Big data analytics in high-throughput phenotyping

Abstract

As the global population rises, advancements in plant diversity and crop yield is necessary for resource stability and nutritional security. In the next thirty years, the global population will pass 9 billion. Genetic advancements have become inexpensive and widely available to address this issue; however, phenotypic acquisition development has stagnated. Plant breeding programs have begun to support efforts in data mining, computer vision, and graphics to alleviate the gap from genetic advancements. This dissertation creates a bridge between computer vision research and phenotyping by designing and analyzing various deep neural networks for concrete applications while presenting new and novel approaches. The significant contributions are research advancements to the current state-of-the-art in mobile high-throughput phenotyping (HTP), which promotes more efficient plant science workflow tasks. Novel tools and utilities created for automatic code generation, maintenance, and source translation are featured. Promoted tools replace boiler-plate segments and redundant tasks. Finally, this research investigates various state-of-the-art deep neural network architectures to derive methods for object identification and enumeration. Seed kernel counting is a crucial task in the plant research workflow. This dissertation explains techniques and tools for generating data to scale training. New dataset creation methodologies are debuted and aim to replace the classical approach to labeling data. Although HTP is a general topic, this research focuses on various grains and plant-seed phenotypes. Applying deep neural networks to seed kernels for classification and object detection is a relatively new topic. This research uses a novel open-source dataset that supports future architectures for detecting kernels. State-of-the-art pre-trained regional convolutional neural networks (RCNN) perform poorly on seeds. The proposed counting architectures outperform the models above by focusing on learning a labeled integer count rather than anchor points for localization. Concurrently, pre-trained models on the seed dataset, a composition of geometrically primitive-like objects, boasts improvements to evaluation metrics in comparison to the Common Object in Context (COCO) dataset. A widely accepted problem in image processing is the segmentation of foreground objects from the background. This dissertation shows that state-of-the-art regional convolutional neural networks (RCNN) perform poorly in cases where foreground objects are similar to the background. Instead, transfer learning leverages salient features and boosts performance on noisy background datasets. The accumulation of new ideas and evidence of growth for mobile computer vision surmise a bright future for data-acquisition in various fields of HTP. The results obtained provide horizons and a solid foundation for future research to stabilize and continue the growth of phenotypic acquisition and crop yield.