Data and rule sets
(1) Information extraction from legacy population maps
The procedure to extract population information based on point symbols from legacy maps was described in Kerle and de Leeuw (2009; see article list on main page). Here a subset of the scanned map can be downloaded, together with the ruleset developed in eCognition software (version 7).
- The processing follows the following main steps:
Removal of background. Much of the map is empty (with respect to population symbols). Using a quadtree segmentation (b) light, homogenous patches are found and assigned to background, eliminating most map pixels from further processing.
- Identification of large population symbols.
Candidate population dots were identified using the so-called Maximum Difference (Max.Diff) parameter (c-d), based on their strong local color contrast, as well as information from the green layer. Into the resulting candidate areas circular shapes were fitted using morphological operators (f). Conglutinated segments were dissolved using local extrema (g-h).
- Identification of small population symbols(i). This was done in a similar manner to the once described for large symbols.
- This was followed by some cleaning up (j), and finally by exporting of the results as GIS shapefiles.
- One final objective was to count the population per square tile (some 6300 for the whole map). Hence an additional chessboard segmentation was done.
Note that the original ruleset, developed of the two map sheets of Kenya included a procedure to remove the legend and scale bar that were printed in the same red color. Removal was done by specifying the rectangles to be removed using absolute coordinates (the map was earlier georeferenced).
(2) Characterisation of spectral, spatial and morphometric properties of landslides for semi-automatic detection
The procedure to detect landslides semi-automatically from high resolution satellite and elevation data was described in Martha, T. R. et. al. (2009; see article list on main page). Here a subset of the satellite data and DEM alongwith the DEM derivatives can be downloaded, together with the ruleset developed in eCognition software (versions 8).
The processing follows the following main steps:
A. Landslide recognition
- Identification of landslide candidates. An initial multi-resolution segmentation of the high resolution multispectral image was carried out to derive the objects. Later NDVI was used to separate landslide candidates from vegetated areas.
- Identification of false positives. Landslide false positives, such as roads, built-up areas, barren land and river sands were sequentially eliminated using object features. The leftover objects were recognised as landslides.
- This was followed by several cleaning up stages, whereby recognised landslides were again segmented using a chessboard segmentation technique, and the landslide boundaries and occasional smaller non-failure elements within landslide bodies were removed.
B. Landslide classification
- Segmentation of recognised landslides. The recognised landslide objects were further segmented using multiresolution segmentation, making also use of the terrain curvature layer. This helps in classification based on failure mechanism.
- Classification of landslide types. The objects were classified into five categories of landslides using morphometric and shape criteria.
Download the project (including the image, the .dpr project file, the ruleset)
Also instructions are included to adapt the ruleset to your own data