Automated Soil Classification And Identification Using Machine Vision And Artificial Neural Networks

Abstract

Introduction: In the construction industry, one of the major considerations when designing a superstructure is which foundation should be selected. Foundations provide support to superstructures by transferring the load of the structure evenly into the earth. An inappropriate foundation choice could result in damage to the superstructure, or even the collapse of such a structure. The clay content of soil is a major determining factor when selecting a foundation type. Soil containing clay, has the potential to shrink and swell as the water content changes. This heaving of the soil can cause damage to the superstructure built upon it. Determining the amount of clay in a soil sample, is one of the most important steps in the soil classification process. In South Africa, the Hydrometer method is commonly used to determine the clay content of soil samples. This method is a manual, time intensive soil classification method, with doubtful accuracy. This study was undertaken to develop an Automated Soil Classification System (ASCS) that will classify soil more accurately and more expeditiously, making it cost and time effective. This was achieved by applying a Machine Vision (MV) process to soil samples, to generate unique digital soil sample fingerprints for soil samples. This process was then combined with an Artificial Neural Network (ANN), to automatically classify the soil sample from the fingerprints. Methods: Initially a Machine Vision Instrument (MVI) was constructed for the consistent capturing of high fidelity images during the sedimentation process of a soil sample. Software was then developed to process these captured images and generate unique Soil Sample (SS) fingerprints for different soil constitutions. Four investigations were preformed to validate the consistency of the SS fingerprints generated with the MVI. These investigations were: 1. Validation of the SS fingerprint generation process; 2. Validation of the soil sample preparation procedure; 3. Determination of the differentiation ability of the MVI; and 4. Validation of the MVI by generating SS fingerprints for coded (unknown) soil samples. The generated SS fingerprints were then used to train an ANN to recognise and classify soil samples from their respective SS fingerprints. After the training of the ANN, a fifth investigation was undertaken determine the accuracy of the trained ANN and a final, sixth investigation was undertaken to compare the performance of the ASCS to that of the Hydrometer method. Results: The constructed MVI was able to acquire good quality greyscale images during the sedimentation process of soil samples in a consistent manner. Investigations 1 through 4 showed that correlation amongst SS fingerprints, generated from the same soil sample, was in the order of 97%, while the correlation amongst SS fingerprints, generated from multiple soils samples of the same constitution, was in the order of 95%. Investigation five showed that the training of the ANN was successful as the R values obtained after training were greater than 0,98. The sixth and final investigation showed that the accuracy of the ASCS was in the range of 95% and greatly outperformed the Hydrometer method, who’s accuracy varied from approximately 49 to 89%. The ASCS also delivered these results in 28 hours while the Hydrometer method took approximately seven days.