Boston University
Bernstein polynomial processes for continuous time change detection, and spatially orthogonal factor models for spatial transcriptomics and remote sensing data
Abstract
dc:description.abstractThis dissertation is broken into two main topics. The first is the development of a new class of heterogeneous continuous time Markov chains for continuous time change detection, and the second is the development of spatially orthogonal factor models for modeling data in the spatial transcriptomics and remote sensing sciences. There is a lack of methodological results for continuous time change detection due to challenges of noninformative prior specification and efficient posterior inference in this setting. Most methodologies to date assume data are collected according to uniformly spaced time intervals. This assumption incurs bias in the continuous time setting where, a priori, two consecutive observations measured closely in time are less likely to change than two consecutive observations that are far apart in time. Proposed models have required MCMC sampling which is computationally prohibitive in environmental settings. To address these issues, we derive a new subclass of heterogeneous continuous time Markov chains and prove their equivalence to noninformative change point processes. By construction, change points under this model can be inferred efficiently using the forward backward algorithm and do not require MCMC sampling. These models are then applied to study the effect of drought in semi-arid regions, agricultural crop rotations, and deforestation in the Amazon rainforest. Principal component analyses are fundamental tools for uncovering latent spatial patterns in spatial transcriptomics and environmental/climate sciences. However, there are a few drawbacks to current solutions. First, available approaches do not provide an explainable MAP estimator. To solve this, we derive a new sampling distribution of spatially orthogonal loadings within the principal component analysis framework and prove the MAP estimator under this model yields the eigendecomposition of $S + \frac{1}{n}\Sigma$, where $S = YY'/n$ is the sample covariance, $\Sigma$ is the prior spatial covariance, and $n$ is the number of times/genes. To address scalability, we develop a minorization-maximization-within-EM (MM-EM) algorithm yielding a computational complexity that is linear in the number of spatial locations. Finally, we provide a methodology for estimating the prior covariance $\Sigma$ under nonstationarity. For application, we estimate continental scale nonstationarity of phenological dynamics in sub-Saharan Africa. Further, we extend our method to the multiple CCA setting and estimate spatially orthogonal regions within 12 brain tissue slices that correspond to common modes of genetic expression across slices.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Mathematics & Statistics
- Grantor
- Boston University
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Cunha, Dan
- Advisors dc:contributor.advisor
-
- Carvalho, Luis E.
- Friedl, Mark A.
Subjects
dc:subject × 1Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/2144/53776