Land-cover classification and estimation of land-cover proportions at a global scale

Vous êtes ici

Context and objectives

A good knowledge of the spatial distribution of major biome types on a global scale is of critical importance for global environmental research. To accurately measure changes in global biome distribution, sensors on satellite platforms may be used. One of these sensors is the VEGETATION instrument, which was launched on the SPOT 4 platform in March ’98. The objective of this project is to conduct research in the prospect of the development of a future VEGETATION biome product. More in particular, research is conducted: (a) to identify the needs of present and future users of global biome information; (b) to develop a strategy for the characterisation and automatic classification of important biome types, using spatially aggregated sensor data; (c) to define and evaluate methods for the reduction of proportional bias in biome class area estimates, caused by aggregation of data from 1-km resolution to coarser resolutions. As no VEGETATION data for a full yearly vegetation cycle will be available until mid 1999, methodological developments in this study have been based on the use of NOAA-AVHRR data.

Project outcome

Expected scientific results

A supervised approach for global biome classification is proposed, based on five phenological attributes, derived from monthly NDVI growth curves, and the altitude above sea level, derived from a digital elevation model. The approach has been applied to spatially aggregated NDVI-data for the entire African continent, using the Simple Biosphere Model classification key. Depending on the level of spatial aggregation and the type of classifier the overall accuracy of the classification is between 78% and 87%. Accuracy is the highest for the NPB-classifier, which also has a much higher processing speed than the ML algorithm. For the reduction of proportional bias, a class- and context-dependent strategy is proposed, involving separate modelling of class proportion transition curves for groups of pixel blocks (or groups of individual pixels) with a similar class composition at the coarse resolution. Application of this strategy to 8-km and 16-km aggregated NDVI-data, using the block-based calibration approach, leads to an overall reduction of proportional bias by 26% (8km) and 33% (16km). With the sub-pixel correction method proportional bias is reduced by 18% to 44% for the 8-km scenario, and by 27% to 56% for the 16-km scenario, depending on the type of biome class that is considered. It remains to be investigated how well the proposed strategy for the removal of proportional bias will work when more biome classes are involved. Also, the whole procedure for classification and area correction that has been presented, and applied to NDVI data for Africa, needs to be translated into a standard procedure for the production of global biome information, adjusted to the use of VGT-PS data. These issues will be the subject of future research within the TELSAT programme.