As seen from the reviews above, NOAA AVHRR data have been used operationally for providing global scale land-cover maps. However, the estimation of land-cover proportions from these maps will be associated with a systematic bias due to spatial aggregation effects, especially in areas where the amount of spatial fragmentation is high (Nelson 1989; Gervin et. al., 1985). For this reason, a number of researchers have indicated that NOAA AVHRR data should be used in conjunction with data from higher spatial resolution data for mapping and monitoring purposes (e.g. Nelson and Holben 1986; Van Roessel 1990; Jeanjean and Achard 1997)
We have already discussed the use of higher spatial resolution data for calibrating and validating various NOAA AVHRR sub-pixel classifications (see section 4.5). In this section we review the methods suggested for the use of image classifications from high spatial resolution data to improve the results of the classifications obtained at coarse spatial resolution. A few studies have applied a simple “regression estimator” approach. This involves relating the fine scale forest cover directly to the coarse scale forest cover (Nelson 1989; Stone and Schlesinger 1990; Amaral 1992; Stibig 1993). Forest/non-forest proportions are determined over blocks of AVHRR-size pixels, and related to the forest/non-forest proportions for the corresponding areas on the fine scale classifications. A simple linear regression is sought and then applied inversely to the whole of the AVHRR classification to correct for the estimation bias. However, the regression parameters depend heavily on the type of landscape analysed. When calculating the simple regression over large areas, ignoring the effect of spatial organisation leads to a large instability of the results (Nelson 1989).
In more recent work, Mayaux and Lambin (1995) showed that different regressions were obtained between Landsat TM-based and AVHRR-based forest percentages when the spatial pattern of the AVHRR-based forest/non-forest classification was taken into account. Mayaux and Lambin (1995) used AVHRR forest/non-forest classifications and Landsat TM classifications from 13 sites within the TREES project area. They aggregated forest proportion amounts in 13 x 13 km blocks, and used a simple spatial index (the Matheron Index as defined by Kleinn et. al., 1993) in a two-step correction function. The various stages of the two-step correction process were to:
- Geometrically co-register fine and coarse resolution data
- Overlay of a grid of contiguous pixel blocks. The block size selected is dependent on the accuracy of co-registration. Kleinn et. al. (1993) demonstrated in simulation studies that, assuming a misregistration of Landsat TM and AVHRR data of around two pixels, the influence of misregistration on the calculation of proportional errors for block sizes of 13 x 13 AVHRR pixels becomes negligible. If better registration can be achieved, smaller blocks may be used.
- Measure in each pixel block: (i) the forest cover at fine resolution; (ii) the forest cover at coarse resolution; (iii) the spatial pattern (e.g. Matheron Index) at coarse resolution
- Partition the population of pixel blocks in equal-sized subsets of similar spatial patterns
- Make First step regressions: Linear regressions within each subset between forest cover proportions at fine and coarse resolution. The observations are the pixel blocks.
- Make Second step regressions: Regressions between the spatial measure and (i) the intercept and (ii) the slope of the first-step regression. The observations are the subsets of the blocks.
Mayaux and Lambin (1995) were able to conclude that the integration of spatial information in a correction model significantly improves the retrieval of proportions of cover types from coarse resolution data compared with a simple correction function relating directly proportions at coarse and fine scale resolutions. This was true even is a simple spatial measure such as the Matheron Index was applied to the coarse resolution classification.
In a subsequent study, Mayaux and Lambin (1997) improved the two-step correction model by including textural measures derived from the coarse resolution data (e.g. minimum local variance for 3 x 3 AVHRR channel 2 pixels) instead of the Matheron Index. In a detailed study that tested a number of hypotheses they were able to show that:
- Area estimates of land-cover categories should not be extracted directly from coarse spatial resolution classifications because of the biases inherent in assigning a single class to a large pixel area;
- The correction of broad scale maps for the proportional errors introduced by spatial scaling drastically improves the reliability of the estimation of forest-cover areas;
- If only AVHRR data are available, the two-step procedures integrating around pixel texture with proportions estimated from a coarse resolution classification gave an equivalent performance to a mixed pixel regression-based estimator. The mixed pixel estimator, however, did not require the aggregation to larger block sizes (of say 9 x 9) but was dependent on the application of a suitable bulk correction for atmospheric effects;
- The two-step calibration procedure required a very large sample of high spatial resolution data since it depended on two levels of nested regressions;
- A more advanced concept of a land-cover map should consist of a multilayered database representing, for every layer, the calibrated proportion of a given land-cover class in a coarse resolution sampling unit. Or, pixel’s attributes should be vectors of within-pixel land-cover proportions rather than single land-cover categories.