Deriving saltwater concentrations from the light intensity values of dyed saline solutions is a long-established image processing practice in laboratory scale investigations of saline intrusion. The current paper presents a novel methodology that employs the predictive ability of machine learning algorithms in order to determine saltwater concentration fields. The proposed approach consists of three distinct parts, image pre-processing, porous medium classification (glass bead structure recognition) and saltwater field generation (regression). It minimizes the need for aquifer-specific calibrations, significantly shortening the experimental procedure by up to 50% of the time required. A series of typical saline intrusion experiments were conducted in homogeneous and heterogeneous aquifers, consisting of glass beads of varying sizes, to recreate the necessary laboratory data. An innovative method of distinguishing and filtering out the common experimental error introduced by both backlighting and the optical irregularities of the glass bead medium was formulated. This enabled the acquisition of quality predictions by classical, easy-to-use machine learning techniques, such as feedforward Artificial Neural Networks, using a limited amount of training data, proving the applicability of the procedure. The new process was benchmarked against a traditional regression algorithm. A series of variables were utilized to quantify the variance between the results generated by the two procedures. No compromise was found to the quality of the derived concentration fields and it was established that the proposed image processing technique is robust when applied to homogeneous and heterogeneous domains alike, outperforming the classical approach in all test cases. Moreover, the method minimized the impact of experimental errors introduced by small movements of the camera and the presence air bubbles trapped in the porous medium.
|Number of pages||21|
|Publication status||Published - 26 Oct 2020|
FingerprintDive into the research topics of 'Optimizing Laboratory Investigations of Saline Intrusion by Incorporating Machine Learning Techniques'. Together they form a unique fingerprint.
Application of computer vision on laboratory sandbox investigations
Etsias, G. (Creator), Queen's University Belfast, Mar 2021
Machine learning assisted, laboratory and numerical study of saltwater intrusion in stratified and fractured coastal aquifersAuthor: Etsias, G., Dec 2022
Supervisor: Hamill, G. (Supervisor) & Flynn, R. (Supervisor)
Student thesis: Doctoral Thesis › Doctor of PhilosophyFile