<p>High-throughput phenotyping using high spatial, spectral, and temporal resolution remote sensing (RS) data has become a critical part of the plant breeding chain focused on reducing the time and cost of the selection process for the “best” genotypes with respect to the trait(s) of interest. In this study, the potential of accurate and reliable sorghum biomass prediction using hyperspectral and LiDAR data acquired by sensors mounted on UAV platforms is investigated. Experiments comprised multiple varieties of grain and forage sorghum, including some photoperiod sensitive varieties, providing an opportunity to evaluate a wide range of genotypes and phenotypes. </p><p>Feature extraction is investigated, where various novel features, as well as traditional features, are extracted directly from the hyperspectral imagery and LiDAR point cloud data and input to classical machine learning (ML) regression based models. Predictive models are developed for multiple experiments conducted during the 2017, 2018, and 2019 growing seasons at the Agronomy Center for Research and Education (ACRE) at Purdue University. The impact of the regression method, data source, timing of RS and field-based biomass reference data acquisition, and number of samples on the prediction results are investigated. R2 values for end-of-season biomass ranged from 0.64 to 0.89 for different experiments when features from all the data sources were included. Using geometric based features derived from the LiDAR point cloud and the chemistry-based features extracted from hyperspectral data provided the most accurate predictions. The analysis of variance (ANOVA) of the accuracies of the predictive models showed that both the data source and regression method are important factors for a reliable prediction; however, the data source was more important with 69% significance, versus 28% significance for the regression method. The characteristics of the experiments, including the number of samples and the type of sorghum genotypes in the experiment also impacted prediction accuracy. </p><p>Including the genomic information and weather data in the “multi-year” predictive models is also investigated for prediction of the end of season biomass. Models based on one and two years of data are used to predict the biomass yield for the future years. The results show the high potential of the models for biomass and biomass rank predictions. While models developed using one year of data are able to predict biomass rank, using two years of data resulted in more accurate models, especially when RS data, which encode the environmental variation, are included. Also, the possibility of developing predictive models using the RS data collected until mid-season, rather than the full season, is investigated. The results show that using the RS data until 60 days after sowing (DAS) in the models can predict the rank of biomass with R2 values of around 0.65-0.70. This not only reduces the time required for phenotyping by avoiding the manual sampling process, but also decreases the time and the cost of the RS data collections and the associated challenges of time-consuming processing and analysis of large data sets, and particularly for hyperspectral imaging data.</p><p>In addition to extracting features from the hyperspectral and LiDAR data and developing classical ML based predictive models, supervised and unsupervised feature learning based on fully connected, convolutional, and recurrent neural networks is also investigated. For hyperspectral data, supervised feature extraction provides more accurate predictions, while the features extracted from LiDAR data in an unsupervised training yield more accurate prediction. </p><p>Predictive models based on Recurrent Neural Networks (RNNs) are designed and implemented to accommodate high dimensional, multi-modal, multi-temporal data. RS data and weather data are incorporated in the RNN models. Results from multiple experiments focused on high throughput phenotyping of sorghum for biomass predictions are provided and evaluated. Using proposed RNNs for training on one experiment and predicting biomass for other experiments with different types of sorghum varieties illustrates the potential of the network for biomass prediction, and the challenges relative to small sample sizes, including weather and sensitivity to the associated ground reference information.</p>

<p>Improved phenotyping technologies and data analytic strategies have the potential to reduce the phenotyping bottleneck in breeding programs, increase the number of genotypes that can be evaluated, and improve genetic gain of maize. Ear photometry and remote sensing were evaluated in this dissertation for their integration into breeding programs to understand the development of grain yield in diverse germplasm and to better predict yield performance. Ear photometry was used in Chapter 2 to characterize the testcross performance of temperate and tropical inbred lines. The effect of heterosis among the temperate heterotic groups was more noticeable in the ear-related characteristics rather than kernel-size characteristics. Yield components were generally more heritable than grain yield per ear, so they were explored for their use in multi-trait genomic prediction for grain yield on a plot or ear basis in Chapter 3. Multi-trait genomic prediction of grain yield was improved where ear characteristics were known in the testing set compared with single-trait genomic prediction.  Additionally, single-trait genomic prediction was more accurate in the temperate germplasm compared to the tropical germplasm. Thus, ear photometry is an efficient method to quickly assess yield components in maize and improve yield prediction in certain circumstances. In Chapter 4, the effect of row selection, plot size, and plot trimming on remote sensing trait repeatability and prediction accuracy of biomass yield in sorghum or grain yield in maize was evaluated. Decreased plot size and configuration has been suggested to increase the number of genotypes that can be evaluated per unit area. In this study, larger plot sizes were favorable for increasing repeatability and excluding outer rows improved predictive modelling. Plot trimming was never shown to be significantly different from non-trimmed plots in this study. Genomic prediction is another way to minimize experimental size and phenotypic data collection and was evaluated in Chapter 5. A reaction norm was used to model the trajectory of hybrid yield performance across a gradient of 86 environments. The heritability and prediction accuracy of grain yield were both improved in the higher-yielding environments compared to the lower-yielding environments. Single nucleotide polymorphisms with the highest magnitude of effects were selected in each environment. Twenty-one SNPs were selected indicating many SNPs were selected in multiple environments. Candidate genes in linkage disequilibrium with many of these SNPs were previously reported as stress adaptions. Genomic prediction and remote sensing were integrated for prediction of grain yield in Chapter 6. Heritability of remote sensing traits generally improved throughout the growing season. Prediction accuracy of BLUPs were improved through an integrated phenomic and genomic prediction model for all scenarios tested. In summary, ear photometry and remote sensing are technologies to evaluate large populations for unique plant trait characteristics that can be used in combination with genomic prediction to improve understanding of grain yield development and grain yield prediction.</p>

Peanut (Arachis hypogaea L.) is an important food crop in the USA and worldwide with high net returns but yield in excess of 4500 kg ha-1 is needed to offset the production costs. Because yield is limited by biotic and abiotic stresses, cultivars with stress tolerance are needed to optimize yield. The U.S. peanut mini-core germplasm collection is a valuable resource that breeders can use to improve stress tolerance in peanut. Phenotyping for plant height, leaf area, and leaf wilting have been used as proxies for the desired tolerance traits. However, proximal data collection, i.e. measurements are taken on individual plants or in the proximity, is slow. Remote data collection and machine learning techniques for analysis offer a high-throughput phenotyping (HTP) alternative to manual measurements that could help breeding for stress tolerance. The objectives of this study were to 1) develop HTP methods using aerial remote sensing; 2) evaluate the mini-core collection in SE Virginia; and 3) perform a detailed physiological analysis on a sub-set of 28 accessions from the mini-core collection under drought stress, i.e. the sub-set was selected based on contrasting differences under drought in three states, Virginia, Texas, and Oklahoma. To address these objectives, replicated experiments were performed in the field at the Tidewater Agricultural Research and Extension Center in Suffolk, VA, in 2017, 2018, and 2019, under rainfed, irrigated, and controlled conditions using rainout shelters to induce drought. Proximal data collection involved physiological, morphological, and yield measurements. Remote data collection was performed aerially and included collection of red-green-blue (RGB) images and canopy reflectance in the visible, near infra-red, and infra-red spectra. This information was used to estimate plant characteristics related to growth and drought tolerance. Under objective 1), we developed HTP for plant height with 85-95% accuracy, LAI with 85-88% accuracy, and wilting with 91-99% accuracy; this was done with significant reduction of time as compared to proximal data collection. Under objectives 2) and 3), we determined that shorter genotypes were more drought tolerant than taller genotypes; and identified CC650 less wilted and with increased carbon assimilation, electron transport, quantum efficiency, and yield than other accessions. / Doctor of Philosophy / Peanut is a profitable food crop in the USA but has high input costs. Pod yield over 4500 kg ha-1 is required for a profitable production, which is challenging in dry and hot years, and under disease pressure. Varieties tolerant to dry weather conditions (drought) and disease presence are required to sustain production. A collection of 112 peanut varieties is available for researchers to study the mechanisms of tolerance to drought and disease, and identify tolerant varieties to these stresses. Plant characteristics including height, leaf area, and leaf wilting can be used as proxies to estimate stress tolerance and yield, and identify tolerant varieties. How to measure these characteristics is very important. We think that using images collected by a drone and automated analysis by specific computer programs is the easiest, fastest, and most accurate way. Therefore, the objectives of my study were to 1) use drones and cameras to collect images, and computer programs to derive plant characteristics from these images, 2) evaluate the peanut collection to identify varieties with tolerance to drought and disease, and 3) evaluate in depth a sub-set of 28 varieties from this collection under controlled drought conditions to further learn about peanut mechanisms of tolerance to drought and diseases. Field experiments were conducted in 2017, 2018, and 2019, at the Tidewater Agricultural Research and Extension Center in Suffolk, VA. For some tests, we used rainout shelters to mimic drought. We measured plant height, leaf area, color, and wilting, canopy temperature, photosynthesis, and pod yield. From a drone, we collected images in the visible and invisible radiation and, using specific computer programs, estimated plant characteristics with 95% accuracy for height, 88% for leaf area, and 91% for leaf wilting under drought. We concluded that taller varieties were more susceptible to drought than shorter varieties. Peanut varieties CC650 and CC068 had higher end of season yield. The study showed that drought reduced several key mechanisms of photosynthesis including electron transport; and reduced the end of season yield. Variety CC650 performed better under drought than other varieties of the collection.

Quantitative genetics aims to bridge the genome to phenome gap. The advent of high-throughput genotyping technologies has accelerated the progress of genome to phenome mapping, but a challenge remains in phenotyping. Various high-throughput phenotyping (HTP) platforms have been developed recently to obtain economically important phenotypes in an automated fashion with less human labor and reduced costs. However, the effective way of designing HTP has not been investigated thoroughly. In addition, high-dimensional HTP data bring up a big challenge for statistical analysis by increasing computational demands. A new strategy for modeling high-dimensional HTP data and elucidating the interrelationships among these phenotypes are needed. Previous studies used pedigree-based connectetdness statistics to study the design of phenotyping. The availability of genetic markers provides a new opportunity to evaluate connectedness based on genomic data, which can serve as a means to design HTP. This dissertation first discusses the utility of connectedness spanning in three studies. In the first study, I introduced genomic connectedness and compared it with traditional pedigree-based connectedness. The relationship between genomic connectedness and prediction accuracy based on cross-validation was investigated in the second study. The third study introduced a user-friendly connectedness R package, which provides a suite of functions to evaluate the extent of connectedness. In the last study, I proposed a new statistical approach to model high-dimensional HTP data by leveraging the combination of confirmatory factor analysis and Bayesian network. Collectively, the results from the first three studies suggested the potential usefulness of applying genomic connectedness to design HTP. The statistical approach I introduced in the last study provides a new avenue to model high-dimensional HTP data holistically to further help us understand the interrelationships among phenotypes derived from HTP. / Doctor of Philosophy / Quantitative genetics aims to bridge the genome to phenome gap. With the advent of genotyping technologies, the genomic information of individuals can be included in a quantitative genetic model. A new challenge is to obtain sufficient and accurate phenotypes in an automated fashion with less human labor and reduced costs. The high-throughput phenotyping (HTP) technologies have emerged recently, opening a new opportunity to address this challenge. However, there is a paucity of research in phenotyping design and modeling high-dimensional HTP data. The main themes of this dissertation are 1) genomic connectedness that could potentially be used as a means to design a phenotyping experiment and 2) a novel statistical approach that aims to handle high-dimensional HTP data. In the first three studies, I first compared genomic connectedness with pedigree-based connectedness. This was followed by investigating the relationship between genomic connectedness and prediction accuracy derived from cross-validation. Additionally, I developed a connectedness R package that implements a variety of connectedness measures. The fourth study investigated a novel statistical approach by leveraging the combination of dimension reduction and graphical models to understand the interrelationships among high-dimensional HTP data.

<p dir="ltr">Meeting the nutritional requirements of an increasing population in a changing climate is the foremost concern of agricultural research in recent years. A solution to some of the many questions posed by this existential threat is breeding crops that more efficiently produce food with respect to land and water use. A key aspect to this optimization is geometric aspects of plant physiology such as canopy architecture that, while based in the actual 3D structure of the organism, does not necessarily require such a representation to measure. Although deep learning is a powerful tool to answer phenotyping questions that do not require an explicit intermediate 3D representation, training a network traditionally requires a large number of hand-segmented ground truth images. To bypass the enormous time and expense of hand- labeling datasets, we utilized a procedural sorghum image pipeline from another student in our group that produces images similar enough to the ground truth images from the phenotyping facility that the network can be directly used on real data while training only on automatically generated data. The synthetic data was used to train a deep segmentation network to identify which pixels correspond to which leaves. The segmentations were then processed to find the number of leaves identified in each image to use for the leaf-counting task in high-throughput phenotyping. Overall, our method performs comparably with human annotation accuracy by correctly predicting within a 90% confidence interval of the true leaf count in 97% of images while being faster and cheaper. This helps to add another expensive- to-collect phenotypic trait to the list of those that can be automatically collected.</p>

Doctor of Philosophy / Genetics Interdepartmental Program / Jesse A. Poland / A central question in modern biology is to understand the genotype-to-phenotype (G2P) link, that is, how the genetics of an organism results in specific characteristics. However, prediction of phenotypes from genotypes is a difficult problem due to the complex nature of genomes, the environment, and their interactions. While the recent advancements in genome sequencing technologies have provided almost unlimited access to high-density genetic markers, large-scale rapid and accurate phenotyping of complex plant traits remains a major bottleneck. Here, we demonstrate field-based complex trait assessment approaches using a commercially available light-weight Unmanned Aerial Systems (UAS). By deploying novel data acquisition and processing pipelines, we quantified lodging, ground cover, and crop growth rate of 1745 advanced spring wheat lines at multiple time-points over the course of three field seasons at three field sites in South Asia. High correlations of digital measures to visual estimates and superior broad-sense heritability demonstrate these approaches are amenable for reproducible assessment of complex plant traits in large breeding nurseries. Using these validated high-throughput measurements, we applied genome-wide association and prediction models to assess the underlying genetic architecture and genetic control. Our results suggest a diffuse genetic architecture for lodging and ground cover in wheat, but heritable genetic variation for prediction and selection in breeding programs. The logistic regression-derived parameters of dynamic plant height exhibited strong physiological linkages with several developmental and agronomic traits, suggesting the potential targets of selection and the associated tradeoffs. Taken together, our highly reproducible approaches provide a proof-of-concept application of UAS-based phenomics that is scalable to tens-of-thousands of plots in breeding and genetic studies as will be needed to understand the G2P and increase the rate of gain for complex traits in crop breeding.

<p>The presented work quantifies the impacts of inundated land area (ILA) on streamflow and crop development in the Upper Midwest, which is experiencing a changing climate with observed increases in temperature and precipitation. Quantitative information is needed to understand how upland and downstream stakeholders are impacted by ILA; yet the temporal and spatial extent of ILA and the impact of water storage on flood propagation is poorly understood. Excess water in low gradient agricultural landscapes resulting in ILA can have opposing impacts. The ILA can negatively impact crop development causing financial loss from a reduction or total loss in yield while conversely, ILA can also benefit downstream stakeholders by preventing flood damage from the temporary surface storage that slows water movement into channels. This research evaluates the effects of ILA on streamflow and crop development by leveraging the utility of remotely sensed observations and models.</p><p> </p><p>The influence of ILA on streamflow is investigated in the Red River basin, a predominantly agricultural basin with a history of damaging flood events. An inundation depth-area (IDA) parameterization was developed to parameterize the ILA in a hydrologic model, the Variable Infiltration Capacity (VIC) model, using remotely sensed observations from the MODIS Near Real-Time Global Flood Mapping product and discharge data. The IDA parameterization was developed in a subcatchment of the Red River basin and compared with simulation scenarios that did and did not represent ILA. The model performance of simulated discharge and ILA were evaluated, where the IDA parameterization outperformed the control scenarios. In addition, the simulation results using the IDA parameterization were able to explain the dominant runoff generation mechanism during the winter-spring and summer-fall seasons. The IDA parameterization was extended to the Red River basin to analyze the effects of ILA on the timing and magnitude of peak flow events where observed discharge revealed an increasing trend and magnitude of summer peak flow events. The results also showed that the occurrence of peak flow events is shifting from unimodal to bimodal structure, where peak flow events are dominant in the spring and summer seasons. By simulating ILA in the VIC model, the shift in occurrence of peak flow events and magnitude are better represented compared to simulations not representing ILA.</p><p> </p><p>The impacts of ILA on crop development are investigated on soybean fields in west-central Indiana using proximal remote sensing from unmanned aerial systems (UASs). Models sensitive to ILA were developed from the in-situ and UAS data at the plot scale to estimate biomass and percent of expected yield between the R4-R6 stages at the field scale. Low estimates of biomass and percent of expected yield were associated with mapped observations of ILA. The estimated biomass and percent of expected yield were useful early indicators to identify soybean impacted by excess water at the field scale. The models were applied to satellite imagery to quantify the impacts of ILA on soybean development over larger areas and multiple years. The estimated biomass and percent of expected yield correlated well with the observed data, where low model estimates were also associated with mapped observations of ILA and periods of excessive rainfall. The results of the work link the impacts of ILA on streamflow and crop development, and why it is important to quantify both in a changing climate. By representing ILA in hydrologic models, we can improve simulated streamflow and ILA and represent dominant physical process that influence hydrologic responses and represent shift and seasonal occurrence of peak flow events. In the summer season, where there is an increased occurrence of peak flow events, it is important to understand the impacts of ILA on crop development. By quantifying the impacts of ILA on soybean development we can analyze the spatiotemporal impacts of excess water on soybean development and provide stakeholders with early assessments of expected yield which can help improvement management decisions.</p>

<p>High-throughput phenotyping is a recent multidisciplinary research field that investigates the accurate acquisition and analysis of multidimensional phenotypes on large and diverse populations of plants. High-throughput phenotyping is at the crossroad between plant biology and computer vision, and profits from advances in plant modeling, plant reconstruction, and plant structure understanding. So far, most of the data analysis is done on 2D images, yet plants are inherently 3D shapes, and measurements made in 2D can be biased. For example, leaf angles change when they are reprojected in 2D images. Although some research works investigate the 3D reconstruction of plants, high-throughput phenotyping is still limited in its ability to automatically measure a large population of plants in 3D. In fact, plants are difficult to 3D reconstruct because they look self-similar, feature highly irregular geometries, and self-occlusion. </p> <p><br></p> <p>In this dissertation, we investigate the research question \textit{whether we can design and validate high-throughput phenotyping algorithms that take advantage of the 3D nature of the plants to outperform existing algorithms based on 2D images?} We present four contributions that address this question. First, we show a voxel 3D reconstruction pipeline and measure phenotypic traits related to canopy architecture over a population of 351 sorghum plants. Second, we show a machine learning-based skeletonization and segmentation algorithm for sorghum plants, which automatically learns from a set of 100 manually annotated plants. Third, we estimate individual leaf angles over a population of 1,098 sorghum plants. Finally, we present a sparse 3D reconstruction algorithm that can triangulate thousands of points of interest from up to 15 views without correspondences, even in the presence of noise and occlusion. We show that our approach outperforms single-view methods by using multiple views for sorghum leaf counting.</p> <p><br></p> <p>Progress made towards improving high-throughput phenotyping has the potential to benefit society with a better adaptation of crops to climate change, which will limit food insecurity in the world.</p>

The use of nitrogen (N) fertilizers is both costly to farmers and contributes to environmental degradation. N applied to wheat accounts for 18% of N applied to farmland globally, making it a prime target for reducing and optimizing N application. Chapter I is a review on nitrogen use efficiency (NUE) in wheat, with emphasis on breeding efforts and genetic resources available to increase NUE. The concept of effective use of nitrogen (EUN) as yield per unit N applied as a measure of N use, is also introduced. Chapter II is a study using two bi-parental double haploid families to evaluate genetic variability of both the genetic main effects (intercept) and linear response to N (slope) and determine the feasibility of selection for EUN in wheat. Using cross validation, a genomic prediction accuracy of 0.68 for intercept and 0.50 for slope was found, indicating that EUN is under genetic control and can be selected for. The prospect of breeding for EUN under limited resources, i.e., using fewer N rates and fewer experimental plots, is also explored. It was found that two different N treatments can be used to produce accurate predictions of intercept and slope as high as 0.98 and 0.95, respectively. Chapter III uses the same population described in chapter II to further investigate feasibility of selection for EUN using a normalized difference vegetation index (NDVI) obtained from multi-spectral aerial images gathered throughout the growing season. Cumulative photosynthesis across the growing season was estimated by integration across the NDVI curve, and compared to grain yield estimates to determine the efficacy of aerial imaging to identify high EUN lines. NDVI values and the area under the NDVI curve were able to predict yield and had the strongest ability to predict yield in moderate to low N treatments, with R2 values as high as 0.81 and 0.78 respectively. / Master of Science / Chapter I is a review on nitrogen use efficiency (NUE) in wheat, with emphasis on breeding efforts and genetic resources available to increase NUE. The concept of effective use of nitrogen (EUN) defined as grain yield per unit N applied, is contrasted to NUE as a more economic breeding goal. Chapter II uses two bi-parental mapping populations to evaluate genetic variability of both the genetic main effects and the linear response to N and determine the feasibility of selection for EUN in wheat. The efficacy of genomic prediction for EUN is explored and the prospect of breeding for EUN under limited resources is also explored. Chapter III uses the same populations described in chapter II to further investigate the feasibility of selection for EUN using a normalized difference vegetation index (NDVI) obtained from multi-spectral aerial images gathered throughout the growing season. Cumulative photosynthesis across the growing season was estimated and found to be predictive of grain yield estimates at accuracies ranging from 0.31 to 0.78.

Crop improvement is necessary for food security as the global population is expected to exceed 9 billion by 2050. Limitations in water resources and more frequent droughts and floods will make it increasingly difficult to manage agricultural resources and increase yields. Therefore, we must improve our ability to monitor agronomic research plots and use the information they provide to predict impacts of moisture stress on crop growth and yield. Towards this end, agronomists have used reductions in leaf expansion rates as a visible ‘plant-based’ indicator of moisture stress. Also, modeling researchers have developed crop models such as AquaCrop to enable quantification of the severity of moisture stress and its impacts on crop growth and yield. Finally, breeders are using Unmanned Aircraft Systems (UAS) in field-based High-Throughput Phenotyping (HTP) to quickly screen large numbers of small agronomic research plots for traits indicative of drought and flood tolerance. Here we investigate whether soybean canopy cover and color time series from high-resolution UAS ortho-images can be collected with enough spatial and temporal resolution to accurately quantify and differentiate agronomic research plots, pinpoint the timing of the onset of moisture stress, and constrain crop models such as AquaCrop to more accurately simulate the timing and severity of moisture stress as well as its impacts on crop growth and yield. We find that canopy cover time series derived from multilayer UAS image ortho-mosaics can reliably differentiate agronomic research plots and pinpoint the timing of reductions in soybean canopy expansion rates to within a couple of days. This information can be used to constrain the timing of the onset of moisture stress in AquaCrop resulting in a more realistic simulation of moisture stress and a lower likelihood of underestimating moisture stress and overestimating yield. These capabilities will help agronomists, crop modelers, and breeders more quickly develop varieties tolerant to moisture stress and achieve food security.