Derivation of land cover continuous fields over Canada from SPOT-VGT imagery

Recent comparisons of coarse (1 km) and fine (30 m) resolution land cover maps across Canada indicate that single cover types rarely occupy more than 40% of a 1 km pixel in forested areas. To address this aggregation problem we develop and apply a method for estimating continuous fields of vegetation structural characteristics using 1 km resolution SPOT-VEGETATION (VGT) imagery. A sample of Landsat TM and ETM+ scenes stratified by ecozone is classified using a standard methodology to generate spatially distributed calibration centres. Neural networks, look-up-table labelling, and regression resulted in biases as large as 35% when we calibrated using centers over 400 km away. Only the linear least squares inversion approach produced a bias under 20%. A estimator based on a linear mixture model regularised by the a priori continuous field distribution over the calibration centres is developed. The regularisation parameter is defined by the spectral and spatial similarity of VGT reflectances between calibration centres and the regions being mapped. A strategy for mapping and validating Canada wide continuous fields using this method is described.