Multiscale quality assessment of Global Human Settlement Layer scenes against reference data using statistical learning

A method for quality assessment of the Global Human Settlement Layer scenes against reference data is presented. It relies on two settlement metrics; the local average and gradient functions that quantify the notions of settlement density and flexible settlement limits respectively. They are both utilized as generalization functions for increasing the level of abstraction of the sets under comparison. Generalization compensates for inaccuracies of the automatic target extraction method and can be computed at multiple scales. The comparison between the target built-up layers and the reference data employs an ordered multi-scale, linear regression computing the goodness of fit measure R²$R^2$. An optimized assessment procedure is investigated in a pilot study and is further employed in a big data exercise. A newly introduced quality metric returns the agreement between automatically extracted built-up from a set of 13605 scenes and the MODIS 500 urban layer, that was found too be as high as 91% for selected sensors. A final experiment attempts a performance increase at lower scales by correlating the target layer with automatically selected training subsets. At 50 m the adjusted R²$R^2$ increases by 3% with a mean squared error improvement of 2% compared to the performance achieved without statistical learning. The experiment suggests that the GHSL assessment at a global scale can be carried out based on limited high resolution reference data of minimal spatial coverage.