Site updated: July 02, 2013
|[wordle generated from my 2009-2011 publications]|
April 12, 2013: I successfully defended my Ph.D. and accepted a position in the Machine Learning and Instrument Autonomy Group at the Jet Propulsion Laboratory, California Institute of Technology, starting June 2013.
I received my Ph.D. from the Electrical and Computer Engineering department, George R. Brown School of Engineering, Rice University in April 2013. My advisor was Erzsébet Merényi of the Department of Statistics.
From 2007 to 2010, I was a NASA Graduate Student Researchers Program fellow. The title of my GSRP research proposal was "Automatic Labeling of High Dimensional Remotely Sensed Imagery via Semantic Modeling." My NASA technical advisor was Kiri Wagstaff of the Jet Propulsion Laboratory.
Before coming to Rice, I was a member of the technical staff in the Machine Learning and Instrument Autonomy Group at the NASA Jet Propulsion Laboratory. Prior to JPL, I worked with Tomasz Stepinski, then of the Lunar and Planetary Institute, now at the Department of Geography, University of Cincinnati, on automated landform classification and crater detection through the analysis of Mars Orbiter Laser Altimeter (MOLA) data.
I was the primary developer (Nov. 2010-Oct. 2011) of the Hyperspectral Image Interpretation and Holistic Analysis Toolkit (Hii-HAT), originally implemented by David Thompson (PI) and Lukas Mandrake.
My CV is available here (last updated: April 15, 2013)
Much of the knowledge we have acquired about the processes that shape the Earth and our Solar System has been derived from imagery captured using such remote-sensing systems. The primary focus of my research involves developing machine learning techniques for the high-dimensional classification and domain adaptation problems that are central to the analysis of remotely-sensed imagery for Earth and planetary science application.
Remotely-sensed hyperspectral imagery has become one the most advanced tools for analyzing the processes that shape the Earth and other planets. Effective, rapid analysis of high-volume, high-dimensional remotely-sensed imagery demands efficient, automated techniques capable of identifying signatures of known material signatures captured in image data. We frame the material identification problem as a multiclass similarity-based classification problem, where our goal is to predict material labels for unlabeled target spectra based upon their similarities to source spectra with known material labels. Because the spectral representations of materials are affected by differences in capture conditions (e.g., differences in sensor type, atmospheric conditions, spatial location), we divide the material identification problem into intra-domain (i.e., source and target spectra captured under identical conditions) and inter-domain (i.e., source and target spectra captured under different conditions) settings.
The first component of our work in material identification develops adaptive similarity measures for intra-domain settings that measure the relevances of spectral features to the given classification task, using small amounts of labeled data. We proposed a technique based on multiclass Linear Discriminant Analysis (LDA) that combines several distinct similarity measures into a single hybrid similarity measure capturing the strengths of each of the individual measures. We also performed a comparative survey of techniques for low-rank Mahalanobis metric learning, and showed empirically that regularized LDA yields competitive results to the state-of-the-art, at substantially lower computational cost in a diverse set of hyperspectral image classification tasks.
The second component of my work in material identification considers inter-domain material identification settings. We proposed the multiclass domain adaptation framework for Relational Class Knowledge Transfer: RelTrans, that reconciles systematic differences between spectra captured under similar, but not identical, conditions. Our framework computes a similarity-based mapping that captures structured, relative relationships between classes shared between source and target domains, allowing us apply a classifier trained using labeled source spectra to classify target spectra. We demonstrated improved domain adaptation accuracy in comparison to recently-proposed multitask learning and manifold alignment techniques in several case studies involving state-of-the-art synthetic and real-world hyperspectral imagery.
Impact craters are among the most studied features on Mars. They are structures formed by collisions of meteoroids with the Martian surface. Their importance stems from the wealth of information that detailed analysis of their distributions and morphology can bring forth. or example, statistics of crater sizes forms the basis for geological stratigraphy of Mars. In addition, knowledge of crater morphologies enables studies of a number of outstanding issues in Martian geomorphology, such as: the nature of degradational processes, regional variations in geologic material, and distribution of subsurface volatiles. Thus, surveying Martian craters is an important task in planetary research.
We developed a novel approach for crater detection that utilizes digital topography data instead of imagery. We use a combination of segmentation and detection algorithms to delineate craters according to their morphological characteristics. We applied our method to a large and technically demanding test site and compare the results to the existing catalog of manually identified craters. Our algorithm finds many small craters not listed in the manual catalog, but it fails to detect heavily degraded craters. A detailed quality assessment of the algorithm is presented. The topography-based crater detection algorithm we developed offers a relatively simple and ready-to-use tool for identification and characterization of fresh impact craters with an adequate performance for an immediate application to Martian geomorphology.
As a precursor to our automated crater detection work, we proposed a numerical method for classification and characterization of landforms on Mars. Our method provided an alternative to manual geomorphic mapping of the Martian surface. We used digital elevation data to calculate several topographic attributes for each pixel in a landscape, and applied an unsupervised learning approach based upon the Kohonen Self-organizing Map, to divide all pixels into mutually exclusive and exhaustive landform classes on the basis of the similarity between their attribute vectors. This results in a thematic map of landforms for which the statistics of attributes are used to assign semantic meaning to the classes. We applied our method to produce a geomorphic map of the Terra Cimmeria region on Mars, and assess the quality of the approach in terms of established results for the Terra Cimmeria region.
My Google scholar page is here.
For each Fall semester from 2008 to 2010, I assisted Prof. Devika Subramanian with her course "Introduction to Computational Thinking."
While at Purdue, I was a teaching assistant for Systems Programming Laboratory (C), and Compilers: Principles and Practice, both with Dennis Brylow (now with the Dept. of Computer Science, Marquette University ).