How local polynomial interpolation works

While global polynomial interpolation fits a polynomial to the entire surface, local polynomial interpolation fits many polynomials, each within specified overlapping neighborhoods. The search neighborhood can be defined by using the size and shape, number of neighbors,and sector configuration. Alternatively, the Exploratory Trend Surface Analysis slider can be used to simultaneously vary the Bandwidth, Spatial Condition Number (if enabled) and Searching Neighborhood values. Parameter optimization based on the cross validation statistics can also be done for LPI via the Optimize button Optimize.

A first-order global polynomial fits a single plane through the data; a second-order global polynomial fits a surface with a bend in it, allowing surfaces representing valleys; a third-order global polynomial allows for two bends; and so forth. However, when a surface has a varying shape, such as a landscape that slopes, levels out, and slopes again, a single global polynomial will not fit well. Multiple polynomial planes would be able to represent the surface more accurately (see diagram below).

Local Polynomial interpolation

Local polynomial interpolation, on the other hand, fits the specified order (zero, first, second, third, and so on) polynomial using points only within the defined neighborhood. The neighborhoods overlap, and the value used for each prediction is the value of the fitted polynomial at the center of the neighborhood.

In the image below, a cross section of sample elevation data is taken (a transect). In the first image, three neighbors (the red points) are used to fit a first-order polynomial and a line (the red line) to predict the unknown value for the location identified by the blue point. In the second image, a second location (the yellow point) is predicted by another first-order polynomial. It is very close to the first location, and the same measured points are used in the predictions, but the weights will be a little different, thus the polynomial fit (the blue line) is slightly different.

First-order polynomial
First-order polynomial

This process continues, centering on subsequent prediction locations, fitting local polynomials to predict the values. The two images below show two more arbitrary points being predicted to create the final surface. The orange point is predicted from the fitted polynomial (the green line) using the green measured sample points, and the brown point is predicted from the light purple polynomial.

Local polynomial
Local polynomial

In the two images below, two more polynomials are fit (the yellow and gray lines) to predict two more locations (the bluish-green and green points).

Local polynomial
Local polynomial

This process continues for each location. You can see how the surface is created (the purple surface line) for the sample points below.

Local polynomial

The optimal parameters are chosen to minimize the root mean square prediction error (RMSPE), similar to the selection of the power (p) value in IDW.

Measures of accuracy

Local polynomial interpolation provides two measures of accuracy that are not available for the other deterministic interpolation methods offered in ArcGIS Geostatistical Analyst.

Local polynomial interpolation relies on the following assumptions:

In practice, most datasets will not conform to these assumptions and in these cases the predicted values will be affected, but not as much as the prediction standard errors. To help you decide if the results in certain areas are reliable or not, LPI provides a spatial condition number surface. Rule of thumb values are shown in the table below and these critical values are rendered in yellow in the Condition Number surface.

Order of polynomial

Critical spatial condition number threshold value

1

10

2

100

3

1000

Greater than 3

Not recommended for most situations

Values that are below the critical spatial condition number thresholds indicate at which locations the solutions are reliable. Values close or equal to the critical values are questionable (should be examined with care), and values above the critical limits are not reliable.

The spatial condition numbers are generated by evaluating how sensitive the predicted value is to small changes in the coefficients of the linear prediction equations. A small spatial condition number indicates that the solution is stable, while a large value indicates that the solution is unstable. Instability of the solution should be a concern if it occurs in areas of particular interest because small variations in the input data (including their values, locations, and spatial arrangement) can cause large variations in the predicted value. This means that any uncertainty associated with the input data (for example, errors in attribute measurements or imprecisions the coordinates where the measurement was made) and especially the data outliers may have a considerable impact on the predicted values. Also, changes in the search neighborhood modify the number of data points (and weights in the case of a smooth search neighborhood) that are used to make a prediction and may affect the spatial condition number for that location.

The spatial condition number surface is created for the polynomials of order 1, 2, and 3. The prediction standard error is estimated assuming that the LPI model is correct (i.e. the local weighted least squares regression is an adequate algorithm and the spatial condition number values are smaller than the spatial condition number threshold values in the table above).

You can choose to exclude areas where high condition numbers occur in the prediction and prediction standard error maps by setting Use Spatial Condition Number Threshold to True on the LPI dialog box. The condition number is dependent only on the data location configuration. In other words, whether the ozone or the elevation values coming from the same dataset are used as input to LPI, the condition number surface will remain the same.

In the case of regularly distributed data, Constant, Epanechnikov and Quartic kernels are the best from a theoretical point of view for polynomials of order 0, 1, and 2 respectively. In the case of irregularly distributed data, selection of the best kernel should be based on the validation and cross-validation diagnostics and the spatial condition number values.

Kernel smoothing as found in Kernel Interpolation with Barriers is a variant of LPI. Local instabilities in these results are corrected by using a technique that is similar to ridge regression. The trade-off is that the predicted values are slightly biased and in most practical situations, the bias is not enough to affect decisions that you make based on the predicted values.

There are "holes" in my surface

The Bandwidth, Spatial Condition Number and the Searching Neighborhood values are typically changed when the Optimize button Optimize is used. If the Spatial Condition Number Threshold flag is set to False then only the Bandwidth and Neighborhood values are varied. If the Neighborhood type is set to Standard then all data up to a maximum of 1000 points are used and the minimum number of neighbors set to zero. Optimization is performed for Sector type equal to 1 Sector and for a smooth neighborhood. This optimization process can cause “holes” to appear in the output surface. These are areas where the Spatial Condition Number Threshold is exceeded or the searching neighborhood is too small. The search neighborhood parameters and the Spatial Condition Number Threshold can be adjusted to fill-in these areas, however, it must be noted that these “holes” were created where there may be instability in the calculation of the prediction values.

When to use local polynomial interpolation

Global polynomial interpolation is useful for creating smooth surfaces and identifying long-range trends in the dataset. However, in earth sciences, the variable of interest usually has short-range variation in addition to long-range trend. When the dataset exhibits short-range variation, local polynomial interpolation maps can capture the short-range variation.

Local polynomial interpolation is sensitive to the neighborhood distance and a small searching neighborhood may create empty areas in the prediction surface. For this reason, you can preview the surface before producing the output layer.


7/11/2012