Multiple Point Geostatistics for automated landform mapping
Abstract
Land-surface processes are often studied at the level of elementary landform units, e.g. geomorphological units. To avoid expensive and difficult field surveys and to ensure a consistent mapping scheme, automated derivation of these units is desirable. However, automated classification based on two-point statistics of topographical attributes (e.g. semivarigram) is inadequate in reproducing complex, curvilinear landform patterns. Therefore, the spatial structure and configuration of terrain characteristics suitable for landform classification should be based on statistics from multiple points. In this study, a generic automated landform classification routine is developed which is based on Multiple Point Geostatistics (MPG) using information from a field map of geomorphology in a training area and a gridded Digital Elevation Model (DEM). Focus is on classification of geomorphologic units; e.g. alluvial fan, river terrace. The approach is evaluated using data from the French Alps. In the first procedural step, spatial statistics of the geomorphologic units are retrieved from a training data set, consisting of a digital elevation model and a geomorphologic map, created using field observations and 37.5 x 37.5 m2 cells. For each grid cell in the training data set, the geomorphological unit of the grid cell and a set of topographical attributes (i.e. a pattern) of the grid cell is stored in a frequency database. The set of topographical attributes stored is chosen such that it represents criteria used in field mapping. These are, for instance, topographical slope gradient, upstream area, or geomorphological units mapped in the neighborhood of the cell. Continuous information (e.g. slope) is converted to categorical data (slope class), which is required in the MPG approach. The second step is to use the knowledge stored in the frequency database for mapping. The algorithm reads a set of attribute classes from a classification target cell and its surrounding cells taking account of the multiple-point attribute patterns. Frequencies of these patterns are retrieved from the frequency database and converted to a categorial probability distribution giving the probability for each geomorphological unit class in the cell. The algorithm draws a possible unit class from this distribution and assigns the outcome to the target cell. This procedure is iteratively done for each cell in the mapping area. Results from the MPG-based classification routine are evaluated against the field map for two test areas:1) the training area itself to investigate the possible highest performance and 2) an independent area. It is shown that the MPG-based classification routine is capable of reproducing the landform units as present in the field map. The results show that including relative elevation to the local base level, slope gradient, and geomorphology of the downstream neighboring cell significantly contributes to the classification accuracy. For the test case with the training area, combination of these variables alone results in 60% classification accuracy. It can be concluded that these topographic attributes are key variables in the automated landform classification.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2011
- Bibcode:
- 2011AGUFM.U22B..06K
- Keywords:
-
- 1824 HYDROLOGY / Geomorphology: general;
- 1906 INFORMATICS / Computational models;
- algorithms;
- 1984 INFORMATICS / Statistical methods: Descriptive;
- 3275 MATHEMATICAL GEOPHYSICS / Uncertainty quantification