Semi-automated mapping of landforms using multiple point geostatistics

• Published in: Geomorphology (Amsterdam, Netherlands) Netherlands), 2014-09, Vol.221, p.298-319
• Main Authors: Vannametee, E., Babel, L.V., Hendriks, M.R., Schuur, J., de Jong, S.M., Bierkens, M.F.P., Karssenberg, D.
• Format: Article
• Language: English
• Subjects: Accuracy; Areal geology; Areal geology. Maps; Automated landform mapping; Buëch; Discrete element method; Earth sciences; Earth, ocean, space; Exact sciences and technology; French Alps; Geologic maps, cartography; Geomorphology, landform evolution; Landforms; Landscape delineation; Mapping; Marine and continental quaternary; Multiple point geostatistics; Slopes; SNESIM; Surficial geology; Training; Trees
• ISSN: 0169-555X; 1872-695X
• DOI: 10.1016/j.geomorph.2014.05.032

This study presents an application of a multiple point geostatistics (MPS) to map landforms. MPS uses information at multiple cell locations including morphometric attributes at a target mapping cell, i.e. digital elevation model (DEM) derivatives, and non-morphometric attributes, i.e. landforms at the neighboring cells, to determine the landform. The technique requires a training data set, consisting of a field map of landforms and a DEM. Mapping landforms proceeds in two main steps. First, the number of cells per landform class, associated with a set of observed attributes discretized into classes (e.g. slope class), is retrieved from the training image and stored in a frequency tree, which is a hierarchical database. Second, the algorithm visits the non-mapped cells and assigns to these a realization of a landform class, based on the probability function of landforms conditioned to the observed attributes as retrieved from the frequency tree. The approach was tested using a data set for the Buëch catchment in the French Alps. We used four morphometric attributes extracted from a 37.5-m resolution DEM as well as two non-morphometric attributes observed in the neighborhood. The training data set was taken from multiple locations, covering 10% of the total area. The mapping was performed in a stochastic framework, in which 35 map realizations were generated and used to derive the probabilistic map of landforms. Based on this configuration, the technique yielded a map with 51.2% of correct cells, evaluated against the field map of landforms. The mapping accuracy is relatively high at high elevations, compared to the mid-slope and low-lying areas. Debris slope was mapped with the highest accuracy, while MPS shows a low capability in mapping hogback and glacis. The mapping accuracy is highest for training areas with a size of 7.5–10% of the total area. Reducing the size of the training images resulted in a decreased mapping quality, as the frequency database only represents local characteristics of landforms that are not representative for the remaining area. MPS outperforms a rule-based technique that only uses the morphometric attributes at the target mapping cell in the classification (i.e. one-point statistics technique), by 15% of cell accuracy. •Introduction of multiple point geostatistics (MPS) for automated landform mapping.•MPS uses DEM information and landforms in surroundings to map unvisited locations.•A geomorphological field map is used to train