The roles of criteria, data and classification methods in designing land cover classification systems: evidence from existing land cover data sets

The Land Cover Classification System (LCCS) is a fundamental element and representative feature for any Land Cover Data Set (LCDS). Although various LCCSs have been proposed during the past few decades, discrepancies of LCCSs have widely existed in various LCDSs, which have caused negative impacts on comprehensive comparison and integrated utilization of multiple LCDSs. This study attempted to summarize the independent diagnostic criteria hidden in the existing LCCSs based on the induction method, and to synchronously discover the roles of data sources and classification methods in designing LCCSs. A total of 13 existing regional- or global-scale LCDSs were chosen. The analysis results show that phenology, coverage rate, vertical structure, and leaf type were the most frequently adopted criteria in the LCCSs of existing LCDSs. The decision of whether to adopt a diagnostic criterion in the LCCS of LCDS depended on the availability, effectiveness, and quality of the relevant data sources and classification methods. Currently, optical remote sensing images are still the prominent data source for regional- or global-scale LCDSs, and the potential of each diagnostic criterion could not be fully played. Multi-source and heterogeneous spatial data, ARD (Analysis Ready Data), and a fusion of optical, LiDAR (Light Detection And Ranging), radar, and other kinds of images have provided practical solutions. A lack of tools with high computing and storage capacities has been an alternative challenge. With the increasing advancement of new technologies, such as big earth data, crowdsourcing, deep learning, and cloud computing, more potential diagnostic criteria may be adopted for designing LCCS, and the richness and flexibility of the LCCS in the planned LCDS will gradually improve. This work not only offers beneficial references and revelations for the design of a new LCCS, but also provides insights for land cover mapping in large regions and the rational utilization of LCDSs.