TY - THES

T1 - Lithologies identification in boreholes by endoscope measures and image analysis in MATLAB

AU - Catalan, Ignacio

N1 - no embargo

PY - 2023

Y1 - 2023

N2 - This master's thesis aims to automate lithology identification in boreholes using image analysis and a predictive Machine Learning model based on borehole videos in a limestone quarry (Valdilecha, Madrid, Spain). Endoscope measurements are taken in pre-selected boreholes, and video recordings enable subsequent image analysis. Color characteristics and irregularities aid in identifying lithologies with low, medium and high clay content. Image analysis tools extract key features, such as RGB color statistics, color count, and texture, serving as input for the Machine Learning model. Seven lithologies that include limestone with different clay content levels are defined (the identification id is given in brackets): limestone with no clay content (20), limestone with medium and high clay content (21 and 22, respectively), brecciated limestone with low, medium and high clay content (11,12 and 13, respectively) and pure clay (33). From them, two combinations are created to optimize the model's performance and address class imbalance: -Combination 1: Limestone with no clay content (Class 0: 20), Limestone and brecciated limestone with medium and high clay content (Class 1: 11,12,21 and 22), and Brecciated limestone and pure clay (Class 2: 13 and 33). -Combination 2: Limestone with no and low clay content (Class 0: 20 and 21), Brecciated limestone with medium and high clay content (Class 1: 11,12 and 22), and Brecciated limestone with high clay and pure clay (Class 2: 13 and 33). The Machine Learning model successfully predicts Classes 0 and 1 and demonstrates considerable accuracy, F1 Score, precision and recall for Class 2. Key variables, including Red vector statistics, color count, homogeneity, and contrast, significantly influence the model's performance in distinguishing lithologies. It is recommended to reinforce the performance of the model by extracting additional real data for minority classes, in particular class 2, in order to address potential challenges in future datasets. This study's promising approach automates lithology identification in boreholes, saving timing and reducing human errors and subjectivity. Implementing image analysis and Machine Learning techniques can benefit geological surveys and blasting operations. The model's ability to predict classes with varying clay content underscores its effectiveness in classifying lithologies from new borehole data. Future work should focus on improving image quality and ensuring the model's robustness across diverse datasets to optimize predictive performance.

AB - This master's thesis aims to automate lithology identification in boreholes using image analysis and a predictive Machine Learning model based on borehole videos in a limestone quarry (Valdilecha, Madrid, Spain). Endoscope measurements are taken in pre-selected boreholes, and video recordings enable subsequent image analysis. Color characteristics and irregularities aid in identifying lithologies with low, medium and high clay content. Image analysis tools extract key features, such as RGB color statistics, color count, and texture, serving as input for the Machine Learning model. Seven lithologies that include limestone with different clay content levels are defined (the identification id is given in brackets): limestone with no clay content (20), limestone with medium and high clay content (21 and 22, respectively), brecciated limestone with low, medium and high clay content (11,12 and 13, respectively) and pure clay (33). From them, two combinations are created to optimize the model's performance and address class imbalance: -Combination 1: Limestone with no clay content (Class 0: 20), Limestone and brecciated limestone with medium and high clay content (Class 1: 11,12,21 and 22), and Brecciated limestone and pure clay (Class 2: 13 and 33). -Combination 2: Limestone with no and low clay content (Class 0: 20 and 21), Brecciated limestone with medium and high clay content (Class 1: 11,12 and 22), and Brecciated limestone with high clay and pure clay (Class 2: 13 and 33). The Machine Learning model successfully predicts Classes 0 and 1 and demonstrates considerable accuracy, F1 Score, precision and recall for Class 2. Key variables, including Red vector statistics, color count, homogeneity, and contrast, significantly influence the model's performance in distinguishing lithologies. It is recommended to reinforce the performance of the model by extracting additional real data for minority classes, in particular class 2, in order to address potential challenges in future datasets. This study's promising approach automates lithology identification in boreholes, saving timing and reducing human errors and subjectivity. Implementing image analysis and Machine Learning techniques can benefit geological surveys and blasting operations. The model's ability to predict classes with varying clay content underscores its effectiveness in classifying lithologies from new borehole data. Future work should focus on improving image quality and ensuring the model's robustness across diverse datasets to optimize predictive performance.

KW - Lithology identification

KW - Image analysis

KW - Machine learning

KW - Clay content

KW - Blasting boreholes

KW - Lithologieidentifikation

KW - Bildanalyse

KW - Maschinelles Lernen

KW - Tonanteil

KW - Sprengbohrungen

M3 - Master's Thesis

ER -