The establishment and implementation of procedures for vegetation classification has a long history. There has been a continuing evolution of widely accepted methods and cross-fertilization between the various disciplines concerned with spatial distributions of vegetation. In a trial-and-error manner, earlier work has been evaluated and if effective has been replicated in areas distinct from its original application. Classes and criteria have been replaced or simply added to improve the fit across a broader variability. As the number of different scientific disciplines grows, this trend towards a set of generic standards has been offset by the need for detail to accommodate special objectives.
With the advent of remote sensing and the computerized handling of maps, these long sequences of careful and scholarly compromise have been accelerated. There is now a more rapid cycling of mapping and checking and remapping, and the mechanisms available for comparing map products are enormously improved. Yet while the technical procedures are undergoing rapid development as both data and methods are refined, some efforts have neglected the knowledge base accumulated through centuries of anecdote and field experience. The digital "melting pot" would benefit from review and incorporation of useful contributions from scientific outputs that predate remote sensing.
Vegetation can be classified by any number of criteria, including all of the properties discussed in Chapter 4 (e.g., height, canopy density, seasonal LAI), and extending to a rich tradition of descriptive alternatives. However, when using remote sensing the patterns of vegetation must be classified according to spectral properties, whether from a single instantaneous image, a time series of images, or a selection from the bands available in the combined image sets. On the concept of data processing with complex arrays of spectral and temporal components (Chapter 5), it should be cautioned that a combination of all possible approaches to classification might quickly become unmanageable. Instead, the great majority of successful vegetation classifications with NOAA/AVHRR, Landsat and SPOT sensors have relied on relatively simple combinations of spectral reflectance properties, thermal contrasts, and multitemporal patterns. New information gained from multitemporal radar coverage will augment the lessons learned over the last 25 years with the earlier generation passive sensor systems.
3.1.1 Spectral reflectance properties -- Land cover classification using remote sensing is predicated on the assumption that different types of land cover have distinct reflectance properties. The unique spectral properties of a land cover class derive from its combinations of canopy geometry, leaf densities, colors, optical properties and moisture content, shadow components, transpiration rates, and non-vegetated reflectances. These factors aggregate within each pixel, and the population of pixels within the class offers a set of mean values and variances about those means. Together the two class attributes of central tendency and spread are defined within those image bands collected for the analysis, and are known as the class "spectral signature." When all signatures are collected and compared, it is optimal if they are well separated according to their means, at distances which are large compared with their internal variances. Accepted sets of signatures are presented to statistical classifiers, to assign each pixel of the image to one of the classes according to some form of best-match algorithm.
It is important to note that these reflectance properties are not constant through time. In fact, they vary greatly with the position of the solar light source (time of day and time of year) and with the position of the sensor view angle relative to the imaged area. Variations in spectral properties with varying viewing angles is illustrated in Figure 6.02. The same process can be observed first hand by viewing a grassy area from different angles at different times of the day. This phenomena is known as bidirectional reflectance (see Chapter 5 Section 2.4), and it has been the focus of research by remote sensing scientists interested in using remotely sensed observations to infer vegetation properties and processes (discussed in Chapter 4). Whereas bidirectional reflectance may often be considered a problem, it can be very useful for discriminating subtle differences between classes where surface texture or roughness is distinctive. There is an added benefit in that multiangle viewing, or pointability, allows more frequent observation of a given area by an orbiting satellite sensor.
Classification of vegetation using spectral reflectance properties in different spectral bands is a reasonably straightforward procedure, but a computer is essential for handling the statistical calculations carried out on each and every pixel (picture element) of an image, in each of the spectral wavelength bands available. For example, an AVHRR image of the continental United States alone contains about 350,000 cells (each cell representative of a 64 km2 area), for each of the five spectral bands. Thus nearly 2 million spectral reflectance values would typically be processed in a classification procedure. Relative to these processing loads, the higher resolution imagery from Landsat, SPOT, and future systems require pixel/band volumes greater by five or six orders of magnitude.
The array of values at each pixel must be compared with the statistics generated from the training classes (in the case of supervised classification) or with multiple clusters in an iterative fashion (in the case of unsupervised classification). These computationally intensive calculations quickly result in many billions of computer operations and hours of computation time, depending on how many classes are desired and the type of classification algorithm used. In many programs past and future the computer processing cycle is daily for the entire globe. Regardless of the chosen classification approach this would require at least a lifetime using a hand calculator!
3.1.2 Vegetation Indices (e.g., NDVI) -- Because vegetation indices condense the data from two (or more) spectral bands into one level of information, they can reduce the amount of computation time required for image classification. However, this benefit comes at a cost of information loss (most usually in the distinctions associated with overall brightness), and the classification results may be degraded. Vegetation indices derived from band combinations might better be reserved for monitoring, while classifications in support of monitoring would gain from use of full multiband original imagery.
Exceptions to this rule have emerged in studies where dozens and even hundreds of dates of vegetation indices are clustered to classify broad areas (usually of continental scale) according to the seasonalities of their greenup/senescence sequences. Not only are reduced image volumes preferable, but the difference between soil brightnesses or snow effects are undesirable, so that the brightness normalization performed by vegetation indices is a plus.
With respect to standard cover classification, multitemporal datasets are often very worthwhile, improving the accuracy of classification to a significant degree. Figure 6.03 shows the classification accuracy involving seven land cover classes using a single image vs. a multitemporal sequence of images acquired approximately six months apart (winter and summer). The use of multitemporal images not only resulted in higher classification accuracy than the single image approach, but also obtained consistent accuracy in all classes. In contrast, the single image classification had higher accuracy for some classes (e.g., 4-7) than others (e.g., 1-3).
Use of multitemporal data is especially advantageous in areas where vegetation or land use changes rapidly. For example, the two monsoon systems of East Africa offer four rising and falling limbs of moisture pulse, from which to discriminate the early and late producers for each rainy season. This offers many opportunities for more complete vegetation description than could be achieved with a single image. Similarly, a midsummer single date classification of productive vegetation in the eastern United States would be a poor choice for maximum discrimination, but should be considered an important member of a multitemporal set. Consider the differences between evergreen and deciduous trees. The former may appear quite uniform throughout the year (at least when not covered with snow), whereas the latter will vary widely between leaf-on and leaf-off periods. The discriminant power of multitemporal observations is based on their characterization of vegetation seasonal dynamics (phenology).
3.1.3 Incorporating Thermal Properties -- Thermal information can also contribute to the classification of land cover. Areas with sparse or no vegetation experience wider temperature fluctuations than more densely vegetated areas and generally are warmer than vegetated areas if plants are transpiring. These differences offer the potential to discriminate between contrasting types and densities of vegetation. There are many circumstances, however, where similar thermal responses are observed for quite distinct vegetation types. On the whole, it is not advisable to undertake a vegetation classification based exclusively on thermal imagery.
Combining visible and near-infrared imagery with thermal data series by time of day and by season results in excellent discrimination of cover classes. Figure 6.04 illustrates how NDVI and surface temperature vary with both time (July to September) and density of vegetated cover. For each date, datapoints toward the left of the plot are from areas at the site with sparse grassland vegetation, and datapoints toward the right represent areas of denser grassland cover. The datapoints from sparse grass cover show higher surface temperatures and lower NDVIs. The trend of the three different clumps (July to September) reflects overall cooling surface temperatures and some drying of the grassland cover.
Typically, satellite images are transformed into useful land cover classifications by thorough reference to site observations, background maps, and other supporting local knowledge (collectively termed ancillary data). Consider how much better the work could be designed by a savvy resident compared with someone with equal skills who has never been to the region of interest. This intimate knowledge need not be comprehensive throughout the region, but should identify areas where the classes of interest are juxtaposed and exhibit most of their local variations. These focus areas will aid the design and implementation of the classification by providing examples for the classifier and test areas in which to check the results.
This could be done in several possible ways. Two common methods are referred to as "supervised" and "unsupervised" techniques. Using supervised techniques, representative areas of known cover types are delineated and extracted from a satellite image (e.g., two types of forest, grassland, and four types of marshland) and the statistical properties of these "training areas" are then used to classify the entire satellite image. In an unsupervised approach, the image is classified using information from the satellite image alone, typically through a statistical technique called "clustering." The clustering process groups together pixels with similar spectral properties and then an analyst interprets and labels the resulting clusters.
Although the two techniques may sound similar, in actuality they are quite different. In supervised classification, classes to be delineated are specified exactly. Unsupervised classification algorithms statistically derive groupings based on information and variability contained in the data. It is up to an analyst to figure out the meanings of the groupings after they have been output. This is not to say the analyst has no control over the results. Minimum and maximum number and size of clusters may be stipulated beforehand.
Choosing between a supervised and unsupervised approach to classification of satellite image data usually depends on the availability of ancillary data. With good opportunities to define representative sets of training areas, the supervised approach might be the preferred choice, for reasons of control, replicability, and efficiency. However, if there is doubt about how the cover types will appear in the image data (especially with respect to their special-case appearances), then an unsupervised examination will show an overview of the spectral variability of the study region, and ultimately may allow a better set of categorical objectives to be achieved. If there is insufficient local knowledge to identify multiple examples of each class, then an unsupervised approach would generally be used.
This description of classifications is a simplified overview; within both supervised and unsupervised classifications, different methods are available. As shown in Figure 6.05, the simplest method of supervised classification, called "parallelepiped", uses decision rules in two (or more) dimensions by essentially "boxing in" the values that determine a given vegetation (or land cover) class. In contrast, a "maximum likelihood" classification technique uses decision rules based on the likelihood that a given value belongs to one particular class rather than another that may be similar. Thus maximum likelihood can be thought of as "circling" the values that determine a given vegetation (or land cover) class. Various methods may involve different computation times.
The relationship between vegetation index and surface temperature is used for other purposes besides land cover classification. Figure 6.06 shows that the slope of a linear relationship between NDVI and surface temperature (Ts) is related to the wetness of the surface, with a steep slope indicating drier conditions and potentially water-stressed vegetation, whereas a shallow slope indicates a relatively wetter surface and well- watered vegetation. Therefore, the slope of the relationship can be used as an indicator of surface soil moisture, a variable that has important implications for vegetation growth rates and crop suitabilities.
A land cover classification result is not complete until it has been evaluated by comparison to reference data. The goal is to understand how accurately and comprehensively the classified image represents actual conditions. It is easy to think a classified map represents "truth," yet in nearly every case, categorized patterns contain some error. Particular classes can be confused with one another, so that their combination may result in a better description. Another type of error occurs if a category is either underfavored by the classification procedure and underrepresented on the final map, or else overfavored and overrepresented. Accuracy assessments of classified images are important to identify the nature and extent of such errors.
Two sources of information are needed to perform an accuracy assessment: the classified image and some sort of reference data about the area that was classified. The reference information can be gathered by going physically to different sites and taking note of the land cover characteristics, or, when field visits are not practical, by extracting indicators of correct class from aerial photographs, maps, and higher spatial resolution satellite imagery. The best comparisons are from point sets scattered throughout the class types and over the geographic extent of the classified area, so that each class of vegetation in the classified map is well represented by reference data. The number of reference samples needed for good inference of the performance of each class is dictated by the sampling design.
Information from the classified image and the reference data can be compared in a table called an error matrix. Table 3 shows a simple error matrix. Reading across the first row of this table we can see that of 40 sample points that were selected from the classified image, 32 points agreed with the reference data (actual forest was classified correctly as forest); 7 were classified as forest but according to the reference data were grass; and 1 point classified as forest was actually bare soil according to the reference data.
|
|
|||||
|
|
|||||
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|