Biological Systems Engineering, Department of

 

Document Type

Article

Date of this Version

2007

Citation

Transactions of the ASABE Vol. 50(1): 287−293

Comments

Copyright 2007 American Society of Agricultural and Biological Engineers

Abstract

Subsurface conditions can be non-intrusively mapped by observing and grouping patterns of similarity within ground-penetrating radar (GPR) profiles. We have observed that the intricate and often visually indiscernible textural variability found within a complex GPR image possesses important parameters that help delineate regions of similar subsurface characteristics. In this study, we therefore examined the feasibility of using textural features extracted from GPR data to automate subsurface characterization. The textural features were matched to a “fingerprint” database of previous subsurface classifications of GPR textural features and the corresponding physical probings of subsurface conditions. Four textural features (energy, contrast, entropy, and homogeneity) were selected as inputs into a neural-network classifier. This classifier was tested and verified using GPR data obtained from two distinctly different field sites. The first data set contained features that indicate the presence or lack of sandstone bedrock in the upper 2 m of a shallow soil profile of fine sandy loam and loam. The second data set contained columnar patterns that correspond to the presence or the lack of vertical preferential flow paths within a deep loessial soil. The classifier automatically grouped each data set into one of the two categories. Comparing the results of classification using extracted textural features to the results obtained by visual interpretation found 93.6% of the sections that lack sandstone bedrock correctly classified in the first set of data, and 90% of the sections that contain pronounced columnar patterns correctly classified in the second set of data. The classified profile sections were mapped using integrated GPR and GPS data to show ground surface boundaries of different subsurface conditions. These results indicate that textural features extracted from GPR data can be utilized as inputs in a neural network classifier to rapidly characterize and map the subsurface into categories associated with known conditions with acceptable levels of accuracy. This approach of GPR imagery classification is to be considered as an alternative method to traditional human interpretation only in the classification of voluminous data sets, wherein the extensive time requirement would make the traditional human interpretation impractical.

Share

COinS