Document Type

Thesis - Open Access

Award Date

2026

Degree Name

Master of Science (MS)

Department / School

Geography

First Advisor

Maitiniyazi Maimaitijiang

Abstract

Wheat grain yield is determined by the number of spikes per unit area, the number of grains per spike, and kernel weight. However, spike number and spikelets per spike are still measured predominantly through manual counting, a labor-intensive and subjective process that limits phenotyping throughput and selection intensity in modern wheat breeding programs. This thesis develops and evaluates a deep learning framework for automated, high-throughput phenotyping of wheat spike and spikelet traits across complementary imaging scales and implements the framework as a deployable platform. First, two large-scale benchmark datasets with oriented bounding box (OBB) annotations were established, comprising 48,521 spike instances from 903 close-range images and 60,404 spikelet instances from 4,009 spike images. YOLOv11 and YOLOv12 models were then evaluated for spike and spikelet detection and counting. Pretrained YOLOv11 achieved the highest spike detection performance (mAP@0.5 = 0.958) and counting accuracy (Pearson's r = 0.993), whereas the non-pretrained YOLOv11 model achieved the highest spikelet detection performance (mAP@0.5 = 0.990). Transfer learning improved spike detection but did not benefit spikelet detection, and detection accuracy did not consistently correspond to counting accuracy, demonstrating that both metrics should be evaluated jointly. Second, four YOLO-OBB architectures (YOLOv8, YOLOv11, YOLOv12, and YOLO26) were evaluated using seven UAV image modalities derived from MicaSense Altum-PT imagery acquired at four field sites in South Dakota during the 2025 growing season. Image modality had a greater influence on performance than model architecture. The six-band composite achieved the highest detection accuracy (mAP@0.5 up to 0.668), whereas panchromatic-enhanced imagery provided the best performance among three-band inputs and the most reliable counting results. Recent and heavier architectures did not consistently outperform YOLOv8 and YOLOv11. Finally, the trained models were integrated into WheatAI (wheatai.net), a cloud-native, browser-based platform that supports single-image and batch inference while generating outputs compatible with wheat breeding databases. Collectively, the benchmark datasets, comparative model evaluations, and deployable platform provide a reproducible and scalable framework for AI-enabled phenotyping of wheat yield components.

Publisher

South Dakota State University

Download

Share

COinS
 

Rights Statement

In Copyright