0.6 to 1 meter (2 to 3 feet) above the ground and having a height greater than 6 meters (20 feet).
Includes areas in which more than 67 percent of the trees remain green throughout the year. Both coniferous and broad-leaved evergreens are included in this category.
Areas dominated by woody vegetation less than 6 meters in height. This class includes true shrubs, young trees, and trees or shrubs that are small or stunted because of environmental conditions.
Includes all nontidal wetlands dominated by woody vegetation greater than or equal to 6 meters in height, and all such wetlands that occur in tidal areas in which salinity due to ocean-derived salts is below 0.5 parts per thousand (ppt).
Includes all nontidal wetlands dominated by woody vegetation less than or equal to 6 meters in height, and all such wetlands that occur in tidal areas in which salinity due to ocean-derived salts is below 0.5 ppt.
Includes all nontidal wetlands dominated by trees, shrubs, persistent emergents, emergent mosses, or lichens, and all such wetlands that occur in tidal areas in which salinity due to ocean- derived salts is below 0.5 ppt.
Includes all tidal wetlands dominated by woody vegetation less than or equal to 6 meters in height, and all such wetlands that occur in tidal areas in which salinity due to ocean-derived salts is above 0.5 ppt.
Characterized by erect, rooted, herbaceous hydrophytes (excluding mosses and lichens) that are present for most of the growing season in most years. Perennial plants usually dominate these wetlands. All water regimes are included except those that are subtidal and irregularly exposed.
Characterized by substrates lacking vegetation except for pioneering plants that become established during brief periods when growing conditions are favorable. Erosion and deposition by waves and currents produce a number of landforms, such as beaches, bars, and flats, all of which are included in this class.
Includes persistent snow and ice persist for greater portions of the year.