Laboratory and in-situ soil characterization by computer vision

The gain size distribution is one of the fundamental features used for classification of soil. It provides a preliminary indication of a soil's engineering behavior. Recent development in computer vision and digital image technology has provided the means for development of soil classification methods based on digital images of the soil.; There are two approaches to treat the problem of soil classification from digital images of soil grains in an assembly form, a deterministic approach and a statistical approach. In the deterministic approach, every soil grain in the image is individually segmented and its size determined. A grain size analysis is the performed to develop an image-based grain size distribution curve. In this study, two deterministic methods for grain segmentation and size determination are presented. The first method uses the circular Hough transformation for grain center location followed by active contouring for grain segmentation. The second method utilizes watershed morphology analysis for grain segmentation and pixel counting for size determination. A shape correlation factor, λ_d, is introduced to explain the difference between the standard sieve-based and the alternative image-based grain size distributions. A mosaic imaging technique is used to effectively increase image viewing area while maintaining high pixel resolution. These two methods have been evaluated using laboratory prepared samples and compared with standard sieve-based grain size analysis methods. Excellent accuracy was achieved; The second approach for soil classification from digital images is a statistical texture-based method. In this treatment, correlations are developed between one or more parameters that quantify the grain size distribution, such as the average grain size, and a set of texture indices that measure the coarseness, homogeneity and contrast of the image texture. A supervised back-propagation neural network was utilized to establish this nonlinear correlation for images of uniform grain size soils. The neural network generalization was found to be accurate. This method can be utilized in analyzing soil images with a high degree of occlusion and where individual grain segmentation is not possible, such as in-situ images captured by the vision cone penetrometer (VisCPT). The accuracy of the textural analysis method in delineating highly stratified subsurface soil profiles was evaluated at two test sites. This method had a higher resolution capability for thinly layered soil compared to the standard in-situ classification methods using the standard piezo-cone penetrometer (uCPT).