MACHINE LEARNING MODEL SELECTION FOR PREDICTING GLOBAL BATHYMETRY

MACHINE LEARNING MODEL SELECTION FOR PREDICTING GLOBAL BATHYMETRY MS THESIS DEFENSE PRESENTATION Presented by: Nicholas Moran Advisor: Dr. Md Tamjidul Hoque Committee Members: Dr. Mahdi Abdelguerfi Dr. Shaikh Arifuzzaman Dr. Warren Wood

OVERVIEW � Introduction � Challenges � Motivation � Objective � Metrics � Results � Future Work � Conclusion

QUOTE “Despite the importance of earth’s ocean floor to our quality of life, we have made much better maps of the surfaces of other planets, moons, and asteroids. After five decades of surveying by ships carrying echo sounders, most of the ocean floor(∼ 90% at 1 minute resolution) remains unexplored. ” 1. JJ BECKER ET AL. GLOBAL BATHYMETRY AND ELEVATION DATA AT 30 ARC SECONDS RESOLUTION: SRTM 30 PLUS". IN: MARINE GEODESY 32. 4 (2009), PP. 355 -371.

INTRODUCTION � Aggregated Earth Gravitational Models � MBES and Echo Sounders � 90% of the world’s bathymetry is unknown � Bathymetry Interests � Defense � Commercial Shipping � Maps Figure 2: Single-Beam Echo Sounder (SBES) Illustration Figure 1: Multi-Beam Echo Sounder (MBES) Illustration

EARTH GRAVITATIONAL MODELS Figure 3: Illustration of correlation between sea surface altimetry and geoid height 1. BABULA JENA ET AL. PREDICTION OF BATHYMETRY FROM SATELLITE ALTIMETER BASED GRAVITY IN THE ARABIAN SEA: MAPPING OF TWO UNNAMED DEEP SEAMOUNTS". IN: INTERNATIONAL JOURNAL OF APPLIED EARTH OBSERVATION AND GEOINFORMATION 16 (2012), PP. 1 -4. 2. WALTER HF SMITH AND DAVID T SANDWELL. GLOBAL SEA OOR TOPOGRAPHY FROM SATELLITE ALTIMETRY AND SHIP DEPTH SOUNDINGS". IN: SCIENCE 277. 5334 (1997), PP. 19561962. 3. SIEMENS, C. W. , ON DETERMINING THE DEPTH OF THE SEA WITHOUT THE USE OF A SOUNDING LINE, PHILOS. TRANS. R. SOC. LONDON, 166, 671 -692, 1876. 4. DIXON, TIMOTHY H. , ET AL. "BATHYMETRIC PREDICTION FROM SEASAT ALTIMETER DATA. " JOURNAL OF GEOPHYSICAL RESEARCH: OCEANS 88. C 3 (1983): 1563 -1571.

SATELLITE-DERIVED BATHYMETRY Figure 4: Illustration of Satellite-Derived Bathymetry (SDB)

CHALLENGES � Global data coverage is limited � Ship Soundings are expensive � Predictions have inherently large error � Scope and Size of Problem Figure 5: Illustration of global echo sounder coverage for bathymetry

MOTIVATION � Dr. Warren Wood NRL Base funded project for predicting bathymetry � Use Ocean Features and Machine Learning to improve existing models � How can Machine Learning benefit global bathymetry prediction?

OBJECTIVE � Objective: Identify if there is a best fit model for predicting bathymetry, and is there a way to optimize predictions?

PLAN OF ACTION � Use Regression and Classification to compare effectiveness of models � Bin Bathymetry into class ranges of 150 meters � Select best features for prediction � K Fold Cross Validation � Find best performing model globally � Optimize by location

CLASSIFICATION MODELS Table 1: Classifiers used in thesis research Random Forest Classifier Ada Boost Classifier Gradient Boosting Classifier Quadratic Discriminant Analysis Decision Tree Classifier Voting Classifier Bagging Classifier Artificial Neural Network K Nearest Neighbors Classifier Naïve Bayes Classifier

FEATURE SELECTION • Genetic Algorithm for Feature Selection [1] • Selects best features in a relatively quick timeline 1. JIHOON YANG AND VASANT HONAVAR. FEATURE SUBSET SELECTION USING A GENETIC ALGORITHM". IN: FEATURE EXTRACTION, CONSTRUCTION AND SELECTION. SPRINGER, 1998, PP. 117 -136.

K FOLD CROSS VALIDATION � 10 Fold Cross Validation � Ideal for ensuring models do not over fit [1] 1. GAVIN C CAWLEY AND NICOLA LC TALBOT. ON OVER-TTING IN MODEL SELECTION AND SUBSEQUENT SELECTION BIAS IN PERFORMANCE EVALUATION". IN: JOURNAL OF MACHINE LEARNING RESEARCH 11. JUL (2010), PP. 2079 -2107.

SELECTED OCEAN FEATURES Table 2: Summary of ocean features aggregated for training Feature Dataset Mantle Density [1] CRUST 1 Crust Thickness [1] CRUST 1 Crust Density [1] CRUST 1 Current Vectors HYCOM Ocean Measurements [2] NASA Studies Sediment Thickness [1] CRUST 1 Bio. Mass Features WEI Geoid Features EGM Wave Height, Period WAVEWATCH 1. GABI LASKE ET AL. UPDATE ON CRUST 1. 0|A 1 -DEGREE GLOBAL MODEL OF EARTH'S CRUST". IN: GEOPHYS. RES. ABSTR. VOL. 15. EGU GENERAL ASSEMBLY VIENNA, AUSTRIA. 2013, P. 2658. 2. THOMAS MEISSNER, FRANK J WENTZ, AND DAVID M LE VINE. THE SALINITY RETRIEVAL ALGORITHMS FOR THE NASA AQUARIUS VERSION 5 AND SMAP VERSION 3 RELEASES". IN: REMOTE SENSING 10. 7 (2018), P. 1121.

CLASSIFICATION FLOW CHART

METRICS • • TP = True Positives FP = False Positives TN = True Negatives FN = False Negatives

CLASSIFICATION RESULTS

OPTIMIZED MODEL CACHE

LOCATION OPTIMIZATION METRICS

OPTIMIZED MODEL FLOW CHART

OPTIMIZED GRID RESULTS F 1 Score Balanced Accuracy 0. 83 0. 85

CONCLUSION � Using Machine Learning will require large amounts of accurate data � Selecting a model by a optimization function can improve prediction accuracy � A single model does not perform best in all instances

FUTURE WORK � Optimize by other factors � Different coverage shapes � Feature Optimization by Geospatial Location � Dynamic Resolution Predictions � Particle Swarm Optimization for Feature Selection � Extend for using physics models in optimization

QUESTIONS? � Thank you for attending!