Remotely sensed images offer a new perspective on the land surface and environment, and a trained image interpreter can draw an enormous amount of information from even a qualitative examination of an enhanced image. However, the demand for statistical measures, and for comparisons with conventionally-derived maps normally leads to a process of classification of remotely-sensed data into a thematic map. It is widely recognised that a thematic map is in itself a considerable condensation of the information available from a survey (Goodchild, 1995). One of the major challenges in the use and adoption of remotely sensed data is to reconcile the type of observations made from satellite-based sensors with classifications that have been derived from direct ground-based eco-floristic classifications. Additionally, eco-floristic definitions and classifications often vary across national or regional boundaries.
The problem is probably more marked with coarser spatial resolution data. Ground-based classification schemes are often based on direct detailed observation and physical measurement of eco-floristic characteristics over quite small and widely distributed sample areas or transects, often just a few metres wide. Conventional vegetation cover maps are often produced by interpolation and generalisation of point or transect data into polygons, with abrupt and distinct boundaries shown between different cover types, i.e. forest to non-forest. Contrast this to a complete coverage of large but remote and indirect samples of light from 1.1km squares that form the basis of NOAA AVHRR data. Satellite-based data are continuous in space, but indirect measures of what is on the ground, averaged to the pixel size sample of measurement. Both datasets have their advantages and disadvantages, and the real benefit is to use the two systems synergistically.
NOAA AVHRR data are normally resampled to a grid with a nominal cell size of 1km (pixel). The classification of these data then involves an attempt to assign a category to this 1km pixel, according to some predetermined legend. Statistics can then be generated by summing the number of pixels assigned to this category and thematic maps can be produced. Some classification methodologies assign more than one value to a pixel, for example, a category, and a reliability, or proportion. The methods used for the categoric assignment are discussed in the next section. However, classification techniques applied to remotely-sensed data, especially to those of coarse spatial resolution such as NOAA AVHRR, have the inherent problem that pixels seldom contain exclusively one distinct ground cover class (e.g. Foody et. al., 1992). Furthermore, classification techniques are not well suited for the estimation of forest cover variables, which are more continuous than discrete in nature. Numerous studies have shown that the accuracy of classification of NOAA AVHRR data is directly related to the purity of the ground cover in the AVHRR size pixels (Cihlar et. al., 1996).
Accuracy assessments of digital land cover classifications are typically based on contingency tables, or confusion matrices, where accuracy is expressed in terms of errors of omission and commission, or in terms of agreement analysis using the Kappa test statistic (Stehman, 1996a; Congalton, 1991; Rosenfeld and Fitzpatrick-Lins, 1986). The contingency table is created by comparing on a class by class basis the land cover classification with an independent data source - field observations, existing maps, higher resolution imagery - collected using a statistically valid sampling strategy (Robinson et. al., 1983; Stehman, 1996b; Rosenfeld, 1986). Whilst such methods are well established they have generally only been applied to local scale classifications, occasionally to regional scale work and only in isolated instances on scales greater than these (Hay, 1988; Manshard, 1993; Estes and Mooneyhan, 1994; Jeanjean et. al., 1996). The general method of validating NOAA AVHRR-based classifications by employing classification of higher spatial resolution remotely-sensed data is now well established (Fitzpatrick-Lins, 1980; Rosenfield et. al., 1981) and has been employed extensively at local and regional scales (Borella et. al., 1982; Estes et. al., 1987). However, it is often based on the assumption that the results derived from the high spatial resolution imagery is completely accurate and reliable. This is obviously not the case – they themselves will include some imprecision and perhaps mis-classification, especially when seasonal patterns are prevalent but the purchase of multi-date high resolution imagery is prohibitive. In general though, it is generally felt that data derived from the higher spatial resolution imagery should be more reliable than those derived from coarser spatial resolution data, at least for the detailed spatial representation of the different land cover patterns. Furthermore, it is much more convenient to carry out field checking with higher spatial resolution imagery (say at scales of 1: 25,000) than with coarse spatial resolution data. Errors in the AVHRR-based classifications due to scale differences are often measured by inter-comparisons of the two co-registered sets of data (see section 6 below).
Because the satellite-based data are in a digital and “raster” form, they could in theory be reproduced at any scale. However, it is generally accepted that the nominally 1km-resolution data are generally suitable for reproduction of paper maps at one to one million scale. Most paper-based products of NOAA AVHRR data tend to be at smaller scales than this, typically 1: 5 million scale (see for example Stone et. al., 1993).
