The objectives of this study were to test the usefulness of various spectral channel combinations of AVHRR multitemporal composites for deriving land cover information in northern environments, and to assess the effect of AVHRR spatial resolution on the classification accuracy. A sequence of operations was carried out to remove radiometric distortions from AVHRR composites (1 km pixel sire) prepared for the landmass of Canada using multidate NOAA-11 data for the 1993 growing season: atmospheric corrections for AVHRR Channels 1, 2, and 4; identification and replacement of cloud-contaminated pixels; bidirectional reflectance corrections of Channels 1 and 2; and principal component (PC) calculations to retain significant Independent PC channels. Input principal components were classified using an unsupervised clustering algorithm, and accuracies were assessed through a comparison to 30 m Landsat TM pixels at five different sites in three biomes. We found that the normalized difference vegetation index (NDVI) was the most effective single spectral dimension to derive land cover types, but other channels (especially 1 and 2) were needed to obtain highest accuracies. Overall, classification accuracies for the 30 m pixels were between 45% and 60%. Mixes of land cover classes within AVHRR pixels were the principal reason for the low accuracies. When considering only AVHRR pixels with one dominant band cover type, the accuracy increased up to 80% or more in proportion to the mixed types retained. The accuracy also increased when a dispersed class (mixed forest) was combined with the more ubiquitous coniferous forest class. The intrinsic AVHRR resolution and the compositing process are the major factors influencing the impact of mixed cover types on the classification accuracy. The impact of these factors is discussed and strategies for optimizing the use of multitemporal AVHRR data in land cover classification are suggested.
