|
Impacts of Aquatic Vegetation Dynamics on Nitrate Removal in Karst Agricultural Streams: Insights from Unmanned Aircraft Systems and In Situ Sensing
Rosalia Agioutanti2, William Isaac Ford1,*, Michael Patrick Sama1, Timothy McGill3
Published in Journal of the ASABE 67(2): 321-336 (doi: 10.13031/ja.15791). Copyright 2024 American Society of Agricultural and Biological Engineers.
1 Biosystems and Agricultural Engineering, University of Kentucky, Lexington, Kentucky, USA.
2 Wildlands Engineering, Raleigh, North Carolina, USA.
3 Geosyntec Consultants, Jacksonville, Florida, USA.
* Correspondence: bill.ford@uky.edu
Submitted for review on 24 August 2023 as manuscript number NRES 15791; approved for publication as a Research Article and as part of the “Digital Water: Computing Tools, Technologies, and Trends” Collection by Associate Editor Dr. Debabrata Sahoo and Community Editor Dr. Kati Migliaccio of the Natural Resources & Environmental Systems Community of ASABE on 8 December 2023.
Highlights
Abstract. Fate and transport of nutrients in karst streams remains a pressing research need. The objective of this study was to couple high-frequency and remote imagery data to quantify spatial and temporal variability of aquatic vegetation and determine the associated impacts on in-stream nitrate removal in karst headwater streams. The study was conducted in a spring-fed karst stream in the Inner-Bluegrass region of Central Kentucky, USA. Ten Unmanned Aircraft System (UAS) campaigns were coupled with three-years of high frequency in situ data of water quality. Automated segmentation and image classification analysis was performed on UAS imagery, and spatial variability in floating aquatic macrophytes was quantified. Results demonstrated the utility of UAS images to capture spatiotemporal variability of floating aquatic macrophytes (duckweed) with high accuracy, but the UAS analysis poorly predicted algal biomass dynamics due to spectral interferences in the water column and shading by duckweed. Further, in situ water quality data was used to estimate stream metabolism using the Bayesian Single Station Estimator (BASE) model, with results demonstrating primary production and ecosystem respiration was driven by algal biomass. Stream metabolism displayed a shorter season (March-August) relative to other agricultural streams in the region (March-October), likely reflecting the hydraulic structures in the channel which reduce flow velocities in the stream reach and promote transition to duckweed cover during the summer. The impact of aquatic vegetation dynamics on nitrate was assessed using diurnal analysis and regression with dissolved oxygen. Results for March through November showed an average daily diurnal variation of 0.25 mgN/L, which was more than two-fold greater than winter months. During the growing season, maxima generally occurred between 6 a.m.-2 p.m., and minima occurred between 2 p.m.-12 a.m., but varied depending on prominent aquatic vegetation and dissolved oxygen dynamics. These results suggest shifting N removal mechanisms as aquatic vegetation changes throughout the year. Implications for numerical modeling of fluvial nitrate fluxes are illustrated through the evaluation of a previously developed AI model simulating nitrate concentrations at the watershed outlet. The findings highlight the importance of integrating novel datasets, such as those presented in this study, to evaluate and improve predictions of numerical models.
Karst agroecosystems cover 7%-10% of the earth’s land surface and are important landscapes for agricultural production; however, they generate excessive loadings of nitrate (NO3-) to receiving waterbodies (Weary and Doctor, 2014; Robertson and Saad, 2021). While many studies have investigated the nutrient flow pathway dynamics to springs in karst agroecosystems in recent years, less work has been conducted on the fate and transport of nitrogen in surface waters downstream of springs, which are recognized as hotspots for biogeochemical processes (Briggs and Hare, 2018). These streams often flow over bedrock or thin surficial sediment deposits and may have limited hyporheic interaction, resulting in aquatic vegetation being the primary driver of nutrient removal (Ford et al., 2019). In part, limitations in developing effective in-stream management strategies reflect a lack of spatially and temporally distributed datasets that can inform numerical models (e.g., O’Hare et al., 2018). Remote sensing data obtained from unmanned aircraft systems (UAS), and in situ water quality sensors (e.g., nitrate N) are increasingly being implemented to overcome this limitation, although there has been limited coupling of these technologies in karst agroecosystem watersheds. The focus of this study is on coupling remote and in situ sensing to quantify spatial and temporal variability in aquatic vegetation in these landscapes and assess the subsequent implications for in-stream nitrate removal dynamics and modeling.
UAS remote sensing can acquire precise qualitative and quantitative data that can be compared over space and time to monitor the progressive changes in the aquatic ecosystem and may be useful for monitoring aquatic vegetation dynamics in karst landscapes (Klemas, 2016; Jia et al., 2017). Visible spectrum red-green-blue cameras have relatively low cost and flexible deployment options, compared to multispectral and hyperspectral cameras, and have been broadly applied in aquatic vegetation studies (Flynn and Chapra, 2014; Husson et al., 2016; Malthus, 2017). Studies have shown that capturing aquatic vegetation dynamics in the fluvial environment is challenging due to the computational demand of high spatial resolution data and the heterogeneity and interferences that occur in an image (Brando and Phinn, 2007; Silva et al., 2008; Nelson et al., 2011; Visser et al., 2018). As a result, emergent vegetation, such as floating aquatic macrophytes, has been predicted with higher accuracy than submerged vegetation.
In situ water quality sensors have been used to estimate whole-stream metabolism and are particularly sensitive to submerged vegetation, such as benthic algae. High frequency measurements of dissolved oxygen and temperature have been used for more than 60 years to estimate rates of gross primary production and ecosystem respiration (Odum, 1956). More recently, dissolved oxygen and temperature measurements have been integrated within single-station estimation modeling frameworks to estimate whole-stream metabolism using Bayesian techniques (e.g., Grace et al., 2015). This methodology has been found to be useful for quantifying the primary production and respiration dynamics of benthic algae and their response to environmental stressors (e.g., Uehlinger, 2006).
Benthic algae and floating aquatic macrophytes (e.g., duckweed) are prominent aquatic vegetation pools in slow-moving headwater agricultural streams and have been found to significantly impact watershed nitrogen budgets (Ford et al., 2017; Kazama and Watanabe, 2018). Algae and duckweed are widely recognized to transiently store nitrogen through biotic uptake; however, duckweed also creates favorable conditions for permanent removal via denitrification, as evidenced by its usage in treating wastewater ponds, stormwater detention basins, and constructed wetlands for N-removal (Perniel et al., 1998; Körner and Vermaat, 1998; Eriksson, 2001; Körner et al., 2003; Peng et al., 2007a,b). The abundance of duckweed and algal biomass is recognized to impact the temporal dynamics of N cycling in agricultural streams (Bunnell et al., 2020). We postulate that coupling methodologies that are more sensitive to submerged aquatic plants (e.g., in situ sensors) and floating aquatic macrophytes (e.g., UAS) will improve our understanding of the impacts of these aquatic vegetation pools on nitrate.
The impacts of aquatic vegetation on surface water nitrate dynamics have been increasingly studied in the past decade by coupling high-frequency nitrate sensors with estimates of GPP (Heffernan and Cohen, 2010; Rode et al., 2016; Burns et al., 2019; Yang et al., 2019,2023). The availability of accurate and robust in situ sensing platforms has allowed researchers to measure nitrate-N at sub-hourly resolutions for long-term deployments (Snyder et al., 2018). Unlike primary production estimates, in situ nitrate sensors directly measure the integrated impacts of submerged and floating aquatic macrophytes pools on the nitrate concentrations in stream, as well as the impact of heterotrophic microbial communities. Previous studies have utilized these datasets to estimate the dynamics of autotrophic nitrate uptake and respiration, denitrification, and heterotrophic uptake (Heffernan and Cohen, 2010; Yang et al., 2019). While studies often assume tight linkages between diel oxygen variability and diel nitrate curves (e.g., Yang et al., 2019), decoupling of uptake and diel variability of N may occur when other drivers (e.g., denitrification) become prominent and exhibit diurnal variations. Cumulatively, these results suggest that analyzing shifting patterns in diurnal variability of nitrate alongside assessment of aquatic vegetation pools can provide insight into temporal controls on in-stream N removal.
The ability of numerical models to capture aquatic vegetation impacts on fluvial nitrate removal dynamics is critical to developing effective solutions in watershed nutrient planning. Artificial intelligence (AI)-based models are increasingly integrated into water resources management (Krishnan et al., 2022). Data-driven AI models may be advantageous over physically or conceptually based models of in-stream nitrate removal dynamics, given that existing models lack representation of the heterogenous removal processes that occur in aquatic vegetation in both space and time (O’Hare et al., 2018). Further, physical or conceptual models are highly parametric and often have great uncertainty due to equifinality (e.g., Ford et al., 2017). One methodology utilized in many hydrologic and water quality studies has focused on adaptations of extreme learning machines (ELMs), which are feedforward neural networks that achieve similar or better generalization performance, scalability, and faster learning times than support vector machines (Huang et al., 2012). Studies have demonstrated the ability of ELM based neural network models to predict dissolved oxygen dynamics (Heddam and Kisi, 2017), and flowrate (Deo and Sahin, 2016), which are key parameters influencing aquatic vegetation and in-stream removal dynamics for nitrate, and thus may be able to capture the aforementioned heterogeneity in nitrate dynamics in space and time without prior information on aquatic vegetation. Existing model evaluation approaches and guidance for nutrients generally focus on broad statistical metrics, or storm event-dynamics, but rarely consider the ability of models to represent daily variations in N-removal processes. Given the importance of dissolved oxygen to reflect stream metabolic processes and impact denitrification, utilizing nitrate concentration vs. dissolved oxygen relationships to evaluate model performance can provide insight into the model’s ability to represent temporal variability in N removal processes.
 |
| Figure 1. (a) Image of DJI Phantom 4 UAS. (b) Image of DJI Phantom 4 UAS capturing preliminary imagery along the stream reach of interest. (c) Image of the dense duckweed population from the surveyed stream in 2019. (d) Image that shows algae biomass (brown area), duckweed, and macrophytes (green emergent plants) collected from the surveyed reach in 2019. (e) In situ sensor housing located at the watershed outlet of Camden Creek. (f) The YSI EXO2 multi- parameter water quality sonde and SUNA V2 sensor being deployed in sensor housing. (g) Camden Creek watershed located within the Inner-Bluegrass region of Central KY. |
The objective of this study was to couple novel high-frequency and remote imagery data to quantify the spatial and temporal variability of aquatic vegetation and determine the associated impacts on in-stream nitrate removal in karst headwater streams. Specific tasks conducted to meet the overarching objective were: (1) assess the utility of visible spectrum UAS imagery to capture spatial coverage of submerged and floating aquatic vegetation pools in a karst headwater stream, (2) estimate gross primary production using high-frequency in situ single-station estimation methodologies and compare with temporal variability in aquatic vegetation coverage maps, and (3) assess the impacts on nitrate fate and transport by assessing diurnal variability in the nitrate dynamics and illustrate the ability of the novel dataset to inform nitrate model predictions.
Study Site
To meet the objectives of this study, data collection was carried out in Camden Creek, which is located within a karst agroecosystem watershed (drainage area of 10.69 km2) in the Inner-Bluegrass physiographic region of central Kentucky (fig. 1g). A temperate Midwestern United States climate prevails in the region, which includes moderately cold winters, warm and humid summers, and moderate transition periods in spring and fall. The surface tributaries in the watershed are shallow and emanate from springs. They flow over limestone bedrock and are mostly unshaded through grazed pasture with riparian vegetation, with low streambed sediment storage on the exposed bedrock, and limited hyporheic exchange (Ford et al., 2019). As a result, the fate and transport of nitrogen are primarily impacted by transformations associated with aquatic vegetation. This watershed was chosen because of an extensive historical hydrologic and water quality dataset at the site (Ford et al., 2019), a previous study of algal and duckweed impacts on water quality (Bunnell et al., 2020), a multi-year in situ high-frequency dataset at the watershed outlet (McGill and Ford, 2024), and AI-based modeling of flow and nitrate concentrations in the watershed (McGill and Ford, 2024).
Aquatic vegetation in Camden Creek is characterized by an abundance of submerged and emergent species. The in-stream vegetation is dominated mainly by floating aquatic macrophytes (duckweed) and benthic filamentous algae. Rooted macrophytes are present to a lesser extent but occur mainly in the near bank regions, due to a lack of sediment storage for roots along the bedrock channels. The aquatic vegetation is highly heterogenous, both spatially and temporally, in the stream channel. Based on recent modeling results, the cumulative aquatic biomass is typically at a minimum in winter, begins to develop in the spring (April-June), reaches a maximum cover during the summer (July-September), and continues to proliferate through the fall, depending upon hydrologic activity (October-December) (Bunnell et al., 2020). More specifically, duckweed dominates the vegetation during the summer months under the lowest flow conditions, while benthic algae are more prominent in the spring and early summer. Flow velocities in the channel provide favorable conditions during low flow summer months for duckweed to proliferate because flow control structures in the form of v-notch weirs were installed throughout the watershed as part of long-term monitoring efforts, creating non-uniform flow conditions. Velocities were difficult to quantify explicitly during duckweed blooms because flow became restricted to narrow passages that formed between dense vegetation.
Materials and Methods
Aquatic Vegetation Classification and Analysis with Visible Spectrum UAS Sensing
A DJI Mavic 2 Pro was the UAS used to collect visible spectrum images. The integrated camera was mechanically stabilized on a 3-axis gimbal and featured a 20-megapixel 12.2 mm by 8.0 mm CMOS sensor with a focal length of 10.27 mm. Two trial surveys were conducted in November 2019 and January 2020 to develop a mission plan and determine the optimal time of day to avoid the sun’s reflection (glitter or glint) from water surfaces, which can diminish the amount of detail in underwater imagery. The flights were typically conducted early in the morning to best constrain changing lighting conditions based on the position of the sun. At the time of this study, the ground control software (DJI Pilot, V1.7.2) lacked a linear mission plan for mapping narrow paths. Therefore, the automated flight path consisted of a waypoint mission with two parallel paths in opposing directions that followed the stream trajectory. The UAS camera was manually configured after takeoff to collect images at a 2 s interval on auto-exposure with the camera facing nadir. UAS takeoff and landing were manually conducted, and the automated flight path was initiated while hovering. The automated flight path traversed 4,532 m at a velocity of 5 m/s, and at a fixed altitude of 60 m above ground level relative to the launch point. The total flight time during the automated portion was 17.1 min. The average spatial resolution was 1.82 cm per pixel and less than 2 cm/pixel for the entire site. Front and side overlap between images exceeded 80%, with a minimum of five overlapping images for each pixel analyzed. The aerial surveys were completed roughly biweekly, over a 4.5-month period (16 June 2020-27 October 2020). In total, ten campaigns were collected. Individual images were processed into an orthorectified mosaic map using PIX4Dmapper software (4.5.6), and the map was imported to QGIS Desktop (3.18) for analysis.
Automatic image segmentation was performed on all collected maps using the Orfeo Toolbox (OTB) in the QGIS environment. The process of segmentation focused on the radiometric information of the pixels, the semantic properties of each segment, the image structure, and on other background information (color, intensity, texture, weft, shape, context, dimensional relations, and position) whose values described the association between adjacent pixels. Segmentation was performed on clips for each of the sampling dates. The mean shift algorithm was applied as it allows for more user-defined thresholds and is a flexible clustering technique. Using the OTB, the input image was divided into tiles, and then the mean shift segmentation was performed within each tile. This was done by grouping together neighboring pixels whose range distance was below the default range parameter and the spatial distance was below the default spatial range parameter. The minimum region size was adjusted since the parameter depends on the size of the clip, the resolution of the clip, and the density of the different features included in the clip. The default input value for minimum region size was 100-pixel units. After conducting a sensitivity analysis with values ranging from 50 to 1000-pixel units, it was concluded that to process images from the visible dataset, a minimum size of a region of 300-pixel units was appropriate to be used as the input value. However, in some cases, a minimum region of 400-700-pixel units was appropriate.
Based on the results of the segmentation analysis, aquatic vegetation was manually classified for all segmented images in QGIS to quantify spatial variability in vegetation dynamics. The main reach of Camden Creek was discretized into ten sub-reaches (fig. 1g). For each reach, the segmented areas of duckweed were summed for each monitoring period. Discretization was performed based on in-stream structural controls (e.g., weirs) located within the stream reach, and surface water inputs including springs, a spring-fed pond, and a large tributary that occurred along the monitored stream reach. To calculate the percent coverage of each reach, the total area of duckweed coverage was divided by the total reach surface area.
Given the workload associated with manual classification, we evaluated the utility of supervised classification to predict vegetation on segmented images for a subsection of the stream reach collected on 28 July 2020. First, the segmented features were classified into six different categories, and an identification number was assigned to each. The number zero was assigned to riparian vegetation, the number one to water, the number two to duckweed, the number three to algae, the number four to macrophytes, and the number five to the white space surrounding the clipped image. Riparian vegetation was present in many of the clips, due to the difficulty of isolating it from the stream boundaries, without clipping portions of the stream. Therefore, riparian vegetation was included as one of the classes in the classification process.
Using a combination of visual interpretation from the high-resolution RGB images and qualitative verification from repeated field surveys and observations on sampling dates, the features were identified and manually assigned for the training and validation datasets. Qualitative evaluation involved ground truth surveys that were performed on a 1 km stretch to verify the remote sensing results. The stream reach was selected based on previous modeling in the reach which suggested high rates of duckweed and algal growth, as well as high temporal variability (Bunnell et al., 2020). Homogenous vegetation pools for duckweed, algae, and macrophytes existed within the segment, which simplified the identification of vegetation species in the aerial images. The segments were surveyed visually by wading the stream reach. Different groups of aquatic vegetation were noted, and the regions where specific vegetation was predominant were identified. A handheld GPS was used during the surveys to note the coordinates of the specific regions.
Out of a total of ten clips that were selected for model evaluation, 666 segments were used for training, and the remaining 498 segments were used for validation, which is consistent with approaches used in similar studies (Husson et al., 2016; Chabot et al., 2018; De Luca et al., 2019). Training and validation of the model was performed using the OTB toolbox. While numerous algorithms are available in OTB, support vector machines (SVM) have generally yielded superior results for classifying aquatic vegetation (Mills, 2008; Lou et al., 2014; Azhar, 2015), and this was therefore the algorithm that was used for this study. The output of the training analysis was a confusion matrix (or error matrix), which is a table used to describe the performance of the selected classification model (Hardin and Shumway, 1997; Jeness and Wynne, 2005; Huang et al., 2017). Results from the confusion matrix were used to calculate goodness of fit indicators including overall accuracy (OvAc), producer’s accuracy (PrAc) for errors of commission, user’s accuracy (UsAc) for errors of omission, and the kappa coefficient. The kappa coefficient, which provides a measure of how the classification results compare to values assigned by chance, equals 0 if there is no agreement between the classified image and the reference image, and scales up to 1 if the classified image and the ground truth image are without error.
In Situ Single-Station Estimation of Metabolism
A YSI EXO2 in situ sonde was deployed at the watershed outlet of Camden Creek over a three-year monitoring period (September 2018-August 2021) and encompassed the UAS data collection period. Flowrate, dissolved oxygen, specific conductance, and water temperature were measured. Flowrates were estimated based on flow depth measurements over a v-notch weir located at the watershed outlet (Ford et al., 2019). Dissolved oxygen (mg/L) and percent saturation were measured using an optical dissolved oxygen sensor. Specific conductance (µS/cm) and temperature (°C) were measured using an EXO conductivity and temperature sensor. For all parameters, data was collected continuously at a 15-minute interval, with some periodic gaps (McGill and Ford, 2024). Data was screened to check if points fell within maximum and minimum threshold values for the sensors per manufacturer specifications and were flagged if they fell outside of bounds. Datasets were manually checked for any abnormalities or outliers and cross-referenced with field notes. All sensors were generally calibrated and maintained per manufacturer specifications using manufacturer standards every four to five weeks.
Stream metabolism was estimated from diel sensor data daily using a single-station estimation approach throughout the monitoring period. Gross primary production (GPP) and ecosystem respiration (ER) were calculated at a daily timestep using the Bayesian Single-station Estimation (BASE) method detailed in Grace et al. (2015). BASE v2.3 was downloaded from github (https://github.com/dgiling/BASE) and compiled in R version 3.6. Input data for the model was incorporated at a 15-minute interval and included photosynthetically active radiation (µmol/m2/s), which was estimated from a pyranometer measuring total solar radiation at the National Centers of Environmental Information’s Climate Reference Network in Versailles, Kentucky, and atmospheric pressure data obtained from the Bluegrass airport. Additionally, measurements of water temperature (Celsius), dissolved oxygen (mg/L), and salinity (ppt) at the watershed outlet, obtained using YSI EXO2 sensors, were implemented as inputs to the model. Posterior solutions for the model were screened to ensure convergence of parameters, predictive p-values were between 0.1 and 0.9, and the effective number of parameters was positive. Only one of the 979 monitored days did not meet these criteria, and it was excluded from the results.
Diurnal Variability in Nitrate
Continuous nitrate/nitrite measurements at the watershed outlet were obtained using a SUNA V2 sensor manufactured by Sea-Bird Scientific. The values of nitrate-nitrite were reported in milligrams of nitrogen per liter, mgN/L. Data was collected from September 2018-August 2021 at 15-minute intervals with some period gaps (McGill and Ford, 2024). Reference spectrum updates were generally carried out every four to five weeks per manufacturer specifications. The sensor was checked for debris or other possible fouling during weekly site visits, although fouling was never noteworthy. Quality control analysis included screening an RMSE value that estimates the goodness of the nitrate spectral fit. The SUNA values were noted and flagged if this RMSE value was reported as 0.001 or higher. Grab samples were used to validate SUNA measurements and ensure the flagged values were accurate. At least two samples were collected each month over a broader monitoring period (August 2018-August 2021), which reflected seasonal and event dynamics and spanned the range of concentrations detected at the watershed outlet. Samples were analyzed by the Kentucky Geological Survey (KGS) for nitrate-nitrogen (NO3-N). KGS follows EPA Method 9056A-Determination of Inorganic Anions by Ion Chromatography when determining the nitrate concentration of surface water samples (EPA, 2007). Linear regression of lab nitrate measurements vs. SUNA estimates (n = 80) yielded R-squared values of 0.98 and a slope of 0.96, indicating the strength of the optical measurement (McGill and Ford, 2024).
Nitrate concentration results were analyzed for diurnal variability to determine the impact of aquatic vegetation pools on N removal. First, days in which storm events occurred within 48 hours of the end of the day were excluded from analysis since concentrations may be more strongly impacted by hydrological processes as compared to in-stream biochemical processes (McGill and Ford, 2024). Differences in daily maxima and minima in nitrate concentrations, the timing of daily maxima, and the timing of daily minima were analyzed for each day in the filtered dataset.
We analyzed concentration vs. dissolved oxygen percent saturation (DO%sat) relationships for a subset of the data, representing periods of greatest GPP, transitional periods with decreasing GPP and ER but increasing duckweed coverage, and periods with low GPP but high duckweed coverage. In total, 18 days were selected for analysis. These periods were found to be characteristic of the aquatic vegetation growing season. For the selected periods (8-23-20 to 8-29-20; 9-22-20 to 9-28-20; and 4-15-21 to 4-21-21), time series of nitrate concentrations and scatter plots of concentration vs. DO%sat were generated. Linear regression analysis was performed for each day in excel and slopes, intercept, and coefficients of determination (R-squared values) were reported.
Results from a previously published, data-driven AI modeling of nitrate concentrations were compared with the measured data to assess the utility of the novel datasets to inform model performance. While the model formulation is described in detail elsewhere (McGill and Ford, 2024), briefly, we utilized a two-layer extreme learning machine (TELM) model that simulates nitrate concentrations as a function of meteorological and soil variables (moisture and temperature from 10-100 cm). The model simulates concentrations at 15-minutes from 29 July 2020 through 4 August 2021. Global model evaluation showed strong agreement between measured and modeled nitrate concentrations, with a Nash Sutcliffe Efficiency of 0.94. Likewise, the model was found to accurately predict concentration discharge relationships during storm events for a broad spectrum of hydrologic conditions through performing quantitative and qualitative hysteresis analyses (McGill and Ford, 2024). To evaluate the ability of the model to capture N removal dynamics under low flow conditions, we analyzed both visual fit and concentration-dissolved oxygen relationships for the same 18 days analyzed for the measured data, including periods with prominence of algal biomass and duckweed, as well as transitional periods.
Results and Discussions
Aquatic Vegetation Detection Using UAS
Collected aerial imagery displayed temporal variability in aquatic vegetation pools throughout the monitoring period, as observed in a magnified section of Camden Creek (fig. 2). We observed the prominence of benthic algal biomass at the start of the monitoring period (figs. 2a and 2b) but decreased biomass in subsequent monitoring periods (e.g., fig. 2e). In July through mid-October, duckweed was prominent in the channel, although abundance varied spatially and temporally throughout the reach. Some locations in the stream reach had duckweed throughout all summer months (e.g., fig. 2), while other sites only had duckweed periodically throughout the summer or early fall. Following storm events in late October, aquatic biomass in the stream channel was observed to decrease substantially.
The automated segmentation procedure was able to distinguish riparian vegetation, duckweed, and macrophytes, but not algae (figs. 3-4). Box and whisker plots of visible band values for the selected classes show that the magnitude of the spectral signatures of duckweed were greater than all other classes and the magnitude of the spectral signatures of riparian vegetation were less than the RGB bands of other classes (fig. 4). As a result, the automated segmentation procedure was able to clearly separate areas with duckweed and riparian vegetation (fig. 3a). The spectral signature of the algae and macrophytes had broadly overlapping ranges that fell between duckweed and riparian endmembers. In addition, the spectral signature of the water overlapped with the spectral signature of algae, which is a likely cause for the lack of segmentation observed between these two classes in figure 3b. An overlap between the ranges of the bands of the water and of the bands of macrophytes is also evident, but it is not as prominent as the overlap between the water and the algae, and the automated segmentation procedure generally did a good job of separating macrophytes (fig. 3c).
Analysis of the successfully segmented images for duckweed displayed spatial and temporal variability in duckweed coverage throughout the monitoring period based on manual classification (table 2). On average, areal coverage was greatest in mid-August through mid-October. However, maximum values varied between the stream reaches. Downstream reaches (e.g., 7, 8, and 9) generally had maximum coverage in late August, whereas upstream reaches (e.g., 1-4, and 6) had maximum values in late September. Duckweed coverage was most prominent (average = 29.7%; range = 0-87.5%) in reach 6, which was downstream of both a nitrogen rich spring (SP1) and a large spring-fed pond with prominence of duckweed throughout the growing season (based on observations). Coverage in reach 6 was more than four-fold greater, on average, than in any other reach of the stream channel. The findings underscore the extensive spatial and temporal variability in duckweed biomass throughout the main stem of Camden Creek.
The trained SVM classifier performed well overall during training and validation, as evidenced by the statistical metrics (table 1), although algae and macrophytes were not well predicted. The overall accuracy achieved was OvAc = 83.5% for training and OvAc = 83.7% for validation. Similarly, the kappa index of the training dataset was equal to 0.74 for training and 0.77 for validation, which illustrates a fairly accurate classification in both cases, based on previously reported values (Chabot and Bird, 2013). Duckweed was generally predicted well by the classifier and had low errors of commission (PrAc = 94% in validation), but there were some errors of omission (UsAc = 81% in validation) in which duckweed was mistaken as water and macrophytes. Riparian vegetation also had low errors of commission and omission. Conversely, the omission and commission errors were high for both algae and macrophytes, as evidenced by UA values of 10%-61% during validation and PA values of 20%-64% during validation. Only one super pixel was correctly classified from the ten super pixels for training algae.
 |
| Figure 2. Visible clips of a magnified area of Camden Creek for (a) 16 June 2020, (b) 1 July 2020, (c) 28 July 2020, (d) 17 August 2020, (e) 31 August 2020, (f) 14 September 2020, (g) 28 September 2020, (h) 15 October 2020. In each image the observations of both algae and duckweed are noted. |
Interferences alter the spectral characteristics of vegetation, which likely contributed to overlapping spectral signatures and poor classification of submerged aquatic vegetationfor the shallow bedrock channel. Water strongly absorbs the electromagnetic radiation in the optical spectral region, resulting in significant dampening of the radiometric signal. Because of this, reflectance measurements for submerged species are usually very low (e.g., Dierssen and Zimmerman, 2003). As can be seen from the results in figure 4, the water spectral signals were similar to macrophytes and algae for the RGB bands. Further, other factors may influence the spectral signatures of vegetation, including the presence of optically active material such as sediment and organic matter (Kirk, 1994; Han and Rundquist, 2003), bottom reflectance and water depth (Ackleson and Klemas, 1987; Gagnon et al., 2008; Klemas, 2013), and the presence of epibiont organisms, especially epiphytes, which can cover the plant surface (Armstrong, 1993; Fyfe, 2003). Collectively, the results of the classification analysis highlight the complexities of detecting submerged vegetation in shallow-water karst agricultural streams.
 |
| Figure 3. Segmentation of images from the 28 July 2020 dataset. The segmented images are grouped into (a) clips that include mostly duckweed, (b) clips containing algae and bedrock, and (c) clips containing mostly rooted macrophytes. |
 |
| Figure 4. Spectral signature of the training and validation segments acquired from the ten clips (five for training/calibration and five for validation) of the visible training dataset on 28 July 2020. Each spectral signature is comprised of three 8-bit intensity values corresponding to the red, green, and blue bands. |
Stream Metabolism Results
High frequency in situ data displayed temporal variability in flow conditions, dissolved oxygen saturation, and temperature at the watershed outlet throughout the data collection period (figs. 5a-c). Flowrates were generally greatest during the winter, least during the summer, and transitional in the spring and fall (fig. 5a). Diel variability of dissolved oxygen percent saturation (DO%sat) varied in magnitude throughout the year but was generally greatest during spring and early summer, then decreased in late summer during the lowest flow conditions and greatest water temperatures (figs. 5a-c). Percent saturation was close to 100% during the cold winter periods, indicating limited in-stream biological activity. These patterns were consistent from year to year during our three-year monitoring period. Dissolved oxygen saturation was also observed to be sensitive to storm events. For example, in early June 2021, a high flow event with peak flowrates exceeding 1 cms substantially reduced diel variability in DO%sat on subsequent days.
| Table 1. Percentage of stream reach area and cumulative stream area covered with duckweed during the ten sampling campaigns in 2020. Bold values represent the date of maximum duckweed coverage for the specified reach. | ||||||||||||||||||||
| Percentage of Reach Area Covered in Duckweed | ||||||||||||||||||||
| Reach | 16-Jun | 1-Jul | 17-Jul | 28-Jul | 17-Aug | 31-Aug | 14-Sep | 28-Sep | 15-Oct | 27-Oct | Average | |||||||||
| 1 | 0.00 | 0.00 | 0.40 | 1.39 | 10.43 | 10.33 | 5.63 | 12.23 | 12.55 | 12.44 | 6.54 | |||||||||
| 2 | 0.00 | 0.00 | 0.00 | 0.00 | 2.93 | 2.73 | 3.12 | 9.46 | 2.27 | 1.28 | 2.18 | |||||||||
| 3 | 0.08 | 0.00 | 0.00 | 0.56 | 2.00 | 5.58 | 3.73 | 10.32 | 4.30 | 2.54 | 2.91 | |||||||||
| 4 | 0.06 | 0.00 | 0.64 | 2.04 | 12.71 | 12.24 | 10.15 | 17.36 | 11.65 | 3.69 | 7.05 | |||||||||
| 5 | 0.00 | 0.47 | 3.77 | 7.86 | 1.65 | 2.40 | 0.64 | 3.20 | 0.75 | 0.15 | 2.09 | |||||||||
| 6 | 0.00 | 0.48 | 74.12 | 6.28 | 22.02 | 31.31 | 30.51 | 87.46 | 32.13 | 12.49 | 29.68 | |||||||||
| 7 | 0.00 | 0.12 | 4.99 | 1.28 | 6.42 | 7.82 | 3.41 | 7.39 | 5.19 | 3.77 | 4.04 | |||||||||
| 8 | 0.00 | 0.17 | 0.86 | 0.28 | 3.21 | 6.45 | 5.44 | 1.51 | 1.60 | 0.21 | 1.97 | |||||||||
| 9 | 0.00 | 0.33 | 0.00 | 0.65 | 7.56 | 12.77 | 7.25 | 2.38 | 10.30 | 1.66 | 4.29 | |||||||||
| 10 | 0.23 | 0.50 | 0.00 | 0.23 | 1.43 | 3.81 | 2.83 | 2.68 | 9.46 | 4.02 | 2.52 | |||||||||
| Total | 0.05 | 0.25 | 4.24 | 2.36 | 5.37 | 7.46 | 5.13 | 9.12 | 6.77 | 2.90 | 4.37 | |||||||||
Table 2. The confusion matrix for the support vector machines (SVM) classification algorithm for the visible dataset. User’s accuracy (UA) and Producers accuracy (PA) are provided. Each row represents the reference classes and each column represents the produced classes. The classes are denoted as follows: 0 = riparian vegetation; 1 = water; 2 = duckweed; 3 = algae; 4 = macrophytes, 5 = white space. | ||||||||||||||||||||
| Produced Classes | Reference Classes | |||||||||||||||||||
| 0 | 1 | 2 | 3 | 4 | 5 | Total | UA | |||||||||||||
| Training | 0 | 329 | 6 | 1 | 2 | 15 | 0 | 353 | 93.20% | |||||||||||
| 1 | 5 | 80 | 0 | 0 | 4 | 0 | 89 | 89.89% | ||||||||||||
| 2 | 4 | 0 | 24 | 0 | 4 | 0 | 32 | 75.00% | ||||||||||||
| 3 | 22 | 3 | 0 | 7 | 11 | 0 | 43 | 16.28% | ||||||||||||
| 4 | 28 | 2 | 1 | 1 | 99 | 0 | 131 | 75.57% | ||||||||||||
| 5 | 1 | 0 | 0 | 0 | 0 | 17 | 18 | 100.00% | ||||||||||||
| Total | 389 | 91 | 26 | 10 | 133 | 17 | ||||||||||||||
| PA | 84.58% | 87.91% | 92.31% | 70.00% | 74.44% | 100.00% | ||||||||||||||
| Validation | 0 | 184 | 0 | 1 | 1 | 5 | 3 | 194 | 94.85% | |||||||||||
| 1 | 12 | 77 | 6 | 0 | 1 | 0 | 96 | 80.21% | ||||||||||||
| 2 | 0 | 15 | 109 | 0 | 9 | 1 | 134 | 81.34% | ||||||||||||
| 3 | 8 | 0 | 0 | 1 | 1 | 0 | 10 | 10.00% | ||||||||||||
| 4 | 9 | 6 | 0 | 3 | 28 | 0 | 46 | 60.87% | ||||||||||||
| 5 | 0 | 0 | 0 | 0 | 0 | 18 | 18 | 100.00% | ||||||||||||
| Total | 213 | 98 | 116 | 5 | 44 | 22 | ||||||||||||||
| PA | 86.38% | 78.57% | 93.97% | 20.00% | 63.64% | 81.82% | ||||||||||||||
Patterns in gross primary production (GPP), ecosystem respiration (ER), and net primary production (NPP) were also consistent from year to year (fig. 6). GPP began increasing in late February/early March and continued to rise through the spring season, coinciding with increasing temperatures and decreasing baseflow conditions (figs. 5a and 5b). GPP was observed to be greatest from late spring to early summer (May-June), aligning with our observations of dense algal biomass coverage from our aerial UAS results in June and early July (fig. 2; fig. 6). ER generally lagged behind GPP, which is reflective of dynamics observed in other stream metabolism studies (e.g., Yang et al., 2023) and suggests a lag in the activity of heterotrophic bacteria. As a result, positive net primary production (NPP) generally occurred in April through May, and negative NPP occurred throughout the remainder of the year. Interestingly, GPP and ER were negligibly impacted by floating aquatic macrophytes, as evidenced by the steady decline of both GPP and ER during the July-October period in which the floating aquatic macrophytes were progressively increasing (table 1).
Aquatic Vegetation Dynamics in Karst Agroecosystem Streams
Coupling the high frequency in situ and aerial image analysis allowed us to capture distinct shifts in aquatic vegetation throughout the year. During late winter through spring, algal biomass increased, and peaked in early summer. In mid-summer, duckweed growth became prominent resulting in the shading of algal biomass. Gross primary production sharply declined throughout the summer, despite otherwise favorable environmental conditions for algae growth. Later, in the fall, storm events scoured and transported aquatic vegetation from the stream reach. A dormant period for stream metabolism was observed following this during the late fall and winter due to low sediment storage in the watershed, high baseflow conditions, and cool stream temperatures. The findings illustrate the importance of aquatic vegetation in driving in-stream metabolic processing and water quality dynamics throughout the year. Taken individually, stream metabolism results were unable to capture floating aquatic macrophyte dynamics, which have previously been highlighted as an important substrate for nutrient removal in the study watershed (Bunnell et al., 2020). Others have highlighted the limitations of accounting for macrophyte impacts using stream metabolism single station methods (e.g., Yang et al., 2023). In our study, this likely stems from floating aquatic macrophytes, at least partially, exchanging oxygen directly with the atmosphere (Pokorny and Rejmankova, 1983). Conversely, UAS imagery was unable to detect and differentiate algae from the water column, particularly when duckweed cover was abundant. As a result, we would have limited information regarding algal biomass growth and subsequent decomposition dynamics without single station estimates. Collectively, our results highlight the importance of coupling these methodologies for detecting timing and spatial variations of in-stream aquatic vegetation dynamics.
 |
| Figure 5. Time series of 15-minute frequency data for (a) flowrate, (b) water temperature, (c) dissolved oxygen saturation (%), and (d) SUNA V2 nitrate-N concentrations from 1 September 2018 to 1 August 2021. |
 |
| Figure 6. Estimates of daily gross primary production (GPP), ecosystem respiration (ER), and net ecosystem productivity (NPP) from the Bayesian Single-station Estimation (BASE) model. Average and standard deviation of MCMC results are presented for GPP and ER, and averages of NPP are provided. |
Stream metabolism results contrast findings from other systems with the prominence of algae, which reflects the use of in-stream structures within Camden Creek that promote duckweed accumulation and proliferation. In studies where benthic algae and phytoplankton are the primary forms of aquatic vegetation, studies have found that primary productivity gradually increases during the spring season, consistent with our results, but GPP and ER remain elevated throughout the summer months (June-September), contrasting the results of our study (Griffiths et al., 2013; Yang et al., 2023). Similarly, findings from a 5-year modeling study in a nearby Central KY agricultural watershed with prominence of benthic algal biomass found that benthic algae growth was consistently high in June-September (Ford and Fox, 2014; Ford et al., 2019). Weirs located throughout the study reach in Camden Creek were installed during previous studies to measure flowrates throughout the watershed. These weirs decrease flow velocities and increase storage, particularly during low-flow conditions. As a result of greater residence times, duckweed populations can establish in the stream reach and rapidly uptake nutrients. The proliferation of duckweed along the surface of the channel subsequently blocks sunlight from penetrating to the bed. Gross primary production in July-September therefore likely decreased because duckweed receives oxygen from both the water column and atmosphere (Pokorny and Rejmankova, 1983), and submerged vegetation biomass decreased from shading. These findings have implications for in-stream management practices that enhance storage and residence times within headwater systems.
In-stream management practices promote favorable conditions for floating aquatic macrophyte prevalence, which suggests that coupled algal-duckweed dynamics will be an important consideration when quantifying ecosystem services in restored channels. Stream restoration practices increase floodplain connectivity and enhance residence times in streams (Griffiths et al., 2012). As a result, studies have found increasing prominence of floating aquatic macrophytes in restored reaches that did not exist prior to restoration (e.g., Lorenz et al., 2012). Our findings may be particularly applicable to streams and headwater systems implementing techniques such as regenerative stormwater conveyance (e.g., Cizek et al., 2018). Given the sensitivity of aquatic vegetation to stormflows, it will be important to consider the coupled hydraulic and biochemical processes in design considerations to optimize ecosystems services. Our results demonstrate that coupling the in situ and UAS sensors is a useful tool for detecting the spatial and temporal variability of aquatic vegetation in these landscapes. The impact of these dynamics on water quality considerations is noteworthy and is discussed in the following section.
| Table 3. Daily average ± standard deviation of diurnal variability in nitrate concentration, as well as time of daily maxima and time of daily minima. Time of minima and maxima are scaled from 0-1 with 0 representing 12 a.m. of the current day and 1 representing 12 a.m. of the following day. | |||
| Month | Diurnal Variability (mgN/L) | Time of Maxima (fraction of day) | Time of Minima (fraction of day) |
| January | 0.15 ± 0.28 | 0.50 ± 0.42 | 0.44 ± 0.32 |
| February | 0.10 ± 0.04 | 0.47 ± 0.39 | 0.44 ± 0.39 |
| March | 0.19 ± 0.06 | 0.31 ± 0.40 | 0.60 ± 0.16 |
| April | 0.25 ± 0.09 | 0.38 ± 0.35 | 0.65 ± 0.17 |
| May | 0.23 ± 0.08 | 0.39 ± 0.33 | 0.63 ± 0.30 |
| June | 0.26 ± 0.10 | 0.23 ± 0.31 | 0.68 ± 0.29 |
| July | 0.24 ± 0.07 | 0.26 ± 0.14 | 0.77 ± 0.25 |
| August | 0.33 ± 0.08 | 0.34 ± 0.11 | 0.80 ± 0.28 |
| September | 0.25 ± 0.11 | 0.44 ± 0.20 | 0.74 ± 0.34 |
| October | 0.26 ± 0.12 | 0.48 ± 0.34 | 0.61 ± 0.39 |
| November | 0.21 ± 0.12 | 0.32 ± 0.34 | 0.73 ± 0.28 |
| December | 0.08 ± 0.04 | 0.59 ± 0.37 | 0.46 ± 0.40 |
Implications for Nitrate Fate and Transport
Nitrate concentrations in the karst agroecosystem stream were observed to vary on event to seasonal timescales (fig. 5d). Seasonality was prominent in the data, with average seasonal concentrations being greatest in the winter, least in the summer, and transitional in the spring and fall. Storm events also had prominent impacts, shifting nitrate concentration by as much as 7 mgN/L over a short period of time (see October-November 2020 in fig. 5d). Previous work in the watershed has demonstrated that variable hydrologic connectivity to the vadose zone during events is a major driver of both event-scale and seasonal variability in nitrate concentration dynamics (McGill and Ford, 2024). While modeling results have illustrated that in-stream processes have a limited impact on nutrient concentrations during stormflows (Bunnell et al., 2020), the seasonality of nitrate is also strongly impacted by in-stream processes (Ford et al., 2019). Previous results from time-series analysis of a 10-year grab-sampling campaign illustrated that local annual minima in springs in the watershed occurred in late spring, while local minima at downstream stream locations occurred in late summer, similar to the results of our high-frequency data at the watershed outlet (Ford et al., 2019). The findings highlight the importance of in-stream removal during the growing season to impact nitrate concentrations and loading from karst agroecosystems.
Diurnal variability in nitrate was prominent and variable throughout the growing season (table 3; fig. 7). When excluding days impacted by storm events, results for March through November showed an average daily diurnal variation of 0.25 mgN/L, while December through February had an average daily diurnal variation of 0.11 mgN/L. The daily timing of maximum and minimum dynamics changed throughout the year. During the winter, average daily maxima, and minima both occurred around 12 p.m., suggesting an absence of recurring diurnal trends (which was also visually validated through inspection of the time series), suggesting in-stream processes were not the primary factor driving variability. This finding is supported by the low stream metabolism (both ER and GPP) observed during the winter months (fig. 6). During the growing season, maxima generally occurred between 6 a.m.-2 p.m., and minima occurred between 2 p.m.-12 a.m., but varied throughout the growing season. Periods with high rates of GPP (March-early July) generally had maxima that occurred early in the morning, and minima that occurred in the late-afternoon, or early evening (see fig. 7a). Later in July and August, when GPP was decreasing, and duckweed coverage began increasing throughout the stream, a shift was observed in diurnal patterns of nitrate concentrations (fig. 7b). Maxima generally still occurred early in the day (similar to the high GPP period), however, minima occurred in the evening, often close to 12 a.m. (see fig. 7b). Later in September, when GPP was low, and duckweed was at its maximum for the monitoring period, we found that diurnal patterns in nitrate concentrations shifted again, with maxima occurring after 12 p.m., and minima occurring around midnight (similar to the transitional period in July/August).
 |
| Figure 7. Timeseries of measured nitrate-N concentrations and modeled results from the two-layer extreme learning machine model for (a) periods representing high rates of gross primary production (GPP), (b) transitional periods where duckweed is growing and GPP and ecosystem respiration (ER) are decreasing, (c) and periods of low GPP and ER with greater duckweed biomass. |
Regression results between dissolved oxygen and nitrate revealed distinct differences throughout the growing season (table 4; Supplemental figs. S1-S3). We found weak negative linear relationships between DO%sat and nitrate during periods of greatest GPP (fig. S1). Processes including uptake by aquatic vegetation and microorganisms, denitrification, and regeneration can impact diurnal nitrate concentrations, and likely resulted in the low R2 values. However, studies have demonstrated that periods of greatest GPP are negatively related with nitrate concentrations (Heffernan and Cohen, 2010; Yang et al., 2019), highlighting the relative importance of algal uptake in influencing nitrate concentrations during this timeframe. Later in summer, as GPP decreased, and duckweed began to become more prominent, the regression between dissolved oxygen and nitrate shifted and had a near-zero slope. Further, the relationship changed again during the September period to a strong positive linear relationship between nitrate and DO%sat. This finding points to the progressive decoupling of primary production and nitrate uptake by autotrophs. This decoupling of N removal and uptake in the late summer/early fall likely reflects denitrification, given the greatest removal shifted to occur under depleted oxygen conditions. Recent studies have suggested that nocturnal rates of denitrification can be 3-fold higher at night than day due to linkages with DO and redox fluctuations (Zhang et al., 2023). The reason for denitrification to become increasingly prominent when duckweed is more prominent reflects differences in the algal and duckweed substrates to support denitrifying bacterial communities (Körner and Vermaat, 1998; Bunnell et al., 2020). Cumulatively, these results highlight that shifting aquatic vegetation dynamics have significant impacts on the removal mechanisms of nitrate in karst agroecosystem streams.
| Table 4. Summary of linear regression results between nitrate concentration (both measured and modeled) and dissolved oxygen (% saturation). Results are provided for periods representing high rates of gross primary production (GPP) (4-15-21 to 4-20-21), transitional periods where duckweed is growing and GPP and ecosystem respiration (ER) are decreasing (8-23-20 to 8-28-20), and periods of low GPP and ER with greater duckweed coverage (9-22-20 to 9-27-20). Values provided are averages ± one standard deviation. Graphs and regression results are provided in Supplemental figures S1-S3. | ||||||||
| Date | Measured Nitrate | Modeled Nitrate | ||||||
| Slope (10-3) | Intercept | R2 | Slope (10-3) | Intercept | R2 | |||
| 15-4-21 to 20-4-21 (n = 6) | -1.4 ± 0.3 | 2.9 ± 0.1 | 0.2 ± 0.1 | -0.8 ± 0.9 | 2.8 ± 0.2 | 0.2 ± 0.1 | ||
| 23-8-20 to 28-8-20 (n = 6) | -0.02 ± 2.7 | 1.7 ± 0.2 | 0.1 ± 0.2 | 0.9 ± 1.0 | 1.4 ± 0.1 | 0.2 ± 0.2 | ||
| 22-9-20 to 27-9-20 (n = 6) | 4.8 ± 1.2 | 0.9 ± 0.1 | 0.7 ± 0.1 | 2.9 ± 0.9 | 1.0 ± 0.1 | 0.5 ± 0.1 | ||
Comparing results to a previously developed TELM model of nitrate concentrations for the watershed showed diurnal patterns were well represented when algal biomass and duckweed were prominent but did not perform well during transitional periods in July/August (table 4; fig. 7; Supplemental figs. S1-S3). During the April monitoring period, when algal growth was prominent, slopes and intercepts of NO3-DO%sat regressions were comparable, R2 values were low for both modeled and measured predictions, and the timing of maxima and minima was well represented. During transition periods where duckweed was increasing in August, slopes of the regression were generally well represented by the model; however, intercepts and R2 values were poorly reflected, and we found that maxima were not well-represented in the model, as maxima often occurred prior to 12 p.m. and the model consistently predicted maxima after 12 pm. During the late-September period, when duckweed was at its maximum coverage and had been established for several weeks, we found the model again predicted maxima and minima well with good agreement between measured and model regression slopes, intercepts, and R2 values. These results demonstrate the limitations of the TELM model in capturing nitrate removal during transitional periods from algal coverage to predominantly duckweed coverage. While the exact mechanisms of why the model does not capture dynamics is outside the scope of this study, our results point to the need for further model refinement and improvement during these periods.
Results of the concentration-DO analysis suggest that the utilization of hysteresis indices for environmental drivers (e.g., dissolved oxygen) may be useful for evaluating the model performance of in-stream processes in future studies. R2 values from the linear regressions were highly sensitive to the magnitude of hysteresis observed in the nitrate-DO plots (figs. S1-S3). Hysteresis analyses are now commonly used for concentration-discharge relationships during storm events when flowrate is the primary driver of nitrate concentration variability (Vaughn et al., 2017; Liu et al., 2021). During low flow conditions, in-stream processes govern N variability and are heavily influenced by dissolved oxygen, given that DO is simultaneously an indicator of biological uptake and respiration, but also significantly influences denitrification rates. Like hydrographs, daily dissolved oxygen curves have distinct rising and falling limbs and therefore are well suited for hysteresis analyses and currently adopted hysteresis metrics (e.g., hysteresis index). Hysteretic patterns on environmental variables such as dissolved oxygen have not been reported at diel time scales to the author’s knowledge, although parameters such as temperature have been used to analyze hysteretic patterns of nutrient concentration dynamics on annual timescales (e.g., Aubert et al., 2013). While beyond the scope of this study, generalized frameworks are needed (analogous to concentration-discharge hysteresis) to infer prominent N removal and regeneration dynamics from concentration-dissolved oxygen relationships as well as concentration-temperature relationships. Thereafter, applying quantitative techniques used in storm event analyses such as hysteresis indices (Lloyd et al., 2016) could be a valuable tool in multi-objective model evaluation procedures for calibrating models and identifying potential needs for model improvements, such as our illustration with the TELM model case study for the Camden Creek watershed.
Conclusion
We found visible spectrum imagery from UAS captured spatial and temporal distribution of floating aquatic macrophytes, while high frequency in situ sensors were complementary in quantifying stream metabolism for the karst agroecosystem stream, although further work is needed to assess the ability of aerial images to quantify biomass dynamics. Aerial image results for duckweed showed extensive spatial and temporal variability in duckweed coverage in Camden Creek, highlighting the need for improved models and spatially explicit input datasets that can adequately represent aquatic vegetation dynamics in decision support tools. Gross primary production was governed by submerged vegetation and displayed a shortened season as compared to other agricultural streams in the region due to shading and competition from duckweed. Diurnal nitrate concentration dynamics shifted because of changes in aquatic vegetation, and the results of the study suggest that diurnal nitrate-DO relationships are useful for informing model evaluation and elucidating areas of model deficiency. Further work is needed to expand such approaches to comprehensive watershed modeling tools to improve representation of in-stream processes and reduce uncertainty in model simulations.
Supplemental Material
The supplemental materials mentioned in this article are available for download from the ASABE Figshare repository at: https://doi.org/10.13031/25020554
Acknowledgments
The authors thank the associate editor and four anonymous reviewers for their constructive comments on the manuscript. We gratefully acknowledge funding support for this research from the National Science Foundation (NSF-1632888) and support from the USDA National Institute of Food and Agriculture Multistate Projects S-1069 under accession number 1539070 and S-1089. The authors also thank the Department of Biosystems and Agricultural Engineering for partial funding support for the graduate research assistant. The authors thank Alex Fogle and the graduate and undergraduate research assistants at the University of Kentucky who assisted with sensor installation, maintenance, data collection, and analysis.
References
Ackleson, S. G., & Klemas, V. (1987). Remote sensing of submerged aquatic vegetation in lower Chesapeake Bay: A comparison of Landsat MSS to TM imagery. Remote Sens. Environ., 22(2), 235-248. https://doi.org/10.1016/0034-4257(87)90060-5
Armstrong, R. A. (1993). Remote sensing of submerged vegetation canopies for biomass estimation. Int. J. Remote Sens., 14(3), 621-627. https://doi.org/10.1080/01431169308904363
Aubert, A. H., Gascuel-Odoux, C., & Merot, P. (2013). Annual hysteresis of water quality: A method to analyse the effect of intra- and inter-annual climatic conditions. J. Hydrol., 478, 29-39. https://doi.org/10.1016/j.jhydrol.2012.11.027
Azhar, R., Tuwohingide, D., Kamudi, D., Sarimuddin, & Suciati, N. (2015). Batik image classification using SIFT feature extraction, bag of features and support vector machine. Procedia Comput. Sci., 72, 24-30. https://doi.org/10.1016/j.procs.2015.12.101
Brando, V. E., & Phinn, S. R. (2007). Guest editorial: Coastal aquatic remote sensing applications for environmental monitoring and management. J. Appl. Remote Sens., 1(1), 011599. https://doi.org/10.1117/1.2835115
Briggs, M. A., & Hare, D. K. (2018). Explicit consideration of preferential groundwater discharges as surface water ecosystem control points. Hydrol. Process., 32(15), 2435-2440. https://doi.org/10.1002/hyp.13178
Bunnell, N. L., Ford, W. I., Fogle, A. W., & Taraba, J. (2020). Reach-scale model of aquatic vegetation quantifies N fate in a bedrock-controlled karst agroecosystem stream. Water, 12(9), 2458. https://doi.org/10.3390/w12092458
Burns, D. A., Pellerin, B. A., Miller, M. P., Capel, P. D., Tesoriero, A. J., & Duncan, J. M. (2019). Monitoring the riverine pulse: Applying high-frequency nitrate data to advance integrative understanding of biogeochemical and hydrological processes. WIREs Water, 6(4), e1348. https://doi.org/10.1002/wat2.1348
Chabot, D., & Bird, D. M. (2013). Small unmanned aircraft: Precise and convenient new tools for surveying wetlands. J. Unmanned Veh. Syst., 1(1), 15-24. https://doi.org/10.1139/juvs-2013-0014
Chabot, D., Dillon, C., Shemrock, A., Weissflog, N., & Sager, E. P. (2018). An object-based image analysis workflow for monitoring shallow-water aquatic vegetation in multispectral drone imagery. ISPRS Int. J. Geo-Inf., 7(8), 294. https://doi.org/10.3390/ijgi7080294
Cizek, A. R., Hunt, W. F., Winston, R. J., Waickowski, S. E., Narayanaswamy, K., & Lauffer, M. S. (2018). Water quality and hydrologic performance of a regenerative stormwater conveyance in the Piedmont of North Carolina. J. Environ. Eng., 144(8), 4018062. https://doi.org/10.1061/(ASCE)EE.1943-7870.0001344
De Luca, G., Silva, J. M., Cerasoli, S., Araújo, J., Campos, J., Di Fazio, S., & Modica, G. (2019). Object-based land cover classification of cork oak woodlands using UAV imagery and Orfeo ToolBox. Remote Sens., 11(10), 1238.
Deo, R. C., & Sahin, M. (2016). An extreme learning machine model for the simulation of monthly mean streamflow water level in eastern Queensland. Environ. Monit. Assess., 188(2), 90. https://doi.org/10.1007/s10661-016-5094-9
Dierssen, H. M., Zimmerman, R. C., Leathers, R. A., Downes, T. V., & Davis, C. O. (2003). Ocean color remote sensing of seagrass and bathymetry in the Bahamas Banks by high-resolution airborne imagery. Limnol. Oceanogr., 48(1 part 2), 444-455. https://doi.org/10.4319/lo.2003.48.1_part_2.0444
EPA. (2007). Determination of inorganic anions by ion chromatography. Environmental Protection Agency.
Eriksson, P. G. (2001). Interaction effects of flow velocity and oxygen metabolism on nitrification and denitrification in biofilms on submersed macrophytes. Biogeochemistry, 55(1), 29-44. https://doi.org/10.1023/A:1010679306361
Flynn, K. F., & Chapra, S. C. (2014). Remote sensing of submerged aquatic vegetation in a shallow non-turbid river using an unmanned aerial vehicle. Remote Sens., 6(12), 12815-12836. https://doi.org/10.3390/rs61212815
Ford, W. I., & Fox, J. F. (2014). Model of particulate organic carbon transport in an agriculturally impacted stream. Hydrol. Process., 28(3), 662-675. https://doi.org/10.1002/hyp.9569
Ford, W. I., Fox, J. F., & Pollock, E. (2017). Reducing equifinality using isotopes in a process-based stream nitrogen model highlights the flux of algal nitrogen from agricultural streams. Water Resour. Res., 53(8), 6539-6561. https://doi.org/10.1002/2017WR020607
Ford, W. I., Husic, A., Fogle, A., & Taraba, J. (2019). Long-term assessment of nutrient flow pathway dynamics and in-stream fate in a temperate karst agroecosystem watershed. Hydrol. Process., 33(11), 1610-1628. https://doi.org/10.1002/hyp.13427
Fyfe, S. K. (2003). Spatial and temporal variation in spectral reflectance: Are seagrass species spectrally distinct? Limnol. Oceanogr., 48(1 part 2), 464-479. https://doi.org/10.4319/lo.2003.48.1_part_2.0464
Gagnon, P., Scheibling, R. E., Jones, W., & Tully, D. (2008). The role of digital bathymetry in mapping shallow marine vegetation from hyperspectral image data. Int. J. Remote Sens., 29(3), 879-904. https://doi.org/10.1080/01431160701311283
Grace, M. R., Giling, D. P., Hladyz, S., Caron, V., Thompson, R. M., & Mac Nally, R. (2015). Fast processing of diel oxygen curves: Estimating stream metabolism with BASE (BAyesian Single-station Estimation). Limnol. Oceanogr. Methods, 13(3), e10011. https://doi.org/10.1002/lom3.10011
Griffiths, N. A., Tank, J. L., Roley, S. S., & Stephen, M. L. (2012). Decomposition of maize leaves and grasses in restored agricultural streams. Freshw. Sci., 31(3), 848-864. https://doi.org/10.1899/11-095.1
Griffiths, N. A., Tank, J. L., Royer, T. V., Roley, S. S., Rosi-Marshall, E. J., Whiles, M. R.,... Johnson, L. T. (2013). Agricultural land use alters the seasonality and magnitude of stream metabolism. Limnol. Oceanogr., 58(4), 1513-1529. https://doi.org/10.4319/lo.2013.58.4.1513
Han, L., & Rundquist, D. C. (2003). The spectral responses of Ceratophyllum demersum at varying depths in an experimental tank. Int. J. Remote Sens., 24(4), 859-864. https://doi.org/10.1080/0143116021000009868
Hardin, P. J., & Shumway, J. M. (1997). Statistical significance and normalized confusion matrices. Photogramm. Eng. Remote Sens., 63(6), 735-739.
Heddam, S., & Kisi, O. (2017). Extreme learning machines: A new approach for modeling dissolved oxygen (DO) concentration with and without water quality variables as predictors. Environ. Sci. Pollut. Res., 24(20), 16702-16724.
https://doi.org/10.1007/s11356-017-9283-z
Heffernan, J. B., & Cohen, M. J. (2010). Direct and indirect coupling of primary production and diel nitrate dynamics in a subtropical spring-fed river. Limnol. Oceanogr., 55(2), 677-688. https://doi.org/10.4319/lo.2010.55.2.0677
Huang, D., Xu, S., Sun, J., Liang, S., Song, W., & Wang, Z. (2017). Accuracy assessment model for classification result of remote sensing image based on spatial sampling.. J. Appl. Remote Sens., 11(4), 046023. https://doi.org/10.1117/1.JRS.11.046023
Huang, G., Zhou, H., Ding, X., & Zhang, R. (2012). Extreme learning machine for regression and multiclass classification. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), 42(2), 513-529. https://doi.org/10.1109/TSMCB.2011.2168604
Husson, E., Ecke, F., & Reese, H. (2016). Comparison of manual mapping and automated object-based image analysis of non-submerged aquatic vegetation from very-high-resolution UAS images. Remote Sens., 8(9), 724. https://doi.org/10.3390/rs8090724
Jenness, J., & Wynne, J. J. (2005). Cohen’s Kappa and classification table metrics 2.0 : An ArcView 3x extension for accuracy assessment of spatially explicit models. Open-File Report 2005-1363. Flagstaff, AZ: USGS. https://doi.org/10.3133/ofr20051363
Jia, Q., Cao, L., Yésou, H., Huber, C., & Fox, A. D. (2017). Combating aggressive macrophyte encroachment on a typical Yangtze River lake: Lessons from a long-term remote sensing study of vegetation. Aquatic Ecol., 51(1), 177-189. https://doi.org/10.1007/s10452-016-9609-9
Kazama, S., & Watanabe, K. (2018). Estimation of periphyton dynamics in a temperate catchment using a distributed nutrient-runoff model. Ecol. Modell., 367, 1-9. https://doi.org/10.1016/j.ecolmodel.2017.11.006
Kirk, J. T. (1994). Light and photosynthesis in aquatic ecosystems (2nd ed.). Cambridge, England: Cambridge University Press. https://doi.org/10.1017/CBO9780511623370
Klemas, V. (2013). Remote sensing of emergent and submerged wetlands: An overview. Int. J. Remote Sens., 34(18), 6286-6320. https://doi.org/10.1080/01431161.2013.800656
Klemas, V. V. (2016). Remote sensing of submerged aquatic vegetation. In C. W. Finkl, & C. Makowski (Eds.), Seafloor mapping along continental shelves: Research and techniques for visualizing benthic environments (pp. 125-140). Cham, Switzerland: Springer. https://doi.org/10.1007/978-3-319-25121-9_5
Körner, S., & Vermaat, J. E. (1998). The relative importance of Lemna gibba L., bacteria and algae for the nitrogen and phosphorus removal in duckweed-covered domestic wastewater. Water Res., 32(12), 3651-3661. https://doi.org/10.1016/S0043-1354(98)00166-3
Körner, S., Vermaat, J. E., & Veenstra, S. (2003). The capacity of duckweed to treat wastewater. J. Environ. Qual., 32(5), 1583-1590. https://doi.org/10.2134/jeq2003.1583
Krishnan, S. R., Nallakaruppan, M. K., Chengoden, R., Koppu, S., Iyapparaja, M., Sadhasivam, J., & Sethuraman, S. (2022). Smart water resource management using artificial intelligence - a review. Sustainability, 14(20), 13384. https://doi.org/10.3390/su142013384
Liu, W., Birgand, F., Tian, S., & Chen, C. (2021). Event-scale hysteresis metrics to reveal processes and mechanisms controlling constituent export from watersheds: A review. Water Res., 200, 117254. https://doi.org/10.1016/j.watres.2021.117254
Lloyd, C. E., Freer, J. E., Johnes, P. J., & Collins, A. L. (2016). Using hysteresis analysis of high-resolution water quality monitoring data, including uncertainty, to infer controls on nutrient and sediment transfer in catchments. Sci. Total Environ., 543, 388-404. https://doi.org/10.1016/j.scitotenv.2015.11.028
Lorenz, A. W., Korte, T., Sundermann, A., Januschke, K., & Haase, P. (2012). Macrophytes respond to reach-scale river restorations. J. Appl. Ecol., 49(1), 202-212. https://doi.org/10.1111/j.1365-2664.2011.02082.x
Lou, X.-w., Huang, D.-c., Fan, L.-m., & Xu, A.-j. (2014). An image classification algorithm based on bag of visual words and multi-kernel learning. J. Multimed., 9(2), 269. https://doi.org/10.4304/jmm.9.2.269-277
Malthus, T. J. (2017). Chapter 9 - Bio-optical modeling and remote sensing of aquatic macrophytes. In D. R. Mishra, I. Ogashawara, & A. A. Gitelson (Eds.), Bio-optical modeling and remote sensing of inland waters (pp. 263-308). Elsevier. https://doi.org/10.1016/B978-0-12-804644-9.00009-4
McGill, T., & Ford, W. I. (2024). Extreme learning machine predicts high-frequency stream flow and nitrate-N concentrations in a karst agricultural watershed. J. ASABE, 67(2), 73-87. https://doi.org/10.13031/ja.15747
Mills, H. (2008). Analysis of the transferability of support vector machines for vegetation classification. Proc. Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XXXVII, Part B7, 557-563.
Nelson, C. E., Alldredge, A. L., McCliment, E. A., Amaral-Zettler, L. A., & Carlson, C. A. (2011). Depleted dissolved organic carbon and distinct bacterial communities in the water column of a rapid-flushing coral reef ecosystem. The ISME Journal, 5(8), 1374-1387. https://doi.org/10.1038/ismej.2011.12
O’Hare, M. T., Aguiar, F. C., Asaeda, T., Bakker, E. S., Chambers, P. A., Clayton, J. S.,... Wood, K. A. (2018). Plants in aquatic ecosystems: Current trends and future directions. Hydrobiologia, 812(1), 1-11. https://doi.org/10.1007/s10750-017-3190-7
Odum, H. T. (1956). Primary production in flowing waters. Limnol. Oceanogr., 1(2), 102-117. https://doi.org/10.4319/lo.1956.1.2.0102
Peng, J.-F., Wang, B.-Z., Song, Y.-H., & Yuan, P. (2007a). Modeling N transformation and removal in a duckweed pond: Model application. Ecol. Modell., 206(3), 294-300. https://doi.org/10.1016/j.ecolmodel.2007.03.037
Peng, J.-F., Wang, B.-Z., Song, Y.-H., & Yuan, P. (2007b). Modeling N transformation and removal in a duckweed pond: Model development and calibration. Ecol. Modell., 206(1), 147-152. https://doi.org/10.1016/j.ecolmodel.2007.03.029
Perniel, M., Ruan, R., & Martinez, B. (1998). Nutrient removal from a stormwater detention pond using duckweed. Appl. Eng. Agric., 14(6), 605-609. https://doi.org/10.13031/2013.19429
Pokorný, J., & Rejmánková, E. (1983). Oxygen regime in a fishpond with duckweeds (lemnaceae) and Ceratophyllum. Aquat. Bot., 17(2), 125-137. https://doi.org/10.1016/0304-3770(83)90109-2
Robertson, D. M., & Saad, D. A. (2021). Nitrogen and phosphorus sources and delivery from the Mississippi/Atchafalaya River Basin: An update using 2012 SPARROW Models. JAWRA, 57(3), 406-429. https://doi.org/10.1111/1752-1688.12905
Rode, M., Wade, A. J., Cohen, M. J., Hensley, R. T., Bowes, M. J., Kirchner, J. W.,... Jomaa, S. (2016). Sensors in the stream: The high-frequency wave of the present. Environ. Sci. Technol., 50(19), 10297-10307. https://doi.org/10.1021/acs.est.6b02155
Silva, T. S., Costa, M. P., Melack, J. M., & Novo, E. M. (2008). Remote sensing of aquatic vegetation: Theory and applications. Environ. Monit. Assess., 140(1), 131-145. https://doi.org/10.1007/s10661-007-9855-3
Snyder, L., Potter, J. D., & McDowell, W. H. (2018). An Evaluation of nitrate, fDOM, and turbidity sensors in New Hampshire streams. Water Resour. Res., 54(3), 2466-2479. https://doi.org/10.1002/2017WR020678
Uehlinger, U. (2006). Annual cycle and inter-annual variability of gross primary production and ecosystem respiration in a floodprone river during a 15-year period. Freshw. Biol., 51(5), 938-950. https://doi.org/10.1111/j.1365-2427.2006.01551.x
Vaughan, M. C., Bowden, W. B., Shanley, J. B., Vermilyea, A., Sleeper, R., Gold, A. J.,... Schroth, A. W. (2017). High-frequency dissolved organic carbon and nitrate measurements reveal differences in storm hysteresis and loading in relation to land cover and seasonality. Water Resour. Res., 53(7), 5345-5363. https://doi.org/10.1002/2017WR020491
Visser, F., Buis, K., Verschoren, V., & Schoelynck, J. (2018). Mapping of submerged aquatic vegetation in rivers from very high-resolution image data, using object-based image analysis combined with expert knowledge. Hydrobiologia, 812(1), 157-175. https://doi.org/10.1007/s10750-016-2928-y
Weary, D. J., & Doctor, D. H. (2014). A preview of “Karst in the United States of America: A Digital Map Compilation and Database”. In U.S. Geological Survey Karst Interest Group Proceedings (pp. 19-27). Reston, VA: USGS. https://doi.org/10.3133/ofr20141156 Yang, X., Jomaa, S., Büttner, O., & Rode, M. (2019). Autotrophic nitrate uptake in river networks: A modeling approach using continuous high-frequency data. Water Res., 157, 258-268. https://doi.org/10.1016/j.watres.2019.02.059
Yang, X., Zhang, X., Graeber, D., Hensley, R., Jarvie, H., Lorke, A.,... Rode, M. (2023). Large-stream nitrate retention patterns shift during droughts: Seasonal to sub-daily insights from high-frequency data-model fusion. Water Res., 243, 120347. https://doi.org/10.1016/j.watres.2023.120347
Zhang, W., Li, H., & Li, B. (2023). Whole-system estimation of hourly denitrification in a flow-through riverine wetland. J. Hydrol., 618, 129132. https://doi.org/10.1016/j.jhydrol.2023.129132