Understanding why and how vegetation classification schemes are designed provides a foundation for informed usage of the various vegetation classification systems commonly used in remote sensing products. Some might think that devising vegetation categories should be relatively straightforward. However, in reality choosing from numerous vegetation criteria and maintaining consistency present substantial challenges that can easily lead to confusion.

There are two common misconceptions. First, there is a tendency to put too much detail into a list of categories. If some categories are defined using excessive detail while other categories are more generalized, this results in a mixed bag of relatively specific classes and more catchall classes, which makes implementation and mapping awkward if not inaccurate. Another drawback to this approach is that it seeks to use every available shred of information in making the classification decisions, which may be fine for the original mapping effort but much less effective for subsequent use in a different time or place. It usually gives better results to simplify definition of the classes, aiming for consistent accuracy rather than great detail.

The second caution highlights the difference between detection and classification (also a consistency issue). It is ill-advised to assume that if an instance of a candidate category is detected on an image, then all areas covered with that vegetation will also be discernable. The temptation presented by a few encouraging cases is a problem that can only be resolved using clearly defined criteria separating classes and a thorough examination to confirm that the cases were detected on the basis of such criterion.

After categories have been defined (examples of some criteria are listed in Chapter 4, Section 1.2, Table 1), it is necessary to determine which methods are best for sampling the vegetation and detecting categories within the area of interest. For example, if we wanted to assemble a global map, it would not be appropriate to use conventional aerial photography because the number of photos needed to provide global coverage is enormous. For global or regional expanses, coarse spacial resolution satellite imagery is more appropriate. The spatial scale of the final map must also be determined, taking into consideration both the demands of the application and limits of the technology being used to carry out the classification.

In summary, the best classifications result from designs incorporating consistent and image-significant criteria, clearly defined and not overly detailed categories, and use of an appropriate combination of image dates, bands, and software procedures. As the effort matures the various components in the sequence are tested and adjusted against each other. Classification should involve a creative and self-learning process that optimizes the results given the materials at hand. Though this iterative process should yield good confidence in the results, field verification is extremely important and necessary to validate the results.

A balance must be struck between the desired classes and the classes that can be accurately and efficiently delimited. It is not always possible to identify all of the classes that are wanted with sufficient accuracy using the technology that is affordable. Ultimately the entire initiative will be judged according to how well it predicted its own unavoidable errors, including those of miscategorization and misdelineation. Despite such errors, there are great benefits from choosing a vegetation classification scheme that represents the information with sufficient accuracy for the intended application.

2.1 Perspectives: Ground vs. Sensor

If you were sitting in the understory of an open canopy woodland, your perspective, view, and thoughts on classification would be very different from those you might have if you were looking down at the same woodland from a satellite. On the ground, you experience and see leaf and soil exposure, shadow components and moisture stress, while the satellite-sensor environment involves signal-to-noise ratios, measurement gain relative to reflectance variability as the satellite sensor's diode array measures reflected light intensity in split-second snapshots. The recorded intensities represent mixtures of the reflectance of dominant components within in each pixel.

If isolated, these two perspectives would entail very different classification criteria and category definitions. Some compromise between the two extremes is usually required, with the goal being to define categories that are both useful and discernable. The mapped results will not interest many decisionmakers or earth scientists if they contain remotely differentiated classes purely from image patterns that do not relate to specific plant properties. On the other hand, if too many plant parameters are listed in the legend, much of this information will be rendered obscure in the remotely sensed results, with inaccuracies that significantly lower confidence.

2.2 Hierarchical and Nonhierarchical Approaches

When developing a classification scheme, a hierarchical or nonhierarchical approach can be used. In a hierarchical approach, classes are nested such that the major classes are broken into subclasses and these subclasses can be further broken into more detailed classes. The advantage of such a system is that it can be easily generalized and adapted to various scales (the more local the area, the larger the scale, and the more detailed the hierarchy can be; see Chapter 5, Section 2.2.1 for a discussion of scale terminology).

Hierarchical systems are often used when conformity with a multiscale classification system is required or when the phenomena being mapped are hierarchical in nature. Hierarchical structure may be favored if there is a need to nest classes spatially or functionally. A disadvantage is that meaningless fill categories may be generated, since some general types break down more easily into progressive detail than others. The hierarchical approach can readily become ornate, having up to 6-10 hierarchical levels. This may be theoretically valid but impractical to use in the field.

A nonhierarchical approach, on the other hand, is usually designed for a specific purpose in a given area with a certain scale in mind. The advantage of nonhierarchical systems is that they are not constrained by hierarchical balance and rules. If there is no requirement to nest within an existing hierarchical system, it may be best to focus on selecting meaningful classes based on project needs, rather than forcing a fit to a hierarchical system. In some cases, however, it may be that the classes that are of interest to a project fit easily into a hierarchical system.

2.3 Some standard classifications

When we consider different classification schemes on a global expanse, it is useful to place them in the context of the vegetation that could exist in any given location if there were not some factors that restricted its development. This hypothetical concept is known as potential vegetation, that is, the vegetation that could potentially exist in idealized conditions. Potential vegetation is largely determined by climate and soils. A potential vegetation map is shown in Figure 6.01. (A digitized form of the well known Kuchler vegetation map, containing full detail, may be available sometime in the future).

In areas of high rainfall, high temperatures and adequate soils for vegetation growth, we find the potential vegetation is rainforest. In other cases where there is adequate soil for vegetation growth but low rainfall and low temperatures, we find the potential vegetation is tundra. In areas of low rainfall and high temperatures or very low precipitation and low temperatures, we find desert vegetation. The difference between potential and actual vegetation gives an indication of, among other things, the influence of disturbance, including human activities.

Keep in mind, however, that potential vegetation is hypothetical and, because it is largely determined by climate, it is not constant. A potential vegetation map of the world at the time of the last ice age would be quite different from one derived for current climatic conditions. Moreover, vegetation changes through time even in the absence of major climatic change. Following disturbance (for example, a fire) vegetation progresses through a series of seral (successional) stages, ultimately culminating in what has been referred to as a "climax" state. Potential vegetation maps use the assumption that for each location there is an actual climax stage which can be achieved in the absence of disturbance.

Most scientists who study vegetation dynamics, however, now consider vegetation to be in a state of continuous flux, marked by the sequences of individuals and species as gaps are created in the canopy, such as where trees fall, rivulets scour, or anthills denude. This concept is known as "dynamic equilibrium." In a forest, for example, a state similar to a climax community may be reached, but the individuals (and some species) will cycle continuously. Thus, the idea of potential vegetation must be recognized as a dynamic hypothetical concept.

One of the better known classification schemes for existing (as opposed to potential) vegetation is known as the Anderson scheme (Table 1) which was developed for use with remote sensing data (both aircraft and satellite based). Note that Anderson's classification is hierarchical, so it can be used for many different applications by selecting the level of detail desired. Even so, Anderson's scheme is not suitable for some applications, and many of the classes are not separable over large areas using remotely sensed observations.

Alternative schemes that utilize remotely sensed data and meet the criteria of global change research (e.g., vegetation functional types) have been developed. An example of one such classification, shown in Table 2, was derived using annual variations of spectral vegetation index observations, and attempts to order vegetation into functional type classes.

Table 1a: Anderson Land Use and Land Cover Classification Levels I and II

Level I

Level II

1 Urban or built-up land

11 Residential

12 Commercial and services

13 Industrial

14 Transportation, communications, and utilities

15 Industrial and commercial complexes

16 Mixed urban or built-up land

17 Other urban or built-up land

2 Agriculture land

21 Cropland and pasture

22 Orchards, groves, vineyards, nurseries, and ornamental horticultural areas

23 Confined feeding operations

24 Other agricultural land

3 Rangeland

31 Herbaceous rangeland

32 Shrub and brush rangeland

33 Mixed rangeland

4 Forest land

41 Deciduous forest land

42 Evergreen forest land

43 Mixed forest land

5 Water

51 Streams and canals

52 Lakes

53 Reservoirs

54 Bays and estuaries

6 Wetland

61 Forested wetland

62 Nonforested wetland

7 Barren land

71 Dry salt flats

72 Beaches

73 Sandy areas other than beaches

74 Bare exposed rock

75 Strip mines, quarries, and gravel pits

76 Transitional areas

77 Mixed barren land

8 Tundra

81 Shrub and brush tundra

82 Herbaceous tundra

83 Bare ground tundra

84 Wet tundra

85 Mixed tundra

9 Perennial snow or ice

91 Perennial snowfields

92 Glaciers

Table 1b: Example of Levels II and III for Urban or Built-Up Land

Level II

Level III

11 Residential

111 Single family units

112 Multifamily units

113 Group quarters

114 Residential hotels

115 Mobile home parks

116 Transient lodgings

117 Other

Table 2: International Geosphere-Biosphere Programme Classification

Evergreen Needleleaf Forest
Evergreen Broadleaf Forest
Deciduous Needleleaf Forest
Deciduous Broadleaf Forest
Mixed Forest
Closed Shrublands
Open Shrublands
Woody Savannas
Permanent Wetlands
Urban and Built-Up
Cropland and Natural Vegetation Mosaic
Snow and Ice
Barren or Sparsely Vegetated
Water Bodies

Many research groups using satellite observations to classify vegetation on a global basis are continuously improving their classification products and tailoring them to the needs of other researchers concerned with global vegetation modeling. This is particularly relevant to scientists interested in predicting the effects of vegetation change on future climate change.

One of the biases of current classification schemes is that they tend to label vegetation as one category or another based on their relative similarity (a problem shared by any attempt at categorization). New techniques are being developed that allow vegetation classification to be more ambiguous and continuous rather than employing discrete classes. For example, a researcher may be interested in vegetation more in terms of relative "woodiness" than in terms of Anderson's classes.

This is not just a hypothetical problem. It is known, for instance, that the transition between vegetation classes is not abrupt, yet nearly every vegetation map in existence draws a line between classes. Observers walking north through the boreal forest would not simply step out of a wall of trees onto a tundra plain. Instead they would see that the trees gradually became smaller and more sparsely distributed until finally there were no trees for miles in any direction. New vegetation classification schemes attempt to take these types of transition zones into account by depicting mapped vegetation boundaries as gradational rather than abrupt.