Tuesday, 11 February 2003
Neural Networks for spatial interpolation of meteorological data
Juan P. Rigol, University of Edinburgh, Edinburgh, Scotland, United Kingdom
Meteorological data such as air temperature and rainfall are typically
recorded at the ground level at a limited number of weather stations
scattered over a region. Therefore, meteorological data values have to
be estimated at intermediate unsampled locations in order to generate
continuous surfaces for a whole area. Traditionally, two different
approaches have been taken to create continuous surfaces of
meteorological observations. The first approach builds on the fact that
the spatial distribution of meteorological data is known to be dependent
on relief. Thus, this approach aims at finding the functional
relationship existing between a meteorological variable and elevation
and other topographical covariates (e.g., slope orientation). The second
approach, that of interpolation, makes use of the potentially existing
spatial association between observations. In addition, several hybrid
methods have been devised in an attempt to incorporate valuable
information both from surrounding observations and terrain variables
into the estimation process. However, the modelling process may become
very cumbersome to apply when one tries to incorporate more than one
terrain variable into a single estimation process when using methods
such as thin plate splines, universal kriging or cokriging. More
commonly, the relationship between relief and meteorological variable
may be computed using linear regression, and then the residuals are
interpolated using a geostatistical method such as ordinary kriging
(e.g., residual kriging). These processes are combined when using
partial thin plate splines. Whatever the method, flexibility is required
when daily meteorological data are analyzed. The purpose of this paper
is to demonstrate a flexible and effective method of mapping
meteorological data which included both geographical and terrain
covariates and neighbor observations in the estimation process. Here,
feed-forward back-propagation neural networks are employed to estimate
daily meteorological data values at unknown locations using terrain
variables (e.g., easting, northing, elevation, aspect, roughness,
distance to sea) derived from a
digital elevation model (DEM) and observations from a long record made a
surrounding weather stations. Minimum air temperature and rainfall data
for
England and Wales are used to illustrate the methodology. Interpolated
surfaces are compared with those obtained using ordinary kriging and
thin plate splines. Results shown by the correlation between the
estimated and observed values at a spatially independent test data set
suggest that neural networks are an effective method for the spatial
interpolation of daily meteorological data. In particular, they have the
important advantage that only one model, incorporating both relief
information and spatial covariance structure, has to be specified for a
long daily meteorological record. The method also shows promise for
spatial interpolation since it neither requires the specification of the
function being modelled (e.g. linearity) nor makes underlying
statistical assumptions about the data (e.g. normality).
Supplementary URL: