Source: UNIVERSITY OF FLORIDA submitted to
Sponsoring Institution
National Institute of Food and Agriculture
Project Status
Funding Source
Reporting Frequency
Accession No.
Grant No.
Project No.
Proposal No.
Multistate No.
Program Code
Project Start Date
Jan 1, 2021
Project End Date
Dec 31, 2024
Grant Year
Project Director
Zare, A.
Recipient Organization
Performing Department
Non Technical Summary
Studying plant roots, the soil surrounding roots, and their interaction in realistic environments over time is challenging. One of the most successful methods to do so is using minirhizotron tubes. These are clear plastic tubes inserted into the ground. Then, cameras placed in the tube can take photographs of the soil and roots that happen to grow alongside the tube over time. Minirhizotron tubes have been very successful in scientific studies to understand temporal root properties. However, they are expensive, tedious to use, and only collect standard color imagery. In this effort, we will build hyperspectral cameras (which also collect information outside of the visible range collected by standard color cameras) that will be inexpensive, have an automated mechanism to move up and down the tubes, and are paired with automated algorithms to process and understand the collected data. The project will open doors for understanding plant roots and their interaction with the soil surrounding them beyond our capabilities today.
Animal Health Component
Research Effort Categories

Knowledge Area (KA)Subject of Investigation (SOI)Field of Science (FOS)Percent
Goals / Objectives
Minirhizotron technology is widely used for studying the development of roots. Such systems collect color imagery of plant roots in-situ using an imaging system within a clear tube inserted into the soil. They have been shown to be very effective at tracking root characteristics, architecture, and growth over time. However, these imaging systems have several significant limitations. First, minirhizotron systems are extremely costly. Each imaging system is generally on the order of tens of thousands of dollars which, in practice, restricts the number of systems that can be deployed by a team. Second, the imaging systems are tedious to deploy. Generally, operators need to manually lower the imager during the imaging process further restricting the throughput of the system. Third, a significant bottleneck of current systems is the post-processing of collected imagery. Each image is manually analyzed and traced to identify and characterize each root, an extremely tedious and time consuming process. Finally, current minirhizotron systems focus primarily on root analysis and monitoring with little regard to the soil. In our effort, we will tackle some of these most critical challenges by developing a low-cost, easy-to-operate hyperspectral imaging system and associated automated analysis algorithms for use with minirhizotron tubes. The minirhizotron hyperspectral imaging systems will provide root and soil characterization beyond what is possible with the current color imagery being collected. Specifically we will focus on characterizing soil moisture, texture, and soil organic matter to allow us to eventually model critical biogeochemical processes. We have assembled a transdisciplinary team with experts on image analysis and computer vision, sensor design, model and data fusion, machine learning, plant physiology, and soil science. In addition to developing an improved minirhizotron imaging system, we will answer fundamental questions about the interaction between root and soil properties.By developing a hyperspectral imaging system for minirhizotron tubes, we will enable the collection of richer data for root and soil analysis over time. Hyperspectral image analysis has been shown to be effective at characterizing soil nitrogen, carbon, carbonate, and organic matter as well as effective at monitoring root decay and providing the potential for physio-chemical root analysis. Providing the ability to perform this sort of analysis over time in field settings would have revolutionary impact in the study of plant roots, soil, and their interactions. The potential for monitoring root and soil properties over time in-situ can open possibilities for significant advances in biogeochemical modeling and understanding relationships between root architecture, soil properties and root function. Furthermore, the algorithm development efforts on this project will enable automated joint spatial-spectral analysis of root and soil interactions.This effort will enable a collection of broader impacts that will expand the accessibility and utility of minirhizotron imaging systems. We will develop and freely share software pipelines for processing and analyzing minirhizotron color and hyperspectral imagery. Further, because of our team's unique approach to achieving full integration of disciplines through co-development of research questions, approaches, data analysis, and interpretation, we will provide a unique opportunity for graduate student and postdoctoral training of convergent research. Lastly, our novel professional development programs, including the A2i, symposium and A2i+ workshops, we will provide all graduate students and postdoctoral associates mentorship from relevant industry professionals. Ultimately, we will provide intensive project-based transdisciplinary, convergent training for graduate students in computer science and engineering, plant physiology, and soil science.
Project Methods
A. Sensor Development Research PlanTask A.1: Characterize ability of low cost, broadband filters in approximating spectral signatures:In this task, we will create a database of broadband filters from Roscolux filter bankand a small set of target soil/root tasks that already have relevant signatures.We will use non-negative optimization to find coefficients to minimize a linear combination of filters using sparse coefficients.Task A.2: Model the ray geometry of 3D printed pinhole imagery through the MR tube material and propose a calibration algorithm: This take will consist of selecting a web-camera that fits in the MR tube and collecting data through the side of the tube onto an LCD display. We will collect de Bruijn patterns mapping the flat, display plane to the pixels on the device. Then, we will test the above with pinholes printed onto the web camera of different sizes (3x3 and 4x4) and using two kinds of optics: (1) basic pinhole; (2) pinhole immersed in refractive slab (for wide-angle viewing).Task A.3: Build a prototype device with a LED ring and apply multi-view geometry to align the imagery to obtain hyperspectral signatures: In this task, we will select an LED signature that is broadband in our targeted reflected radiances. Then, we will attach an LED ring to the web-camera with 3D printed setup. Integrate color-filters onto the device. Using this, we will collect data of a variety of scenes and create a multi-view algorithm to map between different cameras. One version will be to use a geometric-based approach by sliding the camera along a rail in the tube, collecting epipolar plane images (EPIs) that allow for line detections (with Hough transforms) to recover depths from the imagery. Another strategy is to use a learning-based approach by collecting data from the sensor through the tube siding onto scenes that are also captured in the lab using a Kinect. A mapping learned between the imagery and the depth map can be obtained, and used to align the array.Task A.4: Construct robotic system for moving camera through the tube: We will actuate the novel sensor described above using an automatic image capture system. The final goal is to create a setup which can image and seamlessly stitch the imagery into a "panorama-minirhizotron" image that can allow for better user viewing, easier data collection, and also can form a geometrically-salient visual context for the machine learning algorithms to process.B Research Plan for Laboratory, Mesocosm and Field StudiesTask B.1: The Rowland lab has established a rhizobox imaging system (Racette et al., 2019) which enables digital and HSI of roots and associated soils through one side constructed of clear plastic. Seeds are planted within boxes, inclined at 30 degrees and scanned or imaged using cameras through the plastic. Variations of the rhizobox construction have allowed for separation of soil sections in thirds: a top section roughly 12 cm in depth, and two side by side chambers of approximately 7x5 cm. These chambers allow the placement and maintenance of soil at various predefined water contents, SOM, texture, and bulk density. Lab analyses will be conducted testing different soil water contents, textures, SOM, and compaction in the rhizoboxes with peanut (Arachis hypogaea L.) and sesame (Sesamum indicum L.) seedlings. Rhizoboxes will be imaged daily using the lab-based hyperspectral camera.Hyperspectral signatures that are indicative of the manipulated soil properties will be identified.Because the clear plastic front of the rhizoboxes are removable, images can be taken with and without the cover to determine the impact of imaging through this plastic. This information can be used to aid in algorithm development to understand the impact of the plastic on HSI measurements and analysis.Task B.2: Evaluate Soil Water and Root Differences with Hyperspectral Sensor at Mesocosm Scale. MR ground-truthing data will be the ultimate and essential test of the sensor's capabilities. We propose testing in a variety of soil types, conditions, crops, and management treatments. Further, these studies can quantify root water uptake when coupling MR imaging with a separate measurement using a Time Domain Reflectrometry (TDR) sensor that quantifies soil water. Studies will be conducted at both locations using the traditional MR imaging systems and the developed hyperspectral sensor. Visible images will be analyzed for total root length, surface area, root volume, and root diameter using the techniques established in the Zare and Rowland labs. Groundtruth data on soil properties will be collected including soil water, texture, and SOM.Task B.3: Field Scale Studies: Experiments conducted in peanut involving full irrigation (meeting crop demand) and partial irrigation using two cultivars will be conducted under rainout shelters at the Plant Science Research and Education Unit (PSREU) in Citra, FL. The experimental design will be conducted as a Randomized Complete Block (RCB) with four replications and a split plot restriction on randomization, where levels of irrigation (full and partial) are randomized to main plots and levels of cultivar (2) to subplots.C Algorithm Development Research PlanTask C.1: Guided Unmixing Development: In this task, we will study and develop approaches to determine hyperparameter settings and priors based on laboratory, mesocosm and field studies.Task C.2: Leverage Spatial Information during Unmixing: Many methods in the unmixing literature have been developed to leverage spatial information during hyperspectral analysis. Furthermore, spatial information has been leveraged during LDA usage. Of particular note are those that extend LDA to incorporate spatial information and structure (Wang and Grimson, 2008; Cao and FeiFei, 2008). However, many of these approaches are computationally intensive (i.e., methods relying on Markov Random Field models).Task C.3: Supervised ML Mapping Approaches: In parallel with unmixing approaches, neural networks (NN) and support vector machines will be trained to map hyperspectral data to measured root and soil properties.Task C.4: Evaluation of Approaches: Both the unmixing and supervised mapping methods will be evaluated against a suite of standard hyperspectral analysis methods previously published in the literature including linear and non-linear spectral unmixing approaches. These methods will be evaluated by their ability to estimate soil moisture, texture and SOM under cross-validation.D Biogeochemical Modeling Research PlanTask D.1: Formulation of Microscale Models: In order to link root traits to SOM, we will utilize HSI to separate out deposition of litter from exudation (inferred from presence of root tips, root hairs and possibly novel spectral features). We will parameterize and compare both a microbially-explicit model based on the formulation in Georgiou et al. (2017) and a simplified first-order model (based on CENTURY). We will combine model development and simulation with data-model integration using Bayesian Hierarchical Models (BHM) to leverage the unique advantages of the data and experimentation. We will perform posterior predictive checking - including leave-one-out cross-validation - using soil carbon concentration changes as our predictive quantity of interest, while treating microbial pools as latent variables.Task D.2: From Micro to Macroscale: Aggregation over microscale models will be studied analytically where possible, as well as through simulation experiments. The mathematical machinery we develop will be incorporated into our BHM fitting of the PBMs. Given that each rhizotron will provide data on 100 spatial locations during each measurement, we anticipate having rather large datasets with which to constrain these processes.

Progress 01/01/22 to 12/31/22

Target Audience: The target audience for our hyperspectral minirhizotron system, algorithm development, plant physiology and biogeochemical studies are soil and plant scientists/researchers carrying out field experiments and in situ data collection over time.The target audience for our education and outreach activities are graduate students and postdoctoral scientists. Changes/Problems: Nothing Reported What opportunities for training and professional development has the project provided? The students funded on this project were provided with research training through regular interaction with the PIs, participating in bi-weekly project meetings, meetings with their primary advisor, and participation in a lab seminar and journal club. Graduate students also received training on hyperspectral data collection using a hyperspectral lab camera system and imaging chamber. The students were also trained in hyperspectral data calibration and analysis techniques. Upon starting in November of 2021, co-PI's Hammond and Wislon initiated a postdoctoral development plan with postdoctoral research associate Yangyang Song. This included a pre-assessment of Dr. Song's needs for professional development, which are now targets of his postdoctoral development plan for 2022. Additionally, Dr. Song has already learned new techniques in ecophysiology (e.g., Scholander pressure chamber, hydraulic conductivity, optical xylem vulnerability) and spectroscopy, training with instruments to be used in ongoing experiments. How have the results been disseminated to communities of interest?The group has presented one poster presentation on results thus far (see Other Products). We are currently drafting multiple manuscripts and intend to submit them soon. What do you plan to do during the next reporting period to accomplish the goals? A. Physiology During 2023, we will work along with collaborators to refine our spectra model in predicting soil water content and get the full picture of water status and spectra signals of plants and soil during the dehydration process. A manuscript, "Spectra reflectance predicts plant and soil water status during dehydration (temporal title)," will be prepared and submitted. B. Algorithm and Computational 1. Semantic segmentation - We will train a model that incorporates both spatial and spectral HSI data to improve the average precision for segmenting plant roots. We are currently exploring the use of a 3D-UNET. Additionally, we are going to refine the overall segmentation pipeline to facilitate easier analysis of either root or soil pixels when given a plant root image. 2. Water potential problem - The regression model developed for the caps will be applied to the rhizobox dataset to predict water potential/content in those boxes where other measurements such as box weight and leaf spectra were available. The purpose of this analysis is to improve our understanding of plant health at different drought stages. C. Modeling The root injection experimental design created in 2022 will be executed. Data gathered from this experiment will be compared to the diffusive model, and microbial dynamics will be studied. We will also determine whether hyperspectral imaging can be used to analyze soil carbon dynamics. Lastly, a manuscript detailing the proposed model will be prepared.

What was accomplished under these goals? A. Physiology: We have completed the Rhizo-box dehydration experiments of peanut and sweet corn. Physiological parameters, including leaf water potential, leaf relative water content (RWC), leaf chlorophyll fluorescence, leaf photosynthesis, stem water potential, stem hydraulic conductivity, root water potential, soil water potential, and soil RWC have been collected during drying down and re-hydration (15 days for peanut and 31 days for sweet corn). During the experiment, leaf spectra reflectance was collected to link spectra signals to the plant using a spectroradiometer (SVC, 400-2400). A partial least square regression model (R2 = 0.9, RMSE = 5.55) using spectra reflectance to predict leaf water potential has been established. An independent dataset was used to test the established model (R2 = 0.81, RMSE = 13.38). To better link spectra signals to the soil, soil spectral reflectance was collected using a hyperspectral camera (HinaLea, 400-1000nm) during the whole experiment. In addition, a full soil water retention curve was established using both HYPROP and WP4C soil from the METER group to get an accurate relationship between soil water potential and soil water content. Seventy steel caps containing 2 grams of soil and varying water content ranging from 0 (oven-dry) to 1.5 (fully saturated) g were prepared to establish a preliminary spectra reflectance model. These steel caps were imaged using the HinaLea camera and measured by WP4C for water potential. A preliminary partial least square regression model (R2 = 0.83, RMSE = 13.73) was established using the spectra reflectance and water content of steel caps and tested by an independent dataset (R2 = 0.85, RMSE = 10.14). B. ?Sensor Construction and Calibration Accomplishments We have set up a hyperspectral sensor system for use in the minirhizotron (MR). There are two steps to this. The first was hardware design, integration and testing. We have now a 6 wide band hyperspectral sensor with an LED ring. The device is designed to fit inside an MR. As part of this device testing, we collected more than 50 images of the rhizoboxes with plant roots. For each of these images, we captured a corresponding "ground truth" hyperspectral image using the Headwall camera. We also have shown calibration results, by using a stand UNet network to map between the images collected by our low-cost, compact camera and the larger, expensive ground truth camera. We found that we are able to map between these two, enabling the use of our camera in an MR to obtain hyperspectral imagery. In addition to the in-lab results with rhizoboxes, we expect to test in a real MR within the next month. Finally, we have also shown algorithmic results on how to select the right LED bands. We have demonstrated first an algorithm to select the best 10 bands for reconstructing ground truth HSI images of rhizoboxes. Then we showed an adaptive algorithm that selects the best 10 bands per scene, by taking a color (RGB) picture of the scene and then using that to obtain the 10 bands. All these algorithms have been compared favorably with standard baselines methods. C. Modeling AccomplishmentsUsing R's deSolve package we have created a model to track simple diffusion of root exudate carbon through the soil over time in 1D, 2D, and 3D. In order to compare results and understand microbial dynamics, a root injection experiment has been designed to accompany it. In this experiment, 40 rhizoboxes have been outfitted with one artificial root each and will be filled with field soil from the University of Florida's beef research unit. The boxes will be divided into one control group and three experimental groups with each experimental group receiving an artificial root exudate. The substrates being used to mimic root exudation will be glucose, oxalic acid, and glycine. Over the course of eight weeks, each box will receive their substrate mixture once a week in doses where the substrate carbon is equal to 5% of initial soil organic carbon. Each substrate mixture will include its 13C counterpart for diffusion analysis at the end of the experiment. The controls will receive doses of water or water/food coloring. After each injection, the boxes will be imaged with the HinaLea camera for hyperspectral analysis of SOC. D. Algorithm and Computational Accomplishments 1. Semantic segmentation of roots: Using RGB rhizobox images, we have hand-labeled roots in more than 25 peanut root images to enable supervised image segmentation fine-tuning of a UNET and other variants. First, the model was trained on an existing roots dataset (PRMI, ref?). Then, the model's best learned parameters were used for fine-tuning on a portion of the hand-labeled rhizobox images. The best models show an average precision of 0.84 or higher for masking out the root pixels within these RGB images. Further morphological post-processing of these results enables us to achieve a higher precision. This will allow us to carry out future analysis of the root HSI bands without the interference of soil pixel reflectance values. 2. Water potential prediction:The data collected for soil water potential, as described in the Physiology section above, involved the collection of seventy steel caps and corresponding water potential measurements. These data were used to analyze the signature of different water potential levels in the soil. The ultimate goal was to construct a regression model capable of predicting water potential from hyperspectral imagery of soil.


    Progress 01/01/21 to 12/31/21

    Target Audience:?The target audience for our hyperspectral minirhizotron system, algorithm development, plant physiology and biogeochemical studies are soil and plant scientists/researchers carrying out field experiments and in situ data collection over time. The target audience for our education and outreach activities are graduate students and postdoctoral scientists. Changes/Problems:One of our original Co-PIs on the team (Diane Rowland) transitioned from a faculty position at UF to an administrative role (specifically, Dean) at the University of Maine. To address this, we identified an alternative Co-PI (William Hammond)to take over Prof. Rowlands tasks and aims in the effort. During this reporting year, we obtained the necessary permissions and modifications to enact this change of Co-PI. What opportunities for training and professional development has the project provided?The students funded on this project were provided with research training through regular interaction with the PIs, participating in bi-weekly project meetings, weekly meetings with their primary advisor, and participation in a weekly lab seminar and journal club. Graduate students also received training on hyperspectral data collection using a hyperspectral lab camera system and imaging chamber. The students were also trained in hyperspectral data calibration and analysis techniques. Upon starting in November of 2021, co-PI's Hammond and Wislon initiated a postdoctoral development plan with postdoctoral research associate Yangyang Song. This included a pre-assessment of Dr. Song's needs for professional development, which are now targets of his postdoctoral development plan for 2022. Additionally, Dr. Song has already learned new techniques in ecophysiology (e.g., Scholander pressure chamber, hydraulic conductivity, optical xylem vulnerability) and spectroscopy, training with instruments to be used in ongoing experiments. How have the results been disseminated to communities of interest? Nothing Reported What do you plan to do during the next reporting period to accomplish the goals?Physiology: During 2022, a dry-down experiment will be conducted with rhizoboxes, during which high-frequency hyperspectral imaging is paired with measurements of soil properties and plant physiology. Additionally, canopy reflectance will be measured with a spectroradiometer. Together, this time series of measurements (from a wet to dry state), paired with the data collected in 2021 of plant growth from germination to fully-grown-in rhizoboxes, will provide a rich dataset for algorithm selection and development to link soil process and plant processes with spectra, to identify the most important bands to include in the multi-spectral minirhizotron camera. Sensor development: We will expand the device to add color filters in the optical path. The PCB board is being reworked via Altium to hold more LEDs and possibly implement surface mounted components which would fit better in the compact minirhizotron tube. Currently working on camera calibration and possibly changing to another camera which would require more hardware to operate but would give much more control to the user. We will develop and jointly train a second network that selects for appropriate wavelengths for reconstruction. Communicate with the other team members to change the network towards prediction for tasks deemed appropriate for rhizosphere understanding. Modeling: We will continue to refine our analytical model, and study it numerically to derive hypotheses that can link to exudation and to a time series of spatiotemporal hyperspectral data. This model will be implemented within a Bayesian data assimilation framework to iterate with the outputs of the physiological and sensor development portions of this project. Automated Analysis: We will continue to adapt and extend existing hyperspectral analysis libraries to better address rhizosphere data analysis and incorporate what is learned from modeling, sensor development, and physiology components. We will develop approaches for analysis that leverage prior knowledge from experiments being run to better constrain the solution space.

    What was accomplished under these goals? Physiology Accomplishments: Progress toward determining the bands that may be most informative to include in the final camera was made during preliminary measurements and the setup of a dry-down experiment to be conducted in early 2022. An existing hyperspectral camera (HinaLea, 400-1000nm) was used to image plants grown in rhizoboxes, using the same acrylic material that will be present between the minirhizotron camera and soil/roots. Plants of varied life strategies (annual, perennial) and growth forms (herbaceous, woody) were imaged in rhizoboxes after germination (wet soil), and were then repeatedly imaged in time series as roots developed. To link hyperspectral signals in the soil to meaningful measures of plant physiology, several methods have been developed and project personnel trained to acquire these data coincident with the images being taken (as described in prior paragraph). Plant water content measurements were tested, along with determining the proper rehydration time, and postdoctoral research associate Yangyang Song was trained in these methods, along with measurements of plant water potential using a Scholander pressure chamber. Additionally, above-ground spectral reflectance measurements were taken using a spectroradiometer, another new skill for Dr. Song. These reflectance data, combined with the below-ground imaging of rhizoboxes, will allow us to investigate the relationship between signals above and below the soil. To link soil-specific qualities to the hyperspectral measurements of rhizoboxes, Dr. Song conducted a soil moisture release curve using a dewpoint potentiometer. To infer soil water potential, Dr. Song also developed a protocol for weighing boxes during imaging such that the water content of each imaged box can be used to estimate that plant's soil water potential. In upcoming drought experiments, this will provide the opportunity to investigate how soil moisture changes over space and time, relative to the location of old and newly growing plant roots. Sensor Construction and Calibration Accomplishments: Our goal is to create a compact camera that can fit inside a minirhizotron. The camera will have set of M LEDs set as rings, as well as a set of N color filters inside the optical assembly, and the intersection M*N will be the number of wavelengths measured. This year we built the full system camera without the N color filters, which will be next year's work. We used an extremely compact third-party camera capable of both RGB and infrared photos and videos. A temporary LED system was built to sequentially flash LEDs while the camera took a short video of a color chart. Using MATLAB, the color chart data was broken down to show a clear increase in reflectance when the LED color matched the color chart square. Preliminary testing of the LED system was done on two cotton plants with different soil states: watered and dry. The watered soil results produced a clear reflectance spike for green LEDs on the root, which is in line with previous experiments. To make sure the camera could fit into an MR system, we modeled the entire mechanics in Solidworks and developed an attachable rig to fit onto the height-adjustable metal rod. The rig is designed to hold a 9V battery, an Arduino nano microcontroller, the aforementioned small third-party camera, and a ring-shaped PCB board housing multiple through-hole LEDs of specific wavelengths. Multiple rig iterations were then printed with PLA plastic and slightly adjusted for better fit. Modeling Accomplishments: We developed a simple modeling framework that describes a simple microbial dynamics often used to describe soil organic matter breakdown, in combination with dissolved organic carbon transport. The model has been simplified to obtain closed-form solutions for boundary conditions where the root surface releases exudates suitable for microbial uptake and priming of microbes in the surrounding soil. These analytical solutions show that dissolved organic matter patterns, as well as soil carbon depolymerization, have a function of characteristic distance formed by the ratio of diffusion in soil solution and microbial carbon turnover. While the assumptions in the simple analytical model need to be tested, the framework can help to test hypotheses regarding the role of root exudates on soil carbon turnover, for which the imaging process could provide crucial data points. Algorithm and Computational Accomplishments: During this reporting period, we began development of automated analysis of collected hyperspectral imagery from both laboratory hyperspectral data as well as future sensor data. Analysis algorithms being developed include spectral unmixing approaches (algorithms which decompose hyperspectral signatures into constituent materials (i.e., "endmembers") and the percentage of each pixel associated with material, dimensionality reduction approaches via autoencoders, and automated segmentation into regions of interest. For all of these analysis approaches, we are beginning with the application of existing approaches to collected hyperspectral imagery and existing hyperspectral imagery from other studies (e.g, aerially-collected hyperspectral over peanut and cotton fields). Existing approaches that have been implemented and are being evaluated (and, eventually extended) include the Sparsity Promoting Iterated Constrained Endmembers (SPICE) methods for unmixing, U-Nets deep learning architectures for spatially-driven image segmentation, and standard feed-forward autoencoders for dimensionality reduction.