Global landcover classification validation issues and requirements

Global landcover mapping efforts using multitemporal AVHRR GAC and LAC data are current testbeds for developing continuous landcover characterization and monitoring approaches for use with planned EOS-era sensors, particularly MODIS (Moderate Resolution Imaging Spectroradiometer). While the remote sensing science community can address data handling and algorithm development issues, there are currently no adequate specifications or systems developed for training, validation and accuracy assessment requirements of global landcover. There are over 150 million land pixels for each global acquisition at 1-km resolution. A one-percent sample requires 1,500,000 observations for accuracy assessment. While sampling to support TM-based landcover mapping is well-developed, issues of training and validation at 1-km have not been resolved. Although there is utility in applying simple a priori landcover classification systems, there is a trade-off between simplicity of sampling and parameterization of landcover components so that reference data may be more useful to support different classification systems. It is obvious that traditional approaches to Earth-feature representation, field inventory and sampling for validation and accuracy assessment are inadequate for global mapping. New models of training and testing are required