Geography Reference
In-Depth Information
GIS can play a central role in implementing high-resolution SOMs. While SOM training
functionality is generally not provided in GIS (with the exception of some functions available
in
IDRISI
software), the low-dimensional neuron lattice of a trained SOM can readily be
represented using GIS data models. There is also plenty of flexibility when it comes to using
multiple data models, like vectors and rasters. For example, the vector data model may be
used to represent the locations of vectors mapped on the SOM or the trajectories of features
over time. Component planes - each representing a single variable retrieved for all neurons -
could be represented using a polygon structure, but for very large SOMs it becomes more
efficient to represent component planes as rasters interpolated from neuron centroids. From
the interpolation of these component landscapes to the dissolving of boundaries during
cluster visualization, GIS provides a large toolset readily usable for spatialization. Another
advantageous aspect is the traditional integration of spatial and non-spatial attributes in
GIS databases.
8.4 High-resolution SOM for Climate Attributes
The high-resolution SOM demonstrated here is related to one presented by Skupin (2007),
in which demographic attributes for all 200 000
census block groups were spatialized
in a SOM. Space-time paths captured in different cities and based on different modes of
transport were then projected onto that SOM by tracing the sequence of transected block
groups. One future direction of that effort is to combine demographic attributes with other
human and physical attributes towards a rich, attribute-based model of geographic space.
Ultimately, we would like to create a single SOM combining all of these attributes, which
requires integrating attributes originating in very different domains and stored in different
formats and attaching them to identical geographic features. The latter could be of uni-
form shape and size, like raster cells, or one could use varied features, like polygons of
different shape and size. To that end, the implementation described in this chapter demon-
strates how a set of physical attributes - specifically climate attributes - are summarized for
census block groups, which are then spatialized. This allows experimentation with a num-
ber of interesting aspects, including performing complex overlay procedures for transferring
attributes from raster grids to several hundred thousand polygon features.
Owing to the relatively smooth variation of climate attributes across space, using only
the climate data allows for more detailed observation of differences between the geographic
and attribute space visualization. With dominant spatial autocorrelation effects, neigh-
bouring regions in geographic space will tend to have similar climates and they should
thus remain in close proximity in the spatialization. Where that is not the case, one is
either dealing with pronounced geographic structures, characterized by rapid change of
attribute values across space, or with distortion caused by the dimensionality reduction
technique.
+
8.4.1 Climate source data and preprocessing
The data chosen for this study consisted of 11 climate attribute attached to point lo-
cations in the contiguous states of the United States (48 states plus the District of
Search WWH ::
Custom Search