The performance of four methods for interpolating anisotropic, spatially no
nstationary fields is examined. The methods are optimal interpolation (OI,
also known as objective analysis), spline interpolation, multiquadric-bihar
monic method (MQ-B), and the inverse distance weighted method. The tests we
re performed using multiple realizations of random bivariate fields with kn
own underlying statistics, as well as highly anisotropic and nonhomogeneous
temperature and salinity fields across the Antarctic Circumpolar Current (
ACC).
The results of tests using homogeneous random fields show that all methods
except the inverse distance method have similar performance in the accuracy
. When the interpolated field is sampled adequately and data distributions
are dense, the presence of spatial deviations of the field statistics from
the field average will limit the interpolation skill of OI to be gained fro
m an increase in data density. In contrast, interpolation methods such as s
pline and MQ-B, which adjust the frequency response characteristics so that
the passband of the filter increases as the data spacing decreases, will a
ccount for such spatial variations and provide a more accurate interpolatio
n.
In the case of nonstationary and highly anisotropic processes, the most acc
urate interpolation analysis was obtained by spline interpolation and MQ-B.
As a result of the nonstationary fields encountered in the section crossin
g the ACC, the interpolation skill of the multiscale OI algorithm with an i
sotropic covariance function was lower. The highest relative interpolation
errors were obtained in the case of regular gaps resulting from intersperse
d deep and shallow stations, even though the total number of retained data
points is almost 80%. This is a consequence of inadequate sampling. All con
sidered methods do a poor job of extrapolating data in boundary regions. Fo
r the ACC mapping, extrapolation errors exceeded the standard deviation of
the fields by several times, indicating that the results of any interpolati
on method should be considered very critically in the boundary regions.