Use of vector polygons for the accuracy assessment of pixel-based land cover maps

Identifying appropriate validation sources for large-area land cover products is a challenge, with logistical constraints frequently necessitating the use of preexisting data sources. Several issues exist when comparing polygon (vector-based) datasets to raster imagery: geolocational mismatches, differences in features or classes mapped, disparity between the scale of polygon delineation and the spatial resolution of the image, and temporal discrepancies. To evaluate the potential impact of using vector coverages to assess the accuracy of pixel-based land cover maps, five evaluation protocols are applied to test sites located in British Columbia and Newfoundland and Labrador, Canada. One protocol directly compared the land cover of the sample unit to the land cover of the forest inventory polygon within which the sample unit fell, two protocols used different regions around the sample unit to define the land cover class, and two protocols were based on homogeneity criteria that restricted the selection of sample units. For the protocols tested, the overall accuracy values ranged from 34% to 58%. Given the broad range of accuracies achieved, the results suggest that caution is needed when making spatially explicit comparisons between raster and vector datasets. When possible, the use of purpose-collected validation data is recommended for the accuracy assessment of maps derived from remotely sensed data; if preexisting vector-based data are the only option for the validation, approaches accounting for the heterogeneity of classes within a given polygon are recommended.