TREE COVER AREA ESTIMATION IN EUROPE BASED ON THE COMBINATION OF IN SITU REFERENCE DATA AND THE COPERNICUS HIGH RESOLUTION LAYER ON TREE COVER DENSITY
Keywords: activity data, change detection, EEA, statistical inference, forest cover
Abstract. There is a natural tendency from the remote sensing community to extract area statistics (i.e. “Pixel counting”) from EO based geospatial products to produce statistical indicators for various purposes. However, geospatial map products suffer from misclassification errors and “pixel counting” can only be justified when the accuracy of such map products reaches a level when these misclassification errors can be considered negligible, but this is possible only in very specific circumstances. Nevertheless, there has been some effort in the Remote Sensing community to assess the accuracy of map products against some form of reference data to ensure that the maps could reach a sufficient level of accuracy. However, there is generally a lack of standards and guidelines and how to perform rigorous map accuracy assessment and rigorous methods for assessing map accuracies and extracting statistics are still lacking as highlighted by McRoberts (2011). Despite substantial advances in this topic in the scientific literature in recent years notably with the paper from Olofsson et al. (2014), this has yet to be fully implemented in operational projects. In addition, even if map accuracy assessment is performed correctly, high accuracy does not necessarily mean that area statistics can be directly extracted from a map.
This study is focused on developing the rigorous and appropriate use (i) of geospatial map products from satellite imagery and (ii) statistically sound methods for reporting area estimates and their associated uncertainty.