The Potential of Open Geodata and Multi-Temporal Landsat Data for Automated Large-Scale Land Use and Land Cover Classification

Abstract

With land cover ranked as an ‘essential climate variable’ and its high relevance as product de-rived through Earth Observation techniques, the importance of land cover, and related to this, also land use, is emphasized for various applications reaching from supporting decision-making in environmental planning to improving climate models as input parameter. In view of this, the study aims to examine the potential of existing open land use and land cover (LU/LC) data sets in terms of large-scale and high-resolution LU/LC classification. The proposed approach uses different data sets, i.e. CORINE, Natura 2000, Riparian Zones, Urban Atlas, OpenStreetMap, and LUCAS as input for the generation of training and reference samples, whereby as reference data LUCAS data points were preferred. Multi-temporal Landsat-7 and Landsat-8 scenes rang-ing from 12/2013-11/2014 served as basis for the derivation of temporal metrics and selected spectral indices, which were included in the supervised random forest classification with a reso-lution of 30 m applied to three coastal zones within Europe. An experimental setup was de-signed in order to accomplish the best performance. In this context, not only differing data sets were compared, but also combinations of them. Furthermore, the most promising training sam-pling strategy was identified, the sampling distribution modified, and supplementary input pa-rameters like topography and texture metrics added. In this gradual process, CORINE crystal-lized as the best performing data set for training samples generation, together with a modified sampling distribution and ancillary input features. CORINE achieves overall accuracies ranging from 60.3 % to 72.4 %, depending on the study area. The demonstrated automatable approach is promising and allows for an extension to larger scales, due to today’s advanced data processing resources and the time and cost saving methodology. The potential was also approved by a comparison with two similar LU/LC data sets, i.e. GlobeLand30 and High Resolution Layers. Similar accuracies were found for the High Resolution Layers, while GlobeLand30 showed even less accurate results, with by 12-24 % lower overall accuracies, compared to this thesis.