An evaluation of the Land Cover Classification Product Sentinel 2 Prototype Land Cover 20 m Map of Africa 2016 for Namibia

Abstract

Global information on land cover is a primary output of remote sensing applications due to its importance to global change sciences, but also to governments and international initiatives. Consequently, a variety of land cover datasets have been developed in the past. Today users can choose among different products with various spatial resolutions for applications on global as well as on regional scales. A new classification covering the African continent was released by the European Space Agency only recently. Thanks to its mapping approach of 20 m, the Sentinel 2 Prototype Land Cover 20 m Map of Africa 2016 constitutes a novelty among the freely available products. However, being only a prototype, it is still missing final validation. This study aimed at evaluating the classification for the extent of Namibia by quantitatively and qualitatively comparing it to a selection of four low- to medium-resolution land cover products. Within the framework of an accuracy assessment for four test sites, statistical parameters were calculated which served as indicators for assessing the classification quality. According to the analysis, the overall accuracy of the prototype land cover product is on a medium level attaining approximately 54%. The per class accuracy varies from 2% to 100% suggesting that some classes require considerable reworking whereas others need less improvement. However, compared to the reference datasets, the prototype classification already constitutes a major development in land cover mapping.