The current rapid development trend in Artificial Intelligence (AI) provides a vast selection of high-quality tools to solve complex problems in more efficient ways than before. As a consequence, many fields of science and engineering are starting to explore AI tools, especially Deep Learning (DL) models for visual perception, audio understanding and decision making.

This thesis explores the application of DL in plant science and agriculture to overcome the throughput bottleneck inherent in the currently practiced manual phenotyping in fields. Among many field-based phenotyping challenges, we focus on the problems of pod-counting and flower detection in soybean crop production. To accomplish this objective, we leverage the RetinaNet DL model with different backbones to process the raw image data collected by autonomous ground-robotic platforms. The proposed high-throughput phenotyping framework also involves tracking algorithms for robust decision-making using multiple image frames from the video data collected by the robots. In the thesis, we discuss the training data generation, model building and inference processes in detail. High degree of accuracy results presented in this study demonstrate the promise of DL tools for field-based automated high-throughput plant phenotyping. Hence, a framework such as the one presented here can dramatically transform agriculture in terms of scalability, precision and profitability.