DS Ph.D. Dissertation Proposal | Himan Namdari | Monday, Nov. 25 @ 4pm, Fuller Labs, Beckett CF Room | Soil Characterization Assessment using gprMax and Machine Learning

DATA SCIENCE

Ph.D. Dissertation Proposal

Himan Namdari

Monday November 25 | 4:00PM-5:00PM

Location: Fuller Labs - Beckett Conference Room

Committee: 

• Advisor: Prof. Reza Zekavat, Departments of Data Science and Physics, Worcester Polytechnic Institute (WPI)
• Committee Member #1: Prof. Oren Mangoubi, Assistant Professor, Data Science and Computer Science, WPI
• Committee Member #2: Prof. Emmanuel Agu, Departments of Bioinformatics & Computational Biology, Electrical & Computer Engineering, Robotics Engineering, Data Science, and Biomedical Engineering, WPI
• Committee Member #3: Prof. Doug Petkie, Professor and Department Head, Electrical & Computer Engineering, WPI
• Committee Member #4: Prof. Radwin Askari, Associate Professor, Geological and Mining Engineering and Sciences, Michigan Tech University

Title:  Soil Characterization Assessment using gprMax and Machine Learning 

Abstract: 

Efficient soil root-zone texture, moisture, and composition (TCM) characterization are crucial for optimal mega-farm irrigation, water conservation in the era of climate change, and agricultural productivity. Drone-borne Ground Penetrating Radars (GPR) facilitate rapid scanning of mega-farms and mapping of the received signals into root-zone TCM via Machine Learning (ML) models. Supervised ML models require many labeled data for training that is very expensive to attain via direct farm measurements. This paper uses gprMax to develop a comprehensive synthetic GPR data emulator for realistic heterogeneous soil subsurface channels. The paper details the process of properly setting up the soil environment within the gprMax, which includes proper selection of antenna type and placement, environment size, pixel size, and the transmitted signal waveform. In addition, the paper develops an emulation engine that incorporates diverse heterogeneous soil Peplinski model parameters to generate soil channels realistically. The engine uses the Common Land Model (CLM), soil hydraulic models, soil density model, and the USDA soil texture triangle. Additionally, the paper introduces a novel technique that transforms GPR A-scan data into 2D binary, gray-scale, and heat-map image formats. The paper compares the performance of diverse ML models such as convolutional neural networks (CNNs) and Neural Networks (NN). The paper leverages meta-learning to improve accuracy compared to individual data representations, facilitating enhanced robustness and predictive accuracy.