Factors affecting spatial variation of classification uncertainty in an image object-based vegetation mapping.

摘要

Much effort has been spent on examining the spatial variation of classification accuracy and associated factors on a per-pixel basis. In the past few years, object-based classification has attracted growing interest. This paper examines the factors affecting the spatial variation of classification uncertainty in an object-based vegetation mapping based on airborne images at 1 m resolution in Point Reyes National Seaschore, California, USA. Six categories of factors were studied in an object-based classification: general membership, topography, sample object density, spatial composition, sample object reliability, and object features. First, classification uncertainty (classification accuracy on a per-case basis) was derived with a bootstrap method. Then, six categories of factors were quantified by categorical or continuous variables. In this step, the appropriate radius for calculating the spatial composition metrics of sample objects is also discussed. Finally, classification uncertainty was modelled as a function of those factors using a mixed linear model. The significant factors were identified and their parameters were estimated from restricted maximum likelihood fit. The modelling results showed that elevation, sample object size, sample object reliability, sample object density, and sample spatial composition significantly influence the object-based classification uncertainty. Many of these factors were closely related to the object-based approach. The result of this study helps in understanding classification errors and suggests further improvement of the classification.