Maria Brovelli, Daniele Oxoli and Gorica Bratic
During the last decades, production of Land Cover maps (LC) at a continental and global scale has increased thanks to the progress in Earth Observation capacities, as well as due to high demand for these maps for many applications (e.g. climate change monitoring). However, the usefulness of these maps strongly depends on their accuracy. Therefore, for a fruitful LC maps exploitation, accuracy assessments (i.e. validation and inter-comparison) must antedate. Spatial analysis of the errors should then complement the accuracy assessment to provide insights into the local errors patterns which may not be outlined by traditional accuracy assessment techniques.
According to this, we propose here a comprehensive analysis to target accuracy of LCs by focusing on the African continent. Two datasets, GlobeLand30 (GLC30) at and CCI Land Cover - S2 Prototype Land Cover 20m Map Of Africa (CCI Prototype Africa), at 30m and 20m resolution respectively, were considered.
Inter-comparison was performed by means of traditional accuracy indexes computed from the error matrix (i.e. Overall Accuracy, Producer’s and User’s accuracy etc). Harmonization of the two maps in terms of resolution and classification nomenclature is prerequisite for inter-comparison. This was achieved by taking advantage of QGIS functionalities (e.g. resampling). Results of the accuracy assessment provide overall quality metrics for the map as well as quality indicators for each LC class.
Additionally, spatial association statistics were adopted to investigate local patterns of the errors. The analysis was performed by means of GRASS and custom developed Python scripts exploiting cutting-edge data analysis libraries such as Pandas, Dask, and PySAL. By “virtual” overlaying CCI with GLC30, we computed which classes of CCI are under each pixel of GLC30, and what is the pixel fraction of a CCI class at each GLC30 pixel. Results of the overlay were stored in a vector point file whose coordinates represent centers of pixels of GLC30. If the classes of GLC30 and CCI for a point are not the same this results in an error with magnitude represented by the pixel fraction itself. Error fractions at each GLC30 pixel were analysed by means of Local Indicators of Spatial Association (LISA) to map non-random pattern in the spatial distribution of the errors as well as to assess their intensity and spatial association typology.
By considering the results of accuracy assessment and LISA outputs, a comprehensive comparison of the GLC30 or CCI Prototype Africa is achieved. The results provide a guideline for detecting source of the error, which is potentially useful for future LC production (i.e. sampling design of training data). Lastly, it has been demonstrated that processing of massive datasets for accuracy assessment can be accomplished with Free and Open Source Software (FOSS).