Land Cover Classification

Uncertainty-Aware Visual Analytics for Analysis of Land Cover Classification

This joint project between the Chair of Cartography at TUM (TUM-LfK) and the National Geomatics Centre of China (NGCC) aims to empower land cover data analysis with a visual analytics interface, where patterns with highest geographic and categorical variation might be discovered through an expert input.

The study is applied to the remote sensing data of classified global land cover, as it is one of the crucial variables for the environmental analysis that support such directions of research as climate change, deforestation and urban and population growth. When it comes to the Land Cover classification, the process of deriving and applying the classes is complex. This complexity is due to the possible errors in automated classification algorithms, temporal and spatial heterogeneity of Earth Observation data, variation in availability and quality of ground truth data, variability in ground truth labels, or a combination of these. In the same time the rapid development of classification techniques and large amounts of collected data raise further research questions about the data integrity.

Much research has been carried out to visualise uncertain information along with the data in different spheres of scientific research. Current techniques of handling spatial-temporal uncertainty typically rely on treating data and its uncertainty as separate features to represent visually, through either intrinsic or extrinsic visualization, coincident/adjacent display or static/dynamic views. Therefore, there is a variety of approaches as glyphs, isolines, grid structures, map overlay using distinguishable colour coding or point and polygon highlighting. The visualisation perspective deals rather with visually signifying ambiguous data than reasoning under uncertainty. Thus, the integration of visual analytical tools within domain of land cover mapping could provide an additional dimension for the uncertainty communication and they can facilitate data exploration and analysis. This research aims to contribute to intrinsic/extrinsic visualization approach and propose to incorporate the visualization within a model-based design using Bayesian Networks to describe the data with inherent uncertainty. Our solution accommodates multiple Land Cover (LC) data sets and supports analysts in the task of heterogeneous data interpretation under uncertainty with user-defined quantitative input. The system empowers a human operator to assess multiple LC classifications and uncover uncertain information through interaction with the data. The interaction process includes assigning numerical probabilities based on expert beliefs for performing Bayesian inference.

The project is funded by International Graduate School of Science and Technology (IGSSE).

More information:

Contact person: Ekaterina Chuprikova, M.Sc. (