Source: UNIVERSITY OF GEORGIA 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
Feb 15, 2021
Project End Date
Feb 14, 2024
Grant Year
Project Director
Kim, S.
Recipient Organization
ATHENS,GA 30602-5016
Performing Department
School of Environmental, Civil, Agricult
Non Technical Summary
Coastal salt marshes are very important ecosystems as they protect us from hurricanes and floods and provide critical habitats to many fish and shellfish nurseries. They improve water quality by converting nutrients, such as nitrogen, to forms that are less harmful to the environment. They also play an important role in converting atmospheric carbon dioxide into soil constituents and then storing that carbon. Salt marshes face multiple threats, including droughts, sediment starvation, nutrient pollution, and sea-level rise. In Georgia alone, 20%-45% of salt marshes are at risk of loss to accelerated sea-level rise, which could drastically reduce the amount of carbon stored in these systems. Soil carbon is made by plant parts that do not fully decompose, and that helps salt marshes maintain their elevation against rising seas. Keeping marshes at higher elevations also protects against erosion. Knowing the amount of carbon stored in soils, therefore, is critical to understanding salt marsh vulnerability. To effectively measure salt marsh soil carbon storage, it is crucial to examine changes in carbon storage over space and time. Soil carbon is usually collected by scientists by removing a tube-shaped amount of soil from the marsh and taking it to a laboratory to measure carbon content. This practice is labor-intensive, time-consuming, expensive, and taken at locations and depths specific to individual projects. Such field sampled data, although potentially highly accurate, are inevitably from limited geographic footprints and not representative of the variability of soil from place to place. Site-specific soil carbon data are collected using traditional sampling methods from coastal salt marsh ecosystems as part of individual projects or coordination networks. For example, the National Science Foundation (NSF)-funded Coastal Carbon Research Coordination Network (CCRCN) has been created to compile and house soil carbon contributed by different projects across the globe. This is a very important and useful network, but we still know about soil carbon from only limited areas and times. For example, the entire coast of GA is represented by only 37 sites that are provided at irregular time points. To overcome this shortcoming, we propose a new framework that we call "AWeSOM Sense" (A Wetland Soil Organic Matter Sensor), which will transform salt marsh soil carbon sensing and complement all ongoing and existing efforts with high-quality soil carbon data at many more locations than can be collected manually. AWeSOM Sense will strengthen the current efforts of CCRCN by expanding the footprint of the existing point data to broader scales, depths, and times. A method that enables data to be continuously transferred through existing cellular and WI-FI networks without humans having to visit the sensors allows automated monitoring of salt marsh soil carbon. All of that data can be input into computer models to make predictions that can eventually be translated across large geographical footprints.The scientific merit of this study lies in its potential to radically transform the current landscape of soil carbon monitoring in salt marsh ecosystems by inventing techniques to (1) perform a comprehensive evaluation of the effectiveness and reliability of various, low-cost, soil carbon sensors placed in the field to collect and transmit soil carbon data; (2) build a computing and data transmission architecture that allows data ranging from field sensors to drones to satellite images to be stored and analyzed in new ways that optimize where sensors are placed in the field and then seamlessly integrates all of those data through intelligent collaboration between state-of-the-art computer technology (e.g. cloud and edge computing infrastructures); and (3) develop computer models that can find trends in those large data sets that humans cannot necessarily see to predict both near-surface and deeper soil carbon content for salt marsh ecosystems across coastal GA. The project will impact a broad audience because it will generate new knowledge in many specialized areas, such as field sensor location and data transmission in environments with poor cellular and wi-fi service, soil carbon sensing, and ongoing soil monitoring. These results will be directly beneficial to both academia and government. Among the major stakeholders of the project are research communities working on wetland ecology projects throughout the coastal US, including NSF's GA Coastal Ecosystems Long Term Ecological Research (LTER) project, which is within our proposed study area. Our goal would be to create links between the different coastal ecology working groups to inform them about the project outcomes. The project will engage students, community leaders, resource managers, and the general public via training, workshop, and social media in various aspects of the research, starting from environmental sensor deployments/management, data acquisition, processing, and interpretation to salt marsh ecology and carbon sequestration. Importantly, the GA DNR Coastal Resources Division has prioritized research into the importance of measuring marsh contributions to ecosystem health (or "ecosystem services"), and a broad-scale assessment of soil carbon is essential to achieving their next step of quantifying the ecosystem service of carbon storage. The project has specific tangible steps facilitating seamless transfer of environmental information to resource managers, ensuring early implementation of restoration actions for enhanced sustainability. Our research results will also be disseminated through high-quality scientific journals and presentations in leading conferences in coastal ecology, soil science, and sensing systems.
Animal Health Component
Research Effort Categories

Knowledge Area (KA)Subject of Investigation (SOI)Field of Science (FOS)Percent
Goals / Objectives
Soil organic matter (SOM) data are typically collected using traditional sampling methods from coastal salt marsh ecosystems, which is time-consuming, labor-intensive, and expensive. This traditional sampling method provides discrete (or point-specific) SOM information. Point-specific methods alone are not sufficient to provide comprehensive and continuous evaluation of near-surface and deeper SOM information in salt marshes. The goal of this project is to develop surface and SOM distribution maps across coastal GA at a seasonal timestep. The overall objectives of this project are as follows:First, we will develop an end-to-end, cost-effective, scalable, and robust sensing framework that leverages low-power, low-cost, wireless, in-situ/multifunctional sensors to measure SOM data in salt marshes.Second, we will create a scalable cloud and edge computing system that collects, transmits, and stores sensing data for facilitating sensing data analysis.Third, we will develop new machine learning models that can accurately predict belowground SOM from surface reflectance across multiple spatio-temporal granularities to allow this technology to be applied to additional geographic areas.
Project Methods
Surface and belowground SOM is a critical soil property to salt marsh resilience. Surface and belowground SOM data in salt marshes obtained through remotely and continuous data collection in conjunction with field observations and predictions from machine learning algorithms will provide insight into the challenges faced by salt marsh research groups. Soil color is an important physical property used by soil scientists and promoted by the USDA Natural Resources Conservation Service (NRCS) as a strong indicator of SOM content. The Munsell Soil Color Chart (MSCC), which is commonly used to ascertain soil color, often can be found in soil series descriptions and online databases provided by the NRCS to characterize soils. To accomplish the study objectives, three major tasks (I, II, and III) have been developed:In Task I, the team will deploy Nix™ Pro Color sensors in three coastal salt marshes - near the University of Georgia (UGA) and DNR resources at Skidaway Island, Sapelo Island, and Brunswick. During site visits at six-week intervals, eight sensors will be deployed at each field location along with an already established color and salinity gradient and left to acquire and transmit data for six weeks. Sensors will be attached to floats in a "cage" that moves vertically along a weather-resistant post to ensure it rises and falls with the tides without dislodging from the horizontal plane. After six weeks, the sensors will be collected, defouled, and re-deployed at a nearby site in a different configuration determined by sensor responses from the previous deployment.In Task II, the team will develop the AWeSOMSense platform, a system for smart data collection, storage, and data transmission of field-measured SOM data from Nix color sensors. The AWeSOMSense platform is composed of three layers: sensing, edge, and cloud. The sensing layer includes a set of Nix soil sensors and network infrastructure. The edge layer is responsible for processing sensing streams from SOM sensors and managing sensor deployment/availability. The cloud layer is placed in public clouds, such as NSF funded CloudBank and CloudLab, and this layer stores/analyzes the remote sensing data as well as creates/trains deep learning models used in the edge layer.In Task III, the team will develop and evaluate the predictors that estimate surface and belowground SOM with training data obtained from Task I and Task II. Determination of an accurate prediction model for surface and belowground SOM is a critical component of this project. The prediction model developed in Task III will harness recent advancements in machine learning and deep learning technologies using airborne drone data and spaceborne satellite data (Landsat-8 and Sentinel-2).The performance of the AweSOMSense Prediction framework will be evaluated by comparing its prediction results with the data obtained from Task I. The performance (accuracy) of the prediction framework will be measured by RMSE (Root Mean Squared Error). The baselines will be RF and XGBoost, which were the two most accurate predictors from our preliminary work with a RMSE range of 0.076 ~ 0.124 and R2 of 0.92 ~ 0.97. Ultimately, the model will be cross-validated by estimating the test error.

Progress 02/15/21 to 02/14/22

Target Audience:Our team has foster relationships with the team of the NSF's GA Coastal Ecosystem Long Term Ecological Research (GCE-LTER) project and coordinated our efforts with GCE-LTER's ongoing long-term data collection programs to augment and expand our field data. We also participated in the Georgia Coastal Research Council meeting to network with other in-state researchers.We also have contacted with Georgia Department of Transportation (Natural Resources Division)that is interested in coastal restorationto strategize the dissemination of project information. Changes/Problems:The field data collection continuedto suffer in the wake of the COVID-19 pandemic. Althoughthe research team has remained prepared for field works, some delays and no cost extension (NCE) for USDA-NIFA's approval is acticipated.The research team will submit a NCE request after the first anniversary date. What opportunities for training and professional development has the project provided? Nothing Reported 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?Platform The research team will createa sensing platform to hold and protect the sensors and devices on a fixed spot in wetlands. More specifically, the arms and legs are built with 2" PVC, and the central post is 4''. And, the solar panel is exposed outside. The camera, power management chip (Witty Pi), and Raspberry Pi are installed inside the post. The post is waterproof by sealing the bottom and capping the top.The wildlife camera, action camera, iPhone 7 will be attached to the platformto collect the field marsh soil images. Nix mini and Nix pro will also be attached to the platform toscan the soil color by direct touch on the soil. Raspberry Pi will be used as the edge server to collect the sensing data and take the computation work. Solar panel will be used to generate power that is supplied to Raspberry Pi and sensors.

What was accomplished under these goals? Wetlands store almost 20%-25% of the world's soil organic carbon (SOC) stock in the form of soil organic matter (SOM) with just 4%-6% of the world's land area (Yu et al., 2012). However, recent studies have indicated striking losses of wetlands in places like North America, where wetland loss is estimated to be 53% in the contiguous United States, 16% in Canada, and 62% in Mexico (Cavallaro et al., 2018; Kolka et al., 2018). Wetlands are the most extensive reservoir of soil organic carbon (SOC), and the loss of this carbon often leads to major increases in carbon dioxide in the atmosphere. Therefore, monitoring, modeling, and mapping SOC changes in wetlands become a critical research area. Studies on SOC mapping using remote sensing data and digital soil mapping (DSM) techniques are rare as field-sample-based monitoring of carbon in wetlands can be cost-prohibitive and labor-intensive. To help alleviate this problem, we present a novel cyber-architecture-based data-centric framework referred to as AweSOMSense (A Wetland Soil Organic Matter Sensor). AweSOMSense explores the usability of low bandwidth, continuous, and automated sensing techniques for salt marsh SOC using low-cost and wireless in-situ sensors. AweSOMSense is based on developing remotely operated sensor platforms capable of acquiring data from salt marsh wetlands using Nix color sensors, other soil property sensors for pH, salinity, moisture, Redox potential, temperature, and digital images. All these factors have been linked to SOC in the literature, and there are cheap wireless sensors available that can collect these datasets, log them through Raspberry Pi or Arduino-based computing boards, and transmit them through the cloud to the lab for modeling. Data collected by the sensing platform is sent to an edge layer for sensing stream procurement, sensor deployment management, and sensor fault detection (QA/QC) before the data are uploaded to the cloud. The edge layer sends the refined sensing data to the cloud layer, which stores and analyzes the field sensors and satellite data together in a machine learning framework. The SOC prediction model is validated with field soil core data analyzed for SOC in the lab. We anticipate that a cyberinfrastructure-based data collection approach would help us understand soil processes in lesser-known and data-scarce ecosystems such as salt marsh wetlands. Field sampling and data collection To date, cores have been collected at two locations on three occasions: Skidaway (August 2020, 16 cores), Sapelo (Sept 2021, 15 cores). At each core location and in between them, the soil surface was sensed using SuperGER hyperspectral, Nix color sensor, and other field soil properties (pH, salinity, and redox) via handheld instrument. During the September samples at Dean Creek, sample collection was synchronized with drone data collection that had hyperspectral, multispectral, and Lidar sensors. Laboratory sample processing and data collection Deep cores from both Skidaway and Sapelo were transported from Sapelo Island to the UGA Athens campus. These soil cores were sliced, and hyperspectral and Nix color readings were made on the fresh cuts at 5 cm depth intervals. The sectioned soil samples were then divided in 5 cm intervals. Slices were subsampled for bulk measurements, and the rest of the sample was used for soil organic carbon measurement. Approximately 100 samples have been pulverized; the remaining samples are in progress. Bulk density and soil moisture has been collected from all samples, and some loss on ignition is complete. After pulverizing, these samples will be evaluated for carbon content using the dry combustion method to calibrate the loss-on-ignition method. Soil organic carbon analysis is in progress. Bulk density and soil moisture-based inferences are discussed in later sections. Results The hyperspectral sensor used in the lab for soil depth at 5 cm interval provides a 1 nm bandwidth from the range of 350 nm to 2500 nm. Water bands around 1450 nm and 1950 nm were very apparent, but a subtle dip around 1200 nm. However, around 970 nm there is a break in spectra for the Flux tower site. This could be due to sensor fluctuations and needs to be checked for other locations. Prominent water absorption signals confirm the saturated condition of soils in wetlands. Other than the water absorption band there is also a chlorophyll absorption feature around 650-700 nm, this is because of organic matter, diatoms, and other plant biomass in soils. There is also a weak kaolinite absorption around 2200 nm, however, kaolinite is more prominent in northern Georgia than in Coastal Georgia (Demattê et al., 2017; Woodruff et al., 2015). The influence of organic matter makes the spectra concave and decreases the reflectance in general, in the visible light region of the spectra (Demattê et al., 2017). There is a weak concave shape and less reflectance for lower depth that may reflect for higher organic matter, comparison of spectral signals with soil organic matter measurement would help to understand it more. Soil organic carbon and soil bulk density are related soil properties. As soil organic carbon measurements were still in progress, soil bulk density readings were used for further analysis. The objective was to predict soil bulk density for various soil depths using hyperspectral bands. A correlation plot was made between spectral bands and bulk density measurements from various depths. The Pearson correlation coefficient between spectral bands and bulk density measurements ranges from 25% to 50%. Then spectral bands with higher correlation from the neighboring bands, apparent from a small hump were chosen to be used as covariables for a model. A total of 6 bands were chosen and they were 565-575 nm, 665-675 nm, 1470-1480 nm, 1875-1885 nm, 2205-2215 nm, and 2310-2320 nm. The correlation coefficient between bulk density and six selected bands ranges from 31 to 50%. However, there is high collinearity between some of the bands like Band 565-575 nm and 665-675 nm. All the bands have a higher than 60% correlation with each other that could bring multicollinearity in the models and needs to be accounted for.The linear model between bulk density and six selected bands explained around 31% of the variability. Out of six bands, band 1875-1885 nm, band 2205-2215 nm, and band 2310-2320 nm were significant for the model at a 5% significance level. A random forest-based model did slightly better. The variability explained by the model for training data was 91% (R2) but explained variability drops drastically for testing data to 24%. The drop in R2 for testing data could be because of collinearity or overfitting.


  • Type: Journal Articles Status: Published Year Published: 2021 Citation: Salehi Hikouei, I.; Kim, S.S.; Mishra, D.R. Machine-Learning Classification of Soil Bulk Density in Salt Marsh Environments. Sensors 2021, 21, 4408. 10.3390/s21134408