Multiple methods for detecting land cover/habitat quality change using ADAR 5500 and IKONOS multispectral imagery were tested during the project. These included quick and largely automated methods of change detection such as multidate image overlay compositing, image differencing, and classification of difference images based upon positive and negative difference magnitude thresholds above or below which highlighted features are considered to represent real land cover change. These products were generated using multidate image digital number (DN) values, spectral vegetation indices (SVIs), spatial pattern indices (SPIs), and fraction images derived through spectral mixture analysis (SMA). In addition to these largely automated change detection approaches, change vector classification techniques requiring interactive operator involvement were assessed. For more specific information on theses procedures please refer to the project report (PDF).

 
Automated Change Detection Approaches
Mulitdate Overlay Composite Images

Multidate overlay composite images are generated by viewing corresponding image layers from two dates in separate color planes. For instance, the near-infrared waveband from the first date may be displayed in the red color plane, while the near-infrared waveband from the second date is displayed in the blue and green color plane. Using this technique, features exhibiting a change in the near-infrared waveband will be colored red or cyan (depending upon the direction of change) while areas of no-change appear as gray-scale. This visual analysis method enables an operator to quickly identify change areas within a multitemporal image set. However, changes are only displayed for one image layer at a time and no classification product (or change map) is generated.
Difference Images
Difference images generated by subtracting image values of one date from those of a corresponding layer from a second date highlight areas of changing land cover between dates. Difference images may be viewed a single layer at a time (gray-scale), or three difference image layers can be viewed together, one in each color plane. Viewing three difference image layers together is informative, as colors are indicative of the type of change which occurred. Displaying the red, near-infrared, and NDVI difference images in the blue, green, and red color guns, respectively, yielded the best qualitative results for enhancing land cover changes of interest to habitat managers. This was the case even though there is redundancy in the use of NDVI along with the near-infrared and red bands from which it is derived. However, no change map is generated using this method.

Difference image layers may be used to classify areas within the imagery as change or no-change based upon positive and negative difference magnitude thresholds, above or below which highlighted features are considered to represent real land cover change. The resulting map highlights areas of change (increase or decrease in image layer value between dates) on a per-waveband basis. The primary drawbacks to this approach are that there is no categorization of the type of change which occurred and change areas identified may not be consistent between the individual image layers.

Change Vector Classification
The change vector classification approach to change detection provided the greatest utility in terms of identifying and labeling land cover changes. Change vector classification was performed on multi-layer difference images using unsupervised classification. Since the red, near-infrared, and NDVI layers provided the most information on land cover changes with habitat reserves, change vector classification was generally performed using only difference image layers derived from these three input image types.

The processing and labeling steps for change vector classification included: 1) subset multitemporal image sets so that a common scene area is maintained; 2) mask (set to zero and remove from the analysis) portions of the imagery containing urban land cover (so that changes occurring in urban areas do not influence the change classification of habitat areas); 3) radiometrically normalize multitemporal image sets using the histogram matching approach; 4) create difference image by subtracting time 1 image from time 2 image; 5) classify the multi-layer difference image into 50 or more classes using unsupervised classification; 6) interactively determine and label no change classes; and 7) interactively determine and label change classes. It is worth noting that steps 1 through 3 were similarly performed for image differencing and difference image threshold-based classification change detection techniques discussed in the previous section.

Six primary spectral change classes were derived from change vector classification with ADAR and IKONOS multispectral imagery of southern California Mediterranean-type habitat. These classes were: increase and decrease in brightness, increase and decrease in vegetation cover, and increase and decrease in vegetation greenness associated with differences in plant phenology or timing of precipitation. A representation of the locations of these change types in the red vs. near-infrared difference image features-space is given in the figure below. Changes in image brightness (due to changes in soil exposure and non-photosynthetic vegetation cover/condition) are indicated by an increase or decrease in both the red and near-infrared wavebands. Changes in vegetation cover are indicated by either a increase in red and corresponding decrease in near-infrared, or vice-versa. This is the case because changes in vegetation cover have an inverse effect on the two wavebands; e.g., if vegetation cover increases, the near-infrared waveband values increases while red waveband values decreases. Differences associated with plant phenology and/or precipitation are characterized by an increase or decrease in near-infrared, with little or no change in the red waveband values.

Change vector class locations within the red
vs. near-infrared difference image feature-space.