Research Projects
Regression and Variable Selection in Asynchronous Longitudinal Data
Motivated by Study of Women's Health Across the Nation dataset, we would like to analyze how hundreds of variables from FFQ and physical measurements influence people's cognitve function performance. However, the exposure variables, including nutrient intake, and the outcomes of cognitive tests, were measured asynchronously. We propose a sparse FPCA based calibration of asynchronous longitudinal data for regression and variable selection. Some theoretical work is under progress.
View Project
Improvements for Satellite Images Imputation Algorithm STFIT
We improve STFIT imputation algorithm in several aspects for the purpose of providing more accurate and complete satellite image data in water classification. It enables random forest to classify water pixels and outperform JRC on capturing temporal water area change. An improvement on computation efficiency for imputing global data is under progress.
View Project
Functional Change-point Detection for Multivariate Sparse Functional Data with Application on Urban Dynamics
Detecting an urbanization process and estimating when the urbanization was happening is critical in the study of urbanization and will facilitate the development of urban growth models and the investigation of further environmental impacts. We develop and implement a class of sparse functional change-point detection methods based on FPCA, CUSUM and ensemble to test and estimate the change-year in urbanization process by using Landsat dataset.
View Manuscript
A Geospatial Functional Model for OCO-2 Data with Applications on Imputation and Land Fraction Estimation
The NASA’s CO2 monitoring satellite, Orbiting Carbon Observatory-2 (OCO-2), aims to provide a comprehensive measurement network for the CO2 concentraion by retrieving from high-resolution spectra of reflected sunlight. However, due to large amount of missing radiance and unstable land fraction estimates, the spatial coverage of the retrieval algorithm is limited. We propose an approach to model spatially dependent hyperspectral data such that radiance imputation and land fraction estimation can be tackled well.
View Manuscript
Kaggle Competition: Don’t Overfit! II
We developed a Bayesian logistic regression model for the binary classification using 300 continous features, with only 250 training samples aimed for 19750 cases in test set. It ended up with ranking top 13%, 300/2330.
View Report
Advanced Spatial Statistics Final Project
We reviewed three papers regarding the topic "space-time covariance function on sphere", and successfully implemented the global space-time models for climate ensembles. The work was awarded Meritorious Research in Advanced Spatial Statistics class.
View Project