Land Cover Monitoring: Change Detection Methodology


The Region 5 Land Cover Mapping and Monitoring Program (LCMMP) uses satellite imagery (MSS, TM, and ETM+) and change detection techniques to assess changes in landcover over 5-year intervals. The change detection process relies on the difference in reflectance between the time 1 and time 2 image where landcover has changed during the time interval that the two images were collected. Processing for estimating canopy cover change is accomplished in five primary steps: image preprocessing, GIS database building, change detection processing, thresholding and labeling/editing. These steps are followed by accuracy assessment, cause collection and field verification.


Terrain corrected image pairs are acquired and coregistered using a nearest neighbor resampling method to maintain the spectral integrity of the data. A maximum RMSE of 0.5 is required to minimize or eliminate false change. The image pairs are then radiometrically and atmospherically corrected to at-sensor reflectance. Lastly, the time 2 image is normalized to the time 1 image using an empirical line calibration approach (Schott et al., 1988).

GIS Database Building

GIS database building includes assembling data from multiple sources into a single vegetation layer. This process involves gathering data predominately from the Cooperative Vegetation Mapping Program (USDA-Forest Service, Region 5); however, data from the GAP Project and CDF are used for areas of Region 5 that have not been mapped by the Forest Service. The WHR classification system is used for the final vegetation coverage. Vegetation layers not in this classification system, such as CALVEG (USDA-Forest Service Regional Ecology Group, 1981), are crosswalked to the WHR classification.

Change Detection Processing

Change detection processing begins by applying the Kauth-Thomas (KT) transformation to both dates of co-registered imagery (Kauth and Thomas, 1976). This transformation uses model coefficients to produce three orthogonal axes: brightness, greenness, and wetness (BGW) (Crist and Cicone, 1984). Brightness identifies variation in reflectance, greenness is related to the amount of green vegetation present in the pixel, and wetness correlates to canopy and soil moisture. The regression line between B, G, and W is computed and an image based on the residuals is used for developing thresholds (personal communication, Richard Walker, PhD.-California Department of Forestry and Fire Protection). Using the residuals image helps reduce some of the seasonal variation during the thresholding process.




Thresholding, Stratification, and Classification

Thresholding of the BGW residuals is applied in a model to create a little or no change mask. Additionally, very bright features (e.g. snow, ice, rock, etc.), agricultural, urban, water, clouds, and smoke are masked out in both times. The non-masked areas are then subdivided or stratified. A model based on the BGW residuals and vegetation lifeform (e.g., conifer, hardwood and shrub) produces a GIS layer of potential increase or decrease of vegetation cover for each lifeform. An unsupervised and supervised classification is performed for each individual potential change-lifeform area resulting in pixel groups of similar levels of brightness, greenness and wetness resulting in a change image. These change images are mosaicked into one change image for the TM scene.

Change Labeling

Change labeling converts the change image to a change map that identifies decreases and increases in vegetation cover. These groupings are assigned to one of nine change classes (see metadata). Image appearance, photo interpretation, vegetation data, digital elevation models, bispectral plots (e.g., greenness vs. wetness) and other ancillary GIS layers aid in assigning the change classes. Refinement of the classes proceeds systematically across the entire project area.

Training and Accuracy

A stratified random sample of the change image is performed to measure selected points for quantifying canopy cover change. Digital orthophoto quads are used for the time 1 estimates and digital imagery is used for the time 2 estimates for canopy cover classification. The percent canopy estimates are then used to label the classes into change categories. A portion of the randomly selected sites are used for accuracy assessment. An error matrix is constructed to assess the accuracy of the final classified change map.

Cause Collection

Once the final change map is complete, cause collection for change areas is determined by using GIS overlays, fieldwork and photo interpretation. Areas without a causal agent identified become the focus of further field efforts and aerial photo interpretation. Despite these efforts, full coverage of cause verification is not always possible due to the large number of change areas, insufficient information or inaccessible lands.

Field Verification

A portion of the randomly selected sites are set aside for field verification. Field crews run transects through the polygon of change and measure canopy cover with a densiometer. Canopy cover measurements are then compared against in-house estimates to calibrate the estimates from photo interpretation.