Classification of Land Cover
Methods


Background:
The land cover mapping technique developed by the Florida Fish and Wildlife Cooperative Unit synergizes existing geospatial information with current Landsat imagery. The primary data used in this method are:

For each Landsat scene location, the dates of the imagery are reviewed to determine if two images are available that are approximately one year apart and in different seasons. If this criteria is meet, the multi-date approach is used. This approach capitalizes on the seasonal variation of vegetation that can be detected using Landsat TM imagery. Each scene is classified independently and subsequently, the classified images are merged to create a seamless mosaic. Following is a description of the multi-date technique (see Figure 1. for a graphic overview). Since the single date technique is identical, except for the band combinations used and some pre-processing steps, it will not be presented, but a detailed figure is provided (see Figure 2.).

The method outlined below is divided into a pre-processing and post-processing phase. In the pre-processing phase, the Landsat imagery is made usable for classification by: checking for database consistency with the land use/ land cover maps, correcting for atmospheric affects as needed, and computing the first three spectral rotations of the tasseled cap algorithm (brightness, greenness and wetness). During the processing phase, an iterative unsupervised classification algorithm is used in a knowledge based cluster busting method.

Classification followed The Nature Conservancy, Southeastern Region classification scheme. For detailed descriptions of the TNC scheme ask for the report:

International Classification of Ecological Communities:

Terrestrial Vegetation of the Southeastern United States

The Nature Conservancy

101 Conner Hill Dr., Suite 202

Chapel Hill, NC 27514

(919) 967-5493

 

Methods:
Multi-date Land Cover Mapping Technique:

When adequate information is available, multi-temporal image classification procedures are used. Many different techniques and band combinations can be used to classify Landsat imagery. However, Hill and Megier (1988) found multi-temporal image classification using the Tasseled-Cap algorithm resulted in improved land cover mapping. Using similar procedures, each scene was classified as follows.

Pre-processing:
In the pre-processing phase (
Figure 1. Pre-Processing), the data used in the classification methodology is first checked for database constancy and co-registration. The images are normalized and the tasseled cap transformation for each image is computed. A discussion of each component is presented below.

Database consistency: One component of the classification methodology uses land use/land cover maps. These maps were obtained from Florida’s water management districts and revised with Landsat imagery from 1992 and 1993. However, imagery from 1993 and 1994 were also used in the classification. These images are overlayed with the land use maps and checked for database consistency. If consistent positional errors are present, an affine transformation and nearest-neighbor resampling is used to co-register the image to the land use/land cover maps. Additionally, poor co-registration between multi-date image bands will confuse the classification and therefore, a poor classification will result. Each image, potentially having been geo-rectified by different people and using different ground control, is co-registered as needed.

Normalization: Prior to any multi-date image analysis, it is necessary to correct for differences in sensor offset and gain and also scene illumination caused by different seasons and atmospheric conditions. The difference in overall brightness between the images was normalized using image regression. This method is well suited for multi-temporal analysis where care must be taken not to adjust the image for the seasonal variation of vegetation. A regression model to account for these differences is obtained by first identifying about 15 to 20 bright and dark objects in each scene and, for each band, recording the digital number (DN). An example of a dark object is uniform non turbid man-made lakes. Good bright object are: airport runways, large roads, beaches, dense urban areas, and exposed soils. Once these values are compiled, a linear regression model is computed with the darker of the two images assigned to the X variable. This insures that positive corrections are made such that when applied no negative numbers resulted in the output image however, compression of values near 255 can occur. For each band a linear regression model and an associated scatter plot are computed. If the model has a correlation coefficient (r) higher than 95% and the scatter plot does not have significant outliers, the linear model is used. When outliers are detected, they are removed and the regression model is recomputed.

Tasseled-Cap: The last step in the pre-processing phase is to compute the tasseled cap algorithms. The tasseled-cap transformation provides a mechanism for data reduction and enhanced image interpretation by emphasizing the structures in the spectral data which arise as a result of particular physical characteristics of scene classes (Crist 1985). The equations for this spectral index have been supplied by ERDAS, Inc. Atlanta, Georgia. Brightness, greenness, and wetness are computed for each image and combined into one multi-temporal data set.

Processing:
In the processing phase (
Figure 1. Processing), for each Landsat scene location two classified images are created, a classified image of natural areas, and a classified image of some urban and agricultural areas. This segmentation assistes with reducing the overall spectral variability of the image and hence results in an improved unsupervised classification. Using Florida’s modified land use/land cover classification system (FLUCCS), codes that represent natural areas are used to isolate natural areas in the imagery. Similarly, FLUCCS codes are used to isolate urban and agricultural areas which may contain natural areas (e.g. FLUCCS code for institution lands fall under the urban codes however, these lands many times are natural).


Classification of Natural Areas:
For the natural areas, Imagine’s interative unsupervised classification routine (ISODATA) is used to create 6 signatures. These signatures are then used with the minimum distance to mean classifier to classify natural areas to 6 classes. Next, for each class of the classified natural areas image, 3 to 5 new classes are created using ISODATA and the minimum distance to mean classifier. At this point, up-to 30 classes could exist. These classes are then summarized against ground truth information obtained from videography. This summary is then used in a knowledge based class combining or class "busting" method.

Using this approach, classes with multiple labels are identified and "busted" (classified into more classes). Conversely, multiple classes with the same class name are combined.

When possible, post-classification sorting is used to refine the classification. In some instances certain classes can be separated based on ancillary information such as NWI or soils information. For example, this refinement allowes for the separation of some class that could be differentiated and reclassified based on fresh water or salt water NWI classes. At a pixel level, this could easily be done however, treating contiguous pixels as a group and reclassifying all pixels in that group based on a majority NWI class value will eliminate potential "salt and pepper" and it will create a more natural split of the classes. Splitting the classes involved: isolating and recoding the classes of interest, clumping the classes, summarizing these clumps against the NWI coverage, and splitting the classes through crosstabulation.


Classification of Urban Areas:
The classification of urban areas for the GAP Anaylsis project is being performed by the
Coastal Service Center (C-CAP) Program. However, an unsupervised classification of these areas was performed. For urban/agricultural areas (see Figure 1. for the codes used), 15 to 30 classes were created using ISODATA and the minimum-distance-to-mean classifier. Classes that represented natural areas were isolated and labeled.

Finally, the classified images, with class names, were merge to create a seamless mosaic. Classes from land use/land cover maps were used to populate excluded lands in the mosaic. The descriptive form of the FLUCCS codes were used as labels.

References

Crist, E. P. (1985). "A TM Tasseled Cap Equivalent Transformation for Reflectance Factor Data." Paper, Elsevier Science Publishing Co., Inc., New York, New York.

Hill, J. and J. Megier (1988). "The Use of Multi-Temporal TM Tasseled Cap Features for Land Use Mapping in European Marginal Areas An operational Approach." International Geoscience and Remote Sensing Symposium (IGARSS), v2, pp.798-801.