Street-level Spatial Interpolation Using Network-based IDW and Ordinary Kriging

Transactions in GISTransactions in GIS, August 2011, Volume 15, Issue 4

Narushige Shiode and Shino Shiode

“This study proposes network-based spatial interpolation methods to help predict unknown spatial values along networks more accurately. It expands on two of the commonly used spatial interpolation methods, IDW (inverse distance weighting) and OK (ordinary kriging), and applies them to analyze spatial data observed on a network. The study first provides the methodological framework, and it then examines the validity of the proposed methods by cross-validating elevations from two contrasting patterns of street network and comparing the MSEs (Mean Squared Errors) of the predicted values measured with the two proposed network-based methods and their conventional counterparts.

A cross-validation study using a dense, urban street network containing 148 sample locations (Data 2): (a) the elevation at each sample location shown in metres, (b) prediction accuracy of NT-IDW, (c) prediction accuracy of PL-OK, and (d) prediction accuracy of NT-OK. The prediction accuracy of each method is examined through cross-validation of the elevation at sample locations and the amount of MSE to the closest integer is shown next to each sample location. Darker circles represent larger MSE values (i.e. lower accuracy)

“The study suggests that both network-based IDW and network-based OK are generally more accurate than their existing counterparts, with network-based OK constantly outperforming the other methods. The network-based methods also turn out to be more sensitive to the edge effect, and their performance improves after edge correction. Furthermore, the MSEs of standard OK and network-based OK improve as more sample locations are used, whereas those of standard IDW and network-based IDW remain stable regardless of the number of sample locations. The two network-based methods use a similar set of sample locations, and their performance is inherently affected by the difference in their weight distribution among sample locations.”