Graduate Studies, UNL

 

Dissertations and Doctoral Documents from University of Nebraska-Lincoln, 2023–

First Advisor

Souparno Ghosh

Degree Name

Doctor of Philosophy (Ph.D.)

Committee Members

James Schnable, Mohammad Hasan, Reka Howard, Susan VanderPlas

Department

Statistics

Date of this Version

2025

Document Type

Dissertation

Citation

A dissertation presented to the faculty of the Graduate College of the University of Nebraska in partial fulfillment of requirements for the degree Doctor of Philosophy (Ph.D.)

Major: Statistics

Under the supervision of Professor

Lincoln, Nebraska, December 2025

Comments

Copyright 2025, the author. Used by permission

Abstract

Manual counting of maize tassels remains a time-consuming and labor-intensive process. Automating tassel counting using maize field images has received significant interest, with researchers exploring methods to streamline this task. Accurate tassel detection can substantially reduce both time and labor costs, offering a practical solution for large-scale agricultural management. Current studies in this area predominantly focus on training object detection algorithms to enhance prediction accuracy and efficiency. Images for this task are often captured in sequences, meaning they capture the exact location of the maize fields over time. However, existing tassel detection and counting methods typically ignore this sequential nature, assuming independence between images. In our work, we propose a Bayesian latent AR process model to accurately forecast tassel densities in these image sequences while quantifying the uncertainty in the predictions. Our method leverages established ideas from local count regression approaches for tassel density prediction, adapting these principles to develop a feature extractor within a transfer learning framework. We then train a latent AR process model with Bayesian inference techniques to effectively capture the associations within image sequences of maize fields. We next extend our work to a case when some of the images in the image sequence are missing. We introduce a two-stage framework, where some representations for the missing images are constructed first, and then a Bayesian modeling framework is invoked on the existing and constructed feature representations to forecast tassel densities in a sequence of images. We also compared the performance of the proposed modeling frameworks against some reliable baseline models that heavily rely on deep learning. The results demonstrate that our method not only outperforms the baselines but also provides reliable uncertainty quantification to support informed decision-making.

Advisor: Souparno Ghosh

Download

Share

COinS