Enhancing photosynthesis for increased sorghum grain yield has become a key focus in sorghum breeding efforts. Phenotyping, involving the measurement of various morpho-physiological and physical traits associated with photosynthesis and grain yield, is a time-intensive process. However, the potential of non-invasive leaf-level hyperspectral imaging to swiftly detect plant performance, optimizing photosynthesis and grain yield, is promising. This study aimed to evaluate the feasibility of utilizing hyperspectral reflectance in the 350–950 nm range for the rapid estimation of these traits in intact sorghum leaves. Multiple machine learning regression algorithms were developed using leaf-level hyperspectral reflectance data from nearly 400 sorghum accessions within an association panel. The best-performing prediction models were then considered as potential methods for constructing a prediction model targeting multiple other physiological and yield traits in sorghum accessions. The results indicate that this approach enables the early detection of leaf photosynthetic and yield traits through leaf-level hyperspectral reflectance without the need for a full-range, high-cost leaf spectrometer.
Plot extraction for field trials plays a foundational role in supporting various research and development related to agricultural applications, including the classification of crop types, estimation of crop yields, and monitoring the health of crops. Traditional methods employed to define these plots have a trade-off between the substantial human labor involved and the precision of the delineated plots. In our research, we introduce a semi-automatic framework for plot extraction that requires only two essential inputs: the width and height of the plot. Our framework leverages the Segment Anything Model (SAM) for image segmentation, producing masks that are subsequently converted into polygons. These generated polygons are then filtered based on the user input. We refine the positions and orientations of these polygons by maximizing their overlap with the actual plot field. Experiment results were evaluated by comparing the extracted boundaries with manually digitized ground truth data. The results of our study demonstrate the successful extraction of individual plots across various fields characterized by diverse crop types. It is our expectation that this framework will significantly reduce the manual labor required for precise plot extraction, thus enhancing the ease and efficiency of this critical prerequisite task.
Source-Sink Regulated Senescence (SSRS) in maize is a complex trait involving sugar sensing and carbohydrate partitioning. While the yield effects of SSRS are still under investigation, mapping work in Midwest commercial germplasm has found evidence for QTL on several chromosomes. In this study, we seek to understand the variation of this trait in global maize germplasm and map the genetic loci linked to this senescence response using the Maize Nested Association Mapping population. Wide variation was seen in expression of the senescence phenotype including several lines that are potentially anti-senescent in response to sink removal. Phenotypes of NAM RIL populations support the possibility of anti-senescent or senescence suppressing genetic factors. Joint linkage mapping identified QTL for the SSRS trait on three chromosomes which are distinct from those previously reported. By understanding and mapping the global diversity of this trait, we can better understand the physiology of maize senescence and integrate insights into breeding programs.
Recent progress in proximal remote sensing has elevated both the spatial and temporal resolution of data acquisition, expanding the accessibility of these technologies for digital agriculture applications. These advanced sensors enable the gathering of extensive and novel datasets, proving instrumental in accurately characterizing phenotypes and parameterizing models for crop growth. Despite the distinctive structural, spatial, and spectral information embedded in these data streams, they have predominantly been utilized in isolation. Thus, this research aims to integrate these disparate data sources to improve estimations of agronomically important crop traits, such as yield. Deep learning methods, such as autoencoders, will be used to extract latent phenotypes, which will be used to characterize manually measured traits. We focus on multispectral images (MSIs) collected by unoccupied aerial vehicles and lidar scans collected by unoccupied ground vehicles. MSIs capture canopy-level spectral information, including the red, green, blue, red edge, and near infrared bands. Lidar scans are converted to point clouds to construct the three-dimensional sub-canopy architecture of maize plants. Data were collected on maize hybrids as part of the Genomes to Fields project, from 2018 to 2022, in Aurora, NY. Autoencoder model training on MSIs shows that latent phenotypes are effective image representations, containing relevant and sufficient information to generate image reconstructions. The latent codes are also predictive of the image date and normalized difference vegetation index values. Latent phenotypes were extracted from the lidar point clouds as well, and the prediction accuracies of models using these measurements separately and jointly will be compared.
Forest ecosystems are the largest terrestrial carbon sink and monitoring them effectively, particularly in the context of global change, requires rapid and accurate determination of tree functional traits that indicate forest health. Hyperspectral reflectance has the capacity to predict leaf traits non-destructively using multivariate statistical approaches. The ability of hyperspectral data to estimate tree physiochemical responses is affected by wavelength range selection and the influence of wavelength on accuracy of trait estimation is not well known, especially for more complex physiological processes. To bridge this knowledge gap, this study examined chemical and physiological responses of one-year-old black walnut (Juglans nigra L.) and northern red oak (Quercus rubra L.) to a combination of biotic and abiotic stress events, including pathogen inoculation, water stress, nutrient deficiency, and salt deposition. Leaf photosynthetic-related, water-related, and chemical traits were paired with hyperspectral measurements spanning 350−2500 nm. A total of 100 different wavelength ranges were evaluated using PLSR to determine the variation prediction accuracy. Key findings indicated that incorporating short infrared wavelength ranges (1300−2500 nm) significantly enhanced trait prediction accuracy. In addition, this study also demonstrated that hyperspectral data can detect tree stress responses at fine-scale chemical and physiological levels that are in agreement with reference measurement responses to the different stressors. We suggest that hyperspectral reflectance can potentially offer a solution for monitoring forest health in multi-stress environments with an increase in the accuracy of trait estimations and an expansion of the classes of traits estimated.
Compositional changes within dryland ecosystems in the Intermountain West and projected climatic shifts will negatively impact sagebrush-steppe communities. Restoration of native perennial bunchgrasses can sustainably increase ecosystem resistance to invasive plant species and resilience to environmental stress and disturbance. Bluebunch wheatgrass (Pseudoroegneria spicata ) is a native, perennial grass commonly used for large-scale restoration projects in the western U.S. Creating and utilizing spatially explicit models of phenotypic traits of bluebunch wheatgrass will establish consistent methods for evaluating seedling establishment and plant persistence. These predictive models will aid in the development of plant materials used for restoration, assist management practices that decrease the adverse impact of invasives, and increase ecosystem services of semiarid rangelands. High throughput phenotyping (HTP) uses an unmanned aerial vehicle (UAV) to capture multispectral imagery at incremental periods throughout the growing season. Images are analyzed for phenotypic traits associated with seedling establishment and plant persistence. In-field measurements for phenotypic traits are collected on a random sample of plots. Data mining and exploration will be done to assess relationships between imagery and field data, which will be used to construct models utilizing machine learning algorithms and be validated to determine error and bias. HTP will improve our understanding of seedling establishment and plant persistence traits and facilitate the development of germplasm used to produce plant materials needed for large-scale restoration projects on public and private lands. The re-establishment of native bunchgrasses on public lands could increase vegetative biodiversity within the ecosystem and create more suitable habitats for ruminants and wildlife.
Invasive species have historically disrupted environments by outcompeting, displacing, and extirpating native species, resulting in significant environmental and economic damage. Developing approaches to detect the presence of invasive species, favorable habitats for establishment, and predicting potential spread are crucial for effective management strategies to protect the environment and the economy. Spotted lanternfly (SLF, Lycorma delicatula ) is a phloem-feeding planthopper native to China that poses a severe threat to horticultural and forest production in the United States. Current pest management strategies are being explored to contain the spread and damage caused by SLF, however, methods to rapidly detect novel infestations or low-density populations are lacking. Using spectroscopy, we can detect changes in leaf canopies associated with stress events relatively quickly and potentially over large geographic areas. Here, we hypothesize that SLF infestations change the spectral characteristics of tree canopies. To test this hypothesis, we sampled silver maple (Acer saccharinum ), red maple (Acer rubrum ), black walnut (Juglans nigra), and tree of heaven (Ailanthus altissima ) at a common garden in Berks County, Pennsylvania with varying levels of SLF infestation enclosed on the trees. Composite spectral profiles separated based on SLF infestation level, but the magnitude of separation was different between species. We found multiple regions related with SLF infestation densities in each species, but not tree of heaven. By identifying changes in canopy spectral profiles in response to SLF infestation, we can possibly detect SLF infestations quickly and efficiently to help limit spread and better understand drivers of SLF movement across landscapes.
Water scarcity profoundly affects crop growth in rain-fed regions, including the Pacific Northwest (PNW) of the USA. While unmanned aerial vehicles (UAVs) are integral for crop monitoring in breeding programs, their use is resource-intensive and necessitates pilot presence in the field. Alternatively, Internet of Things (IoT)-based sensor systems offer continuous, remote, and real-time monitoring, but their data integrity requires validation for field applications. This study developed a Raspberry Pi-based sensor system (AGIcam+) and compared its efficacy with UAV in discerning crop responses to drought conditions across various wheat varieties in the PNW region. Multispectral and thermal data were collected across wheat trials (Winter 2023; Spring 2022, 2023) at crucial growth stages – preheading, heading, and post-heading – under varied drought stress conditions. Key vegetation indices and temperature measurements were extracted for a comparative drought performance analysis. Results indicate significant correlations between AGIcam+ and UAV data, more pronounced during the heading and post-heading stages. Pearson’s correlation coefficients for normalized difference vegetation index (NDVI) and 95th percentile temperature data ranged from 0.81-0.88 and 0.81-0.95 (P <0.01), respectively, across all trials during the heading stages. Yield prediction models using multivariate linear regression analysis from both systems’ data underscored AGIcam+’s accuracy in yield estimation, demonstrating performance comparable to UAV, as evidenced in the Spring 2023 trial (AGIcam+: R2 = 0.79, RMSE = 741.7 kg/hectare; UAV: R2 = 0.86, RMSE = 582.0 kg/hectare). These findings underscore AGIcam+ as a resource-efficient crop monitoring alternative, effectively capturing responses to environmental conditions and facilitating accurate yield predictions under drought stress.
Plant response to environmental stresses varies with time and is not uniformly manifested across the entire plant or specific organs. However, in most cases the phenotypic responses are measured at a single time point and lack spatial resolution. In this study, we aimed to develop and test a non-destructive approach to capture the dynamic plant stress responses over a time course and with spatial resolution. We used the rice panicle as the organ with known spatial heterogeneity and heat stress as the environmental perturbation. We used a series of 2D RGB images to reconstruction the rice panicle at high resolution. This analysis was applied to the rice diversity panel and enabled us to identify multiple loci regulating heat stress response by combining the 3D-reconstruction derived-approach digital traits with genome-wide association analysis. We further validated this approach with gene edits to confirm the role of the identified targets genes in heat stress response. In summary, our results present a high spatiotemporal resolution approach to identify digitals traits and underlying genetic variation that is unlikely to have emerged from conventional image-based phenotyping.
The Forage and Range Lab (FRRL) is implementing a high throughput phenotyping (HTP) project using unmanned aerial vehicles (UAV), multispectral sensors, and programmatic pipelines for computational automation. Here we report on different HTP experiences for forage and turf grasses. Phenotypic traits (plant height, biomass, leaf area index, etc.) have been measured during two field seasons (2022 and 2023), and time series of multispectral imagery have been collected, processed, and used in different regression modeling strategies for sparse and dense forage grass canopies. We also provide examples of automatic classification of turf grasses visual ratings using close-range UAV imagery (infrared and multispectral). Accuracy results from our independent validations have been highly variable for the field-measured traits with excellent results for traits like biomass, and moderately acceptable results for other important features such as grain yield. We describe challenges that impact our ability to model certain traits, and expectations from using hyperspectral and light detection and ranging Lidar sensors in the near future to a) expand the number of phenotypic traits, b) simplify workflows, and c) upscale current models from experimental plots to landscapes. Our HTP work aims at accelerating the process of selecting plant material that scientists at FRRL develop to restore disturbed semiarid landscapes in order to augment their resilience to the impacts of climate change and other global processes.
Unmanned aircraft systems (UAS) are now widely used in field phenotyping by industry and academia because they are convenient, fast, and require less labor than manual phenotyping methods. UAS-based phenotyping requires significant effort in data collection, processing, and resource management, as well as specialized expertise in flight operation and remote sensing. The Indiana Corn and Soybean Innovation Center (ICSC) at Purdue University operates multisensor platforms to provide reliable field phenotyping services to many clients in academia and industry. According to the experience with our multisensor platforms to date, we present practical data collection and processing pipelines for (a) creating geospatial datasets from RGB, hyperspectral, lidar, and thermal sensors, (b) extracting plant phenotypes from the multisensor datasets, and (c) combining the resultant datasheets with ground truth and treatment information. While processing techniques may vary depending on the vendor or sensing mechanisms, the general concept is up-to-date and can be relevant to those who are currently operating or planning to adopt UAS phenotyping systems in the near future.
Cadmium (Cd) is a naturally occurring toxic heavy metal found in trace concentrations in most soils. However, some soils have high levels of Cd due to parent materials or anthropogenic factors such as industrial waste and legacy fertilizers contaminated with Cd. Complications arise when Cd bioaccumulates in soil and plant tissues with carcinogenic effects in humans such as kidney and skeletal dysfunction. Leafy greens including spinach and kale are especially susceptible to Cd accumulation due to high translocation of the metal into the edible leaf tissues of the crop. However, while high concentrations of Cd in soil are generally toxic to plants, unlike other plant stresses, Cd toxicity has no distinguishing visual features that indicate stress or uptake of the metal in the plant, making it difficult to determine if remediation strategies are effective. In this study, emergent plant sensing tools such as Hyperspectral Imaging (HSI) and red-green-blue (RGB) wavelengths were utilized as high-throughput screening techniques to quickly distinguish wavelengths reflected by spinach plants grown in Cd contaminated soil. Feature selection of the HSI and RGB bands were used to identify regions of interest that exhibited significant differences of Cd uptake within multiple spinach genotypes previously shown to vary in Cd uptake. Applications of these technologies provide nondestructive sampling of crops to identify wavelengths sensitive to Cd stress and uptake, with potential applications using machine learning to provide a framework to model these interactions, informing plant breeding programs, and ultimately reducing human health risks.
Machine vision techniques like RGB imaging are valuable for measuring plant growth by tracking changes in size and shape over time. By using non-destructive, continuous monitoring, these techniques are especially useful for assessing canopy growth under different experimental conditions like drought. However, canopy monitoring has limitations, like being unable to observe kernel growth hidden in spikes. To address these challenges, we did a study on the growth of a recurrent parent Yecora-rojo (called Yecora) and its 2nd backcrossed progenies under pre-heading and post-heading water limitation. The goal was to evaluate RGB imaging’s capability in quantifying their differences in tolerance to limited watering. The plants were grown in a controlled environment. RGB sensors captured images three times a week until 20 days after heading. The side projected area (SPA) was used as an indicator of growth. We showed the difference in recovery after stress by using the numerical approximation of the area under the curve for two progenies that outperformed Yecora under limited watering. Furthermore, the continuous measurements let us identify specific time points contributing to these differences. The grains from individual plants were phenotyped using a laser-based size determination system, evaluating length, width, and area per kernel. This method provides a complete picture of the total kernels produced by a single plant and helps study varietal differences. This study identified two BC2 progenies, Yecora156 and Yecora190, as targets for further field studies and breeding to develop the next breeding population.
In the quest to enhance citrus yield estimation, this study leverages Multi-Temporal Unoccupied Aerial Systems (UAS) datasets to develop a refined tree-level yield estimation model. Conducted within a 2.22-acre orchard in Weslaco, South Texas, the research integrates data from 56 citrus trees, utilizing RGB and Multispectral UAS imagery collected over critical growth months from June to December (excluding July and August) 2022. The methodology encompasses the collection of UAS data processing using some custom-developed algorithms and automated workflows and the novel application of machine learning algorithms-namely, multiple linear regression, gradient boosting regression, and random forest regression. The study innovatively extracts 11 key phenotypic features, combining tree canopy structural and spectral information to estimate yield with increased accuracy. A significant advancement is proposed in the form of an improved individual tree boundary delineation method. This method addresses the inaccuracies of previous solid shape-based approaches, contributing to more precise feature calculation and improved model performance. Our experimental results indicate the single month whose data is particularly predictive, with the Random Forest model demonstrating robustness and consistency across temporal datasets. The multi-temporal approach confirms that comprehensive data integration yields superior estimation models. This ongoing research promises to bridge the yield estimation gap and set a new standard in precision agriculture methodologies.
The sensitivity, speed and reproducibility of modern mass spectrometers enable deep new function looks into the cellular proteome. Because the dynamics of the protein ensemble link genotype to phenotype, knowledge on protein complex assembly and localization is important for marker-assisted breeding for precision plant breeding. Co-Fractionation Mass Spectrometry (CFMS) enables systems-level analyses of protein complex dynamics. Our CFMS pipeline accepts soluble and membrane-associated cell fractions from agriculturally important cell types and uses orthogonal chromatographic separations, reproducibility filters, and correlation analyses to predict localization and composition based on experimental data. We have applied the CFMS method to analyze protein complex dynamics as a function of dark-induced metabolic stress. We discovered dozens of interesting protein complex rearrangements that likely reflect an adaptive response to reduced energy status. Similar assays have been used to analyze the system of protein complex rearrangements that are observed in a single gene knockout or those that occur in dark grown hypocotyls after short term treatment ethylene. A current project focuses on the protein multimerization and localization dynamics that occur during the development of unicellular cotton fibers. This methodology has great potential in gene discovery and systems-level phenotyping tool. We are organizing the data from our multi-omics studies so that it is findable and useful to the community with the goal of accelerating the genetic engineering of plant traits.
Dynamic process-based plant models are computerized representations of plant growth, development, and productivity that use measurements of environmental and physiological processes as input data to make predictions. However, the use of process-based models in phenotyping programs still faces challenge to parameterization across multiple genotypes, such as (1) the need for extensive and intensive datasets as parameters for running simulations, many of which are obtained using destructive and disruptive sampling, (2) the lack of systematic approaches needed to parameterize across extensive collections of genetically distinct, but often related, individuals that are typical of breeding populations. Remote and proximal sensing are potential alternative sources of data to inform parameters of process-based plant models because they can provide fast and non-destructive estimations of plant biophysical parameters across spatial and temporal scales. Over the years, several approaches have been proposed to leverage sensing for model parameterization, from simple empirical tuning to inverse modeling approaches. Continued inquiry into how best to use remote and proximal sensing data to estimate model parameters is critical to the future scale-out of these models for breeding programs and simulating processes that underlie complex genotype-by-environment interactions. Here, we present a decision flowchart to provide a visual representation of the sources and steps for process-based model parameterization that could be used as a guide for researchers working with remote sensing data and crop modeling across numerous genotypes.
Using hyperspectral technology for high-throughput plant phenotyping is a potentially useful method in crop sciences. To examine its effectiveness, we collected leaf-level hyperspectral and ground-reference data from rice plants grown in controlled-environments under drought and CO2 treatments at Ag Alumni Seed Phenotyping Facility at Purdue University. By applying RReliefF, we found that short-wave infrared region (SWIR) was important for leaf water potential (LWP), near-infrared region was linked with specific leaf area (SLA) and both red-edge and SWIR regions were related to gas exchange traits (net assimilation [An], stomatal conductance to water vapor [gsw] and transpiration [Emm]). Based on those results, we found that LWP and SLA were moderately predictable and gas exchange traits were predictable (R2 \(\geq\) 0.60 and root mean squared error of prediction for An, gsw and Emmwere 7.706 \(\mu\)molm-2s-1, 0.282 molm-2s-1, and 3.906 mmolm-2s-1 in validation datasets, respectively) by using partial least squares regression. Furthermore, treatment effect on An from cross-validation predictions agreed with ground-reference data. In contrast, photosynthetic parameters (Vcmax and Jmax) could not be estimated from hyperspectral data. Hyperspectral data can provide potential insights about plant growth and water status. When the effect of treatments is pronounced, model predictions are consistent with ground-reference data.
Roots are vital for crop development, facilitating water and nutrient uptake, adapting to soil conditions, and supporting above-ground plant growth. Sorghum, a climate-resilient and versatile cereal crop, plays a significant role in global food security and bioenergy production, making its study especially relevant. Analyzing root architecture (RA) traditionally involves destructive and time-consuming methods that preclude longitudinal observations of individual plants. Here we used minirhizotrons (MR), wireless, non-destructive devices developed to capture root imagery. With MR, we aim to identify sorghum varieties that are conducive to prolonged underground carbon storage through larger and deeper roots. Insights into sorghum’s RA could also better elucidate the link between RA and above-ground productivity. We conducted two experiments using the MR cameras. The 2023 summer field experiment at the Danforth Field Research Site (FRS) assessed the impact of tilling and cover crop practices on RA in two sorghum varieties, employing 96 MR tubes across 48 plots. An indoor trial on 35 sorghum lines used 73 MR tubes, with RootSnap software facilitating image analysis. Preliminary findings indicate that MR predicts root biomass with high accuracy, both overall and at each depth, and can determine root color with 90% accuracy. Finally, through the development of a neural network, we aim to utilize image analysis to predict root biomass at various depths, providing a labor-saving alternative to soil coring. We also address challenges, including the limited dataset size and image noise, through enhanced feature engineering and provide a comprehensive comparison of multiple machine-learning techniques.