Machine Learning in Geology A Pipeline for Automatic

Machine Learning in Geology A Pipeline for Automatic Classification of Shear-Sense Indicating Clasts Ari’El Encarnacion, Student Researcher Cinthya Rosales, Student Researcher Kevin Drake, Student Researcher Gurman Gill, Faculty Mentor Matty Mookerjee, Associated Faculty

Overview • Project Goal • Dataset • Machine Learning Overview • Project Challenges • Pipeline Definition • Pipeline components • Classification System Variants • Results • Future Work

The Goal • Use Machine Learning (ML) tools to automatically classify sigma clasts • Sinistral (Counter-Clockwise aka CCW) • Dextral (Clockwise aka CW) • Construct a pipeline that delivers classifications • Allow users to use pipeline via an app

The Data • CW Sigma Clast • CCW Sigma Clast

Machine Learning: What is it? • ML does two things • Extracts “features” from a dataset • Performs some analysis on those features • Computer vision “features” include edges, curves, and more complex shapes via Manning Publications, The Computer Vision Pipeline, Part 4: feature extraction

Challenges • Small starting data set • 70 CW • 32 CCW • Use of Style-GAN 2 -ada to generate new data • 6, 000+ clast images generated • Data always has noisy background • Multiple clasts increase classification difficulty • Possible to have CW & CCW in one image

The Pipeline

Preprocessing • Techniques applied to images prior to model training • Alters image in a way that isolates important features (clast edges, tails, etc. ) • Techniques used vary depending on model

Detection • Classification & Location via Ultralytics, yolov 5 Github repo • Using YOLOv 5 • Convolutional Neural Network based Object Detection Model via Lars Hulstaert, A Beginner's Guide to Object Detection

Detection & Classification • Option 1: • Detect presence of any clast • Isolate detections • Deliver all isolations to classifier • Option 2: • Use YOLOv 5 to detect CW/CCW • Provides detection and classification in one step

Detection & Classification: Option 1 • Classify isolated clasts using CNN + Transfer Learning (TL) • TL leverages an existing CNN to train a small custom CNN

Detection & Classification: Option 2 • Classify isolated clasts using only YOLOv 5 • Reduces complexity of pipeline • Results currently inferior to Option 1

Results: Option 1 • Untested using isolated clasts from detections • CNN + TL has been trained & tested on GAN generated data • Inception. V 3, a CNN by Google, used as base • Trained on 437 CW and 252 CCW clasts • Tested on 117 CW and 53 CCW clasts • CW 86% avg. accuracy • CCW 88% avg. accuracy

Results: Option 2 • Trained on 99 images • Bounding boxes of 151 CW and 65 CCW clasts • 7 training images contained no clast • 11 of 99 images used as validation set • 2 validation images contained no clast • Tested on 11 images • Bounding boxes of 6 CW and 4 CCW clasts as ground truth • 1 test image contained no clast • All experiments detected no clast in “non-clast” image • Best results • 4 out of 6 CW, avg. confidence of 57% • 1 out of 4 CCW, avg. confidence of 46%

Future Work • Use detection isolations as training data for Option 1 (CNN+TL ) • Use GAN generated data to train YOLOv 5 • May improve results greatly due to larger dataset • Assemble pipeline into single ecosystem • Host pipeline on i. OS app (currently in progress) or other public platform

Thank You!