|
Spatial interpolation techniques are a powerful tool for generating visually continuous surfaces from scattered point
data, and the accuracy of interpolation determines the practical values of interpolating surfaces. As the variation of
surface elevation is nonlinear, the conventional spatial interpolation models implemented in many GIS packages
sometime cannot provide appreciate interpolation accuracy for certain application due to their nature of linear estimation.
In this paper, a method of generalized regression neural network residual kriging (GRNNRK) was presented for terrain
surface interpolation. The GRNNRK was a two-step algorithm. The first step included estimating the overall nonlinear
spatial structures by generalized regression neural network (GRNN), and the second step was the analysis of the
stationary residuals by ordinary kriging. And the final estimates were produced as a sum of GRNN estimates and
ordinary kriging estimates of residuals. To test performance of GRNNRK, a total of 1089 scattered elevation data got
from 28.86 km2 area were split into independent training data set (200) and validation data set (889), and the training
data set was modeled for terrain surface interpolation using ordinary kriging and GRNNRK, respectively, while the
validation data set was used to test their accuracies. The results showed that GRNNRK could achieve better accuracy
than kriging for interpolating surfaces. Therefore, GRNNRK was an efficient alternative to the conventional spatial
interpolation models usually used for scattered data interpolation in terrain surface interpolation.
|
