Document Type

Dissertation - Open Access

Award Date

2026

Degree Name

Doctor of Philosophy (PhD)

Department / School

Agronomy, Horticulture, and Plant Science

First Advisor

Sunish Seghal

Abstract

Accurate prediction of complex agronomic traits such as grain yield and yield components remains a central challenge in winter wheat breeding because these traits are controlled by many genes and are strongly influenced by environmental variation. The integration of high-throughput phenotyping (HTP), genomics, and advanced machine learning approaches offers new opportunities to improve predictive accuracy and accelerate genetic gain in plant breeding. This study evaluates the use of an unmanned aerial vehicle (UAV)-based multispectral phenomics combined with genomic information, and deep learning approaches to enhance the prediction of grain yield (GY), test weight (TW), grain protein content (GPC), and tiller density (TD) in a winter wheat breeding program. The research was conducted during the 2022 to 2024 growing seasons at three locations in South Dakota: Brookings, Dakota Lakes, and Winner, across multiple breeding nurseries, including the Elite, Advanced, and Preliminary yield trials. UAV-derived spectral indices collected across multiple developmental stages were first integrated into deep neural network (DNN)-based phenomic prediction models and multi-trait genomic selection (MT-GS) frameworks. Significant associations were observed between UAV-based spectral indices and key agronomic traits. Phenomic prediction using DNN achieved strong accuracy for single-location trials (R² = 0.71, 0.62, and 0.49 for GY, TW, and GPC, respectively), with further improvement when models were trained on multi-location datasets (R² = 0.76, 0.64, and 0.75). Prediction accuracy for GY was highest at the Feekes 11 stage. Forward prediction of preliminary breeding lines using models trained on multi-location advanced lines improved accuracy by 32% relative to single-location training. Incorporating UAV-derived spectral indices as covariates in MT-GS models further improved predictive ability for GY (0.40) compared to single-trait genomic selection models (0.23), demonstrating the value of integrating phenomic information into genomic prediction frameworks. To address challenges associated with model transferability across environments, a deep transfer learning (DTL) strategy based on a one-dimensional convolutional neural network (1D-CNN) was implemented. Baseline cross-year and cross-location predictions showed poor performance (R² as low as -15.3); however, partial fine-tuning with 20-40% of target data substantially improved accuracy, achieving R² values up to 0.83 in cross-year and 0.29-0.69 in cross-location scenarios. Growth stage-specific modeling further revealed predictive performance highest at Feekes 10.5 and 11 (R² = 0.78-0.82), highlighting the importance of developmental timing in UAVbased trait prediction. Beyond predicting primary agronomic traits, UAV-based phenomics also provides opportunities to estimate important yield components that are difficult to measure at scale in breeding programs. Finally, UAV multispectral imagery was used to estimate early-season tiller density (TD). Among evaluated models, an attention-based convolutional neural network achieved the highest predictive performance (R² = 0.82; RMSE% = 15.20), outperforming conventional machine learning and standard deep learning approaches. Collectively, these findings demonstrate that integrating UAVbased HTP, genomic information, and deep learning approaches substantially improves the accuracy, generalizability, and scalability of complex trait prediction in winter wheat breeding. By enabling earlier and more reliable estimation of key agronomic traits across environments, these data-driven phenomic and genomic prediction frameworks accelerate breeding decisions and support the development of high-yielding and climate-resilient wheat cultivars.

Publisher

South Dakota State University

Download

Share

COinS
 

Rights Statement

In Copyright