Smooth function approximation for surface representation of soil sensor data

James Darrell McCauley, Purdue University

Abstract

Surface representations of soil properties greatly add to scientific understanding of spatial behavior as well as the ability of a farm manager to treat fields truly spatially rather than as areal units. Soil property measurements vary greatly at small scales, requiring smoothing in surface representations. Soil data collected through automated sensors are spatially scattered (i.e., non-gridded) and usually in the same spatial pattern that implements make while traversing a field. These two characteristics of data, noisy and traverse, define a unique approximation problem. Therefore, after substantiating these claims of soil property behavior with data from an organic matter sensor, approximation methods were evaluated. Smoothing splines and natural neighbor interpolation with gradient estimation were examined as solutions to this approximation problem. These two methods were viewed as alternatives to the usual covariance estimation and ordinary kriging. In addition to comparisons with ordinary kriging, inverse distance weighting (classical Shepard's method) was included in tests. In tests using an analytical test function, smoothing splines outperformed natural neighbor approximation, ordinary kriging, and inverse distance weighting for both noiseless and noisy data. In cross validation tests with sample data from an organic matter sensor, smoothing splines also outperformed all other methods. The spline approach was recommended for further mapping problems with soil sensor data.

Degree

Ph.D.

Advisors

Engel, Purdue University.

Subject Area

Agricultural engineering|Mathematics|Statistics

COinS