The aim of all Earth Observation is to gather information on the target surface and for land surfaces this generally means information on the status of vegetation and soils. There are two main approaches to this task : the "image-centred" approach and the "data-centred" approach (Schowengerdt, 1997). Broadly, the former uses the spatial relationships between features in the remotely-sensed image and requires high spatial resolution, while the latter uses the dimension of the data itself to infer surface properties. For sensors with a large IFOV (instantaneous field-of-view) and which scan at large angles relative to the surface, such as the AVHRR (NOAA) the ATSR-2 (ESA), POLDER (CNES), VEGETATION (CNES) and MODIS (NASA), the focus is largely on data-centred information extraction over large areas.
For monitoring semiarid grasslands (~20% of terrestrial surface) the objective is to provide reliable and consistent quantitative indicators of vegetation status over large areas and over long periods of time. We might also be interested in vegetation dynamics, that is, short-term changes in grassland condition, which can be related to various forms of disturbance (grazing, drought, fire, rodent infestation, flooding). Spectral vegetation indices calculated as ratios of visible and near-infrared reflectances are often used to provide an indication of vegetation cover and vigour, although none are sensitive to the entire range of values (i.e. both sparse and dense vegetation). It is therefore important to know what kind of grassland biome is being sensed. This can be seen as a hard classification task (classification of each sample into discrete cateogories) or as a soft classification task (estimating proportions of elements or end-members making up the reflectance value sample). Some success has been obtained with the soft classification approach via linear spectral unmixing, although this is less widely adopted with AVHRR since there are only two channels in the reflective domain. Hard land cover classifications have been achieved using data from the AVHRR with some success at various scales; for example, broad land cover classes such as cropland, forest, desert, urban areas and water bodies are easily discriminated and it has even been possible to distinguish subclasses within certain of these broad categories, such as different crop varieties, forest types and grassland biomes. See these examples.
However, problems arise when attempts are made to differentiate more subtle features, such as grasslands with different species composition, an application which may be termed community type differentiation (Kremer and Running, 1993; Trodd et al, 1997). This is particularly important for the monitoring application because one of the most consistent indicators of incipient degradation in semiarid grasslands is invasion of vulnerable areas by shrubs and annuals (see also), reducing the space available for more useful perennial species (which also help to maintain a stable topsoil). If a temporal dimension is added, changes in the locations of the community types as a result of secondary succession will indicate where degradation or restoration are occuring so that appropriate action can be taken. This is important in monitoring trends in desertification and its rehabilitation, especially for susceptible regions of the world such as northern China. See the CIESIN server for more information on the status and extent of land degradation and desertification worldwide.
Major objectives therefore include: