Image Textrual Analysis for Soil Characterization

Seung-Cheol Shin, Ph.D Candidate (shinsc@engin.umich.edu)

and

Roman D. Hryciw , Professor (romanh@engin.umich.edu)

Department of Civil and Environmental Engineering

University of Michigan

Ann Arbor MI 48105

Abstract

With the advent of high-speed general purpose digital computers and the advances in electrical devices, computer vision has been adapted for processing of digital image data in many fields. Recently, the computer vision techniques have been applied to geotechnical engineering for laboratory and in-situ soil characterization.

In geotechnical engineering, one of the basic and most important properties of soil is particle size and its distribution (GSD). It controls the mechanical properties of soils such as shear strength and compressibility as well as a soil’s hydraulic conductivity. Computer vision techniques have led to alternate method for soil particle size analysis based on digital image analysis techniques. Many studies have shown that the size of non-contacting or contacting but not severely overlapping soil particles can be obtained by deterministic image analysis technique. However, the analysis of soil images with severely overlapped soil particles is more difficult task.

In this study, statistical textural analysis based on a spatial gray level dependence method (SGLDM) and a neural network have been used to determine the size of soil particles in gray scale soil images obtained from uniform sub-angular quarry sand. The results of the SGLDM consist of several textural features, which can be extracted from the spatial gray level dependence matrix. These textural indices can be considered as the representation of each digital soil image or a point in n-dimensional textural index space. In order to determine the soil particle size, the textural indices should be correlated to the soil particle size. Relationship between the statistical textural indices for gray scale image of uniform sand and the average particle size have been established using a supervised back-propagation neural network with considerable accuracy.

The information obtained from the statistical textural analysis was successfully implemented in a mapping technique using a supervised back-propagation neural network. Excellent agreement between actual particle size and predicted particle size by the neural network was achieved. The capabilities of the proposed neural network was tested and found to have considerable accuracy.