High-resolution subsurface soil characterization by image analysis and Vision CPT.

Abstract

With advances in sensor technology and optics, computer vision can be applied to in-situ soil testing devices. One of the most notable utilization of computer vision is in the Vision Cone Penetrometer Test (VisCPT). The VisCPT records images of soils as the probe advances through the ground. Subsequently, the collected images are processed using image analysis. The purpose of the image analysis is to extract useful information from the soil images to classify the soil types. An image analysis method for determination of average particle size from images of fairly uniform particle size soil masses was developed. The method utilizes two-dimensional wavelet transformation of images. A general relationship between textural index obtained by the image processing and the perceived soil particle size in pixels per diameter (<italic>PPD</italic>) is established. The VisCPT detects thin layers that were missed by the conventional CPT. To evaluate the significance of these CPT omissions, coefficients of hydraulic conductivity are assigned to inferred soil layers and equivalent horizontal and vertical hydraulic conductivities are computed. The ratio of equivalent horizontal hydraulic conductivity to equivalent vertical hydraulic conductivity is defined as the anisotropy ratio, <italic>R_a</italic>. The conventional CPT showed <italic>R_a</italic> ranging from 10 to 1000 while the VisCPT revealed that <italic>R_a</italic> was actually several orders of magnitude larger. The VisCPT was also used to assess the reliability of conventional CPT based site characterization in a stratified hydraulic fill. A model was also developed for CPT resistance at the interface between two different soils. The tip resistance is generally influenced by soils well ahead of the advancing probe while the sleeve responds only to the soil immediately surrounding it. As such, at interfaces there is a disjoint between the two readings often resulting in pronounced friction ratio spikes near interfaces. The model predicted that the CPT tip resistance lags sleeve friction on average 10 cm at such interfaces. A 10 cm downward shift in the tip resistances corrected the friction ratios and produced a. CPT-based classification in excellent agreement with VisCPT observations.