Application of UAV Imagery to Detect and Quantify Submerged Filamentous Algae and Rooted Macrophytes in a Non-Wadeable River

Imagery from unoccupied aerial vehicles (UAVs) is useful for mapping floating and emerged primary producers, as well as single taxa of submerged primary producers in shallow, clear lakes and streams. However, there is little research on the effectiveness of UAV imagery-based detection and quantification of submerged filamentous algae and rooted macrophytes in deeper rivers using a standard red-green-blue (RGB) camera. This study provides a novel application of UAV imagery analysis for monitoring a non-wadeable river, the Klamath River in northern California, USA. River depth and solar angle during flight were analyzed to understand their effects on benthic primary producer detection. A supervised, pixel-based Random Trees classifier was utilized as a detection mechanism to estimate the percent cover of submerged filamentous algae and rooted macrophytes from aerial photos within 32 sites along the river in June and July 2019. In-situ surveys conducted via wading and snorkeling were used to validate these data. Overall accuracy was 82% for all sites and the highest overall accuracy of classified UAV images was associated with solar angles between 47.5 and 58.72° (10:04 a.m. to 11:21 a.m.). Benthic algae were detected at depths of 1.9 m underwater and submerged macrophytes were detected down to 1.2 m (river depth) via the UAV imagery in this relatively clear river (Secchi depth > 2 m). Percent cover reached a maximum of 31% for rooted macrophytes and 39% for filamentous algae within all sites. Macrophytes dominated the upstream reaches, while filamentous algae dominated the downstream reaches closer to the Pacific Ocean. In upcoming years, four proposed dam removals are expected to alter the species composition and abundance of benthic filamentous algae and rooted macrophytes, and aerial imagery provides an effective method to monitor these changes.


Introduction
The extent, distribution, and assemblage of primary producers can indicate the ecological health of a water body [1,2]. Like vegetation on land, algae and aquatic plants provide a large portion of the nutritional and structural foundation for aquatic ecosystems [3]. In many rivers and clear lakes, primary producer assemblages are dominated by benthic algae and aquatic plants growing attached to the substrate, where filamentous algae and aquatic plants provide structural support to streambeds, as well as improve water quality by storing and processing nutrients and promoting settlement of suspended particles [4,5]. Diverse assemblages of primary producers, including thin films of diatoms, combined with the contribution of smaller tributaries, result in August base flows of about 81 m³ s −1 near the river mouth (Klamath River near Klamath, USGS Gage 11530500). The Klamath River is a non-wadeable river due to both excessive depth and velocity, even during summer baseflow periods. This study examined benthic primary producer assemblages (submerged rooted macrophytes and filamentous algae) in June and July of 2019, as this is peak primary productivity season in the Klamath River [44,45], a time of lower flows, and when water clarity is sufficient to remotely sense benthic and emergent primary producers [50].
The upper watershed which supplies flows at Iron Gate Dam supports heavy agricultural use, including water withdrawals for irrigation and increased nutrient runoff [51], resulting in lower peak flows and increased nutrient concentrations to the Klamath River. A series of 4 hydroelectric dams located between 61 and 113 km below the Upper Klamath Lake outflow ( Figure A1) alters flow within the river and reservoirs and increases hydrologic stability below the dams [52]. High nutrient concentrations and stable flows support high rates of riverine primary productivity and substantial autotrophic biomass in the Klamath River [44], with peak proliferations typically occurring below Iron Gate Dam in early summer. Standing stock of primary producers includes dense beds of filamentous algae and rooted-aquatic plants, such as Cladophora sp., Potamogeton sp., and Elodea sp. Both rooted aquatic plants and filamentous algae create seasonally dense plant and algal patches [53] that are visible from river margins. Primary producer dominance changes longitudinally below the hydroelectric dams, shifting from rooted aquatic plant dominance to filamentous algae dominance downstream.  The upper watershed which supplies flows at Iron Gate Dam supports heavy agricultural use, including water withdrawals for irrigation and increased nutrient runoff [51], resulting in lower peak flows and increased nutrient concentrations to the Klamath River. A series of 4 hydroelectric dams located between 61 and 113 km below the Upper Klamath Lake outflow ( Figure A1) alters flow within the river and reservoirs and increases hydrologic stability below the dams [52]. High nutrient concentrations and stable flows support high rates of riverine primary productivity and substantial autotrophic biomass in the Klamath River [44], with peak proliferations typically occurring below Iron Gate Dam in early summer. Standing stock of primary producers includes dense beds of filamentous algae and rooted-aquatic plants, such as Cladophora sp., Potamogeton sp., and Elodea sp. Both rooted aquatic plants and filamentous algae create seasonally dense plant and algal patches [53] that are visible Remote Sens. 2020, 12, 3332 5 of 24 from river margins. Primary producer dominance changes longitudinally below the hydroelectric dams, shifting from rooted aquatic plant dominance to filamentous algae dominance downstream.

UAV Data
Aerial imagery was captured to classify submerged filamentous algae and rooted macrophytes and estimate percent cover of each class from 32 sites within ten reaches of the Klamath River ( Figure 1) between Iron Gate Dam and Weitchpec, CA. Imagery was acquired using a DJI Phantom 4 Pro rotorcraft (DJI, Shenzhen, China) (Figure 2), using the native RGB 20 megapixel camera. This camera has an 8.8 mm focal length and a resolution of 5472 × 3648 pixels [54]. Flights were conducted for 15 to 20 min over each of the 32 sites, at flying heights between 19 and 104 m. This resulted in estimated ground sampling distances between 0.52 and 2.85 cm (Table A1). Single images per site that encompassed the width of the channel and included portions of each side of the riverbank in every photo were acquired. Images were captured manually (without an automated mission flight plan) at nadir above each site, avoiding sun glint to minimize the relative angle between the sun and surface ripples. The UAV was flown between 10:04 a.m. and 5:11 p.m. to correspond with in-situ transect surveys, from 26 June to 9 July 2019. The time of each UAV flight was converted to solar elevation in degrees (now referred to as solar angle) using the University of Oregon Solar Radiation Monitoring Lab's Sun Chart Program (http://solardat.uoregon.edu/SunChartProgram.html). Solar angle helps predict the amount of solar glare that is reflected off the water's surface and into the aerial sensor. No calibration panel was used for the RGB image analysis, and no atmospheric correction was applied, as this is often unnecessary for low-flying UAV image collection over a small region [55]. Aerial imagery was captured to classify submerged filamentous algae and rooted macrophytes and estimate percent cover of each class from 32 sites within ten reaches of the Klamath River ( Figure  1) between Iron Gate Dam and Weitchpec, CA. Imagery was acquired using a DJI Phantom 4 Pro rotorcraft (DJI, Shenzhen, China) (Figure 2), using the native RGB 20 megapixel camera. This camera has an 8.8 mm focal length and a resolution of 5472 × 3648 pixels [54]. Flights were conducted for 15 to 20 min over each of the 32 sites, at flying heights between 19 and 104 m. This resulted in estimated ground sampling distances between 0.52 and 2.85 cm (Table A1). Single images per site that encompassed the width of the channel and included portions of each side of the riverbank in every photo were acquired. Images were captured manually (without an automated mission flight plan) at nadir above each site, avoiding sun glint to minimize the relative angle between the sun and surface ripples. The UAV was flown between 10:04 a.m. and 5:11 p.m. to correspond with in-situ transect surveys, from 26 June to 9 July 2019. The time of each UAV flight was converted to solar elevation in degrees (now referred to as solar angle) using the University of Oregon Solar Radiation Monitoring Lab's Sun Chart Program (http://solardat.uoregon.edu/SunChartProgram.html). Solar angle helps predict the amount of solar glare that is reflected off the water's surface and into the aerial sensor. No calibration panel was used for the RGB image analysis, and no atmospheric correction was applied, as this is often unnecessary for low-flying UAV image collection over a small region [55].

In-Situ Surveys
To validate the UAV imagery, in-situ surveys were conducted for rooted aquatic plants and filamentous algae along three to four transects (from one side of the river bank to the other) at each of the 10 reaches ( Figure 3) in the Klamath River below Iron Gate Dam between 26 June and 9 July 2019. Reaches were surveyed starting at the reach closest to Iron Gate Dam, and ending near Weitchpec, CA. At each of the 10 reaches, the riverbed was surveyed to assess the coverage of aquatic vegetation. Reaches ranged from 1 to 8 km in length, based on river access points. Kayaks were used to float the 10 reaches, surveying six transects in each reach, with transects randomly chosen among variable habitats represented in the reach. Three to four out of the six transects per reach were selected for analysis based on corresponding UAV image quality and to reduce computational processing time. Each transect was a minimum of 100 m downstream from the previous transect, and attempts

In-Situ Surveys
To validate the UAV imagery, in-situ surveys were conducted for rooted aquatic plants and filamentous algae along three to four transects (from one side of the river bank to the other) at each of the 10 reaches ( Figure 3) in the Klamath River below Iron Gate Dam between 26 June and 9 July 2019. Reaches were surveyed starting at the reach closest to Iron Gate Dam, and ending near Weitchpec, CA. At each of the 10 reaches, the riverbed was surveyed to assess the coverage of aquatic vegetation. Reaches ranged from 1 to 8 km in length, based on river access points. Kayaks were used to float the 10 reaches, surveying six transects in each reach, with transects randomly chosen among variable habitats represented in the reach. Three to four out of the six transects per reach were selected for analysis based on corresponding UAV image quality and to reduce computational processing time. Each transect was a minimum of 100 m downstream from the previous transect, and attempts were made to survey transects with more pool-type and riffle-type characteristics. Each transect's general habitat type was estimated based on river velocity and surface tension characteristics and classified as a run, riffle, or pool (Table A1). Areas of rapids and swift current were under-represented in surveys due to safety concerns and the general inability of a surveyor to access these swift sections. Transect selection was balanced with what was possible to survey safely, as well as what was representative of the reach; thus, transects were selected in the field while conducting these surveys. No transects were fully wadeable across the river, confirming that the mixed survey method of wading and snorkeling was necessary. Each in-situ survey was completed in about an hour per site. became too deep for wading (> 1-1.4 m, depending on river velocity), the surveyor transitioned to snorkeling. Individual quadrants were not surveyed along some transects due to especially swift or deep points (N = 56). At all transects, at least seven quadrants were surveyed. At each quadrant, the surveyor lowered a 40 × 40 cm weighted PVC square to the riverbed to visually estimate information about filamentous algae and rooted macrophytes. Filamentous algae coverage was recorded when filaments were > 2 cm long. For each species or genera observed in a quadrant (generally identified to species for rooted aquatic plants and genus for algae), the surveyor recorded the species or genus code, percent cover of the specific taxa, the substrate that the vegetation type was growing on, plant condition and color, and the average length of the plant or its filaments within the quadrant. Mat thickness was recorded in species without obvious filaments. Water depth was recorded using a marked depth stick (quadrants < 2 m deep) or sonar depth finder (quadrants > 2 m deep, Vexilar LPS-1, Vexilar, Minneapolis, Minnesota). To test the effect of increasing mean river depth on the detectability and classification accuracy of benthic primary producers, in-situ depth measurements were averaged per site. Discrete points within each site were also evaluated to assess the maximum depth detection of filamentous algae and rooted macrophytes in the aerial imagery. Turbidity was measured at deep portions (> 1 m) of a river reach by vertically deploying a Secchi disk from a kayak up to five times and taking an average. Secchi disk observations were not available at each reach due to depth limitations. Finally, the technician and UAV pilot recorded observational field notes regarding the composition of benthic filamentous algae and rooted macrophytes 100 m upstream and 100 m downstream of each transect by kayaking and collecting samples from sizeable patches (> 1 m 2 ) to assist in the image classification and validation processes.  Several variables were measured and collected during the in-situ surveys. At each transect, a shore-based field technician recorded longitude and latitude with a handheld Global Positioning System (GPS) (GPSMAP 76CSx, Garmin Inc., Olathe, Kansas) and a GPS-enabled GoPro Hero 5 (GoPro Inc., San Mateo, California), and measured wetted width with laser rangefinders from the river's edge. The technician calculated the width of the channel using the rangefinders, and divided each transect into 11 evenly spaced quadrants, from one side of the riverbank to the other. When possible, the surveyor waded to each of the quadrant locations ( Figure 2). However, when the river became too deep for wading (> 1-1.4 m, depending on river velocity), the surveyor transitioned to snorkeling. Individual quadrants were not surveyed along some transects due to especially swift or deep points (N = 56). At all transects, at least seven quadrants were surveyed. At each quadrant, the surveyor lowered a 40 × 40 cm weighted PVC square to the riverbed to visually estimate information about filamentous algae and rooted macrophytes. Filamentous algae coverage was recorded when filaments were > 2 cm long. For each species or genera observed in a quadrant (generally identified to species for rooted aquatic plants and genus for algae), the surveyor recorded the species or genus code, percent cover of the specific taxa, the substrate that the vegetation type was growing on, plant condition and color, and the average length of the plant or its filaments within the quadrant. Mat thickness was recorded in species without obvious filaments. Water depth was recorded using a marked depth stick (quadrants < 2 m deep) or sonar depth finder (quadrants > 2 m deep, Vexilar LPS-1, Vexilar, Minneapolis, Minnesota). To test the effect of increasing mean river depth on the detectability and classification accuracy of benthic primary producers, in-situ depth measurements were averaged per site. Discrete points within each site were also evaluated to assess the maximum depth detection of filamentous algae and rooted macrophytes in the aerial imagery. Turbidity was measured at deep portions (> 1 m) of a river reach by vertically deploying a Secchi disk from a kayak up to five times and taking an average. Secchi disk observations were not available at each reach due to depth limitations. Finally, the technician and UAV pilot recorded observational field notes regarding the composition of benthic filamentous algae and rooted macrophytes 100 m upstream and 100 m downstream of each transect by kayaking and collecting samples from sizeable patches (> 1 m 2 ) to assist in the image classification and validation processes.

Image Classification
A supervised classification scheme was employed to categorize image pixels into several classes after images were chosen. First, all aerial photos were reviewed and 3 to 4 images from each of the 10 river reaches were selected based on image quality and in efforts to reduce computational processing time. A total of 32 images (out of 60 images) were chosen for analysis after visually inspecting all photos and selecting the highest-quality images that had reduced sun reflection and glare, variation of in-situ depth measurements, and were captured at a nadir camera angle. In addition to image quality, these 32 images were subjectively selected to reduce image processing time associated with classifications and percent cover estimates. A single-image analysis approach was used in which only one image per site was analyzed (orthomosaics were not constructed for each site) due to challenges in stitching aerial images over a uniform surface such as water. Image metadata was then extracted using the uasimg R package [56] and incorporated to map centroids and footprints of images, as well as to estimate ground sampling distance and above ground altitude of each image. Yaw was manually corrected using Georeference tools in ArcGIS Pro 2.5 (Esri, Redlands, California).
Known GPS locations and visual observation were used to train and classify the UAV imagery into a six-class schema (1. algae, 2. macrophytes, 3. water, 4. land, 5. shadows, and 6. sun glint) with a machine learning algorithm. To detect these classes within the imagery, in-situ validation locations within each image were identified using GPS coordinates extracted from metadata on the GPS devices. It is important to acknowledge that positioning and alignment errors are introduced from the GPS devices used in the river, as well as from the Phantom 4 Pro internal GPS and subsequent EXIF modeling that depends on x, y, and z coordinate values. These errors were minimized by manually aligning images using high resolution basemaps in ArcGIS Pro and Google Earth, comparing river width measurements using ArcGIS Pro tools and in-situ width measurements from laser rangefinders, overlaying several estimated GPS coordinates (GoPro, Garmin, UAV) within the map, and visually detecting the quadrant and surveyor within the photo. When neither the quadrant nor swimmer was visible in the imagery, training sample pixels were selected from regions identified in the visual observation notes that were taken during aerial missions.
Algae and macrophyte classes were identified within UAV photos by incorporating information from in-situ quadrant measurements that had 80% cover or more of either class, resulting in 162 quadrants used for filamentous algae classification and 46 quadrants for rooted macrophyte classification. These data were used to justify the 'pure pixel' assumption employed when selecting samples for classification. All filamentous algae and macrophytes in this study were submerged, and it was assumed that any algae or macrophyte pixels that were visible from above the water's surface represented only one class (thus were considered 'pure pixels'). The algae and macrophyte classes were categorized generally, without classification to the species or genus level, and filamentous algae (no diatoms) composed the algae class. The water class included any portion of the river in which filamentous algae or rooted macrophytes were not visible, including wetted channels and riverbed. The land class included any rocks, grass, trees, sand, or otherwise surrounding environment adjacent to the river. The shadows class included dark shadow pixels throughout an image, while the sun glint class involved highly reflective solar glare pixels on the surface of the water. Not all classes were represented in every image.
After creating the class schema, a Random Trees pixel-based supervised classifier was trained and applied to all image pixels in ArcGIS Pro. Roughly 50-80 training sample polygons were selected per class by visually identifying sample pixels of all classes except algae and macrophytes, which were verified with in-situ quadrant surveys and visual observation notes written concurrently during each aerial mission. To minimize error, 'pure pixel' samples (containing only pixels of the class) were created by selecting pixels from the center of a patch and with a minimum number of 3 pixels. Each image was classified using a Random Trees classifier (30 maximum number of trees, 30 maximum tree depth, 1000 maximum number of samples per class). This algorithm iteratively classifies an image using a random subset of training pixels to form a group of decision trees. Samples are classified Remote Sens. 2020, 12, 3332 8 of 24 based on the majority "vote" of the trained decision tree label outputs [57,58]. This machine learning algorithm was chosen because it reduces computational load, particularly compared to other ensemble classifiers such as boosting and bootstrap aggregating [59], and it is less prone to overfitting data, as artificial neural networks [60] and maximum likelihood classification [61] have been reported to do. Furthermore, Random Trees has low error rates of classification [62] in comparison to support vector machines and artificial neural networks [63], and this method is well suited for studies involving multiple sources of remote sensing imagery [59], as in this study in which each image is analyzed separately. Finally, the Random Trees method was chosen over OBIA because of its ability to detect heterogeneous environments [64], such as the fine-scale combinations of filamentous algae and rooted macrophyte patches that are common in the Klamath River [65].

Validation
UAV imagery classifications were validated using the in-situ survey data and visual observations recorded in the field. The predicted class was compared against the observed class to validate sampling points, and additional variables such as solar angle were analyzed to help understand the factors affecting classification accuracy. The accuracy was assessed by creating an error matrix to calculate errors of omission, errors of commission, and the overall accuracy using the in-situ transect measurements as reference data, and visual observations to increase the spatial distribution of validation. In this context, omission error can be described as how often filamentous algae in the river is omitted from the proper class, while commission error is how often pixels identified as filamentous algae in the image are incorrectly classified [19]. Overall accuracy describes the proportion of correctly mapped pixels out of the total number of reference pixels. In this study, the overall accuracy, and not the kappa coefficient, is reported following best practices to omit the latter metric in imagery classification accuracy assessments [66,67]. ArcGIS Pro 2.5 was used to calculate omission error, commission error, and overall accuracy by assigning ground reference values to 50-100 stratified random points (randomly distributed based proportionally on the area of each class) per image and computing an error matrix.

Percent Cover Estimates
The percent cover of filamentous algae and rooted macrophytes was estimated based on UAV image classifications to ascertain the spatial distribution of these groups within the Klamath River between Iron Gate Dam and Weitchpec, CA. All filamentous algal species that were classified in the imagery were pooled into one percent cover class ("algae"), and all rooted macrophyte species classified in the imagery were grouped to form another separate percent cover class ("macrophytes") [68]. To maintain a standardized number of pixels per image, all images were resampled to 0.014 m (the average resolution of all 32 images), and then clipped using a bounding box of 48.8 m × 48.8 m to mimic a 40 cm × 40 cm quadrant at 160 feet (48.8 m) flight altitude. Next, rasters were converted to polygons (simplified polygons with multipart features), and summary statistics were performed on the area of each class. ArcGIS Pro 2.5 tools, including Resample, Extract by Mask, Raster to Polygon, and Summary Statistics were used within ArcGIS ModelBuilder. Finally, percent cover was calculated by dividing each class' area by the total number of pixels in the image. Although each image was initially classified using the six-class schema described above, the land class was excluded from the filamentous algae and rooted macrophyte percent cover estimates to more accurately depict these classes within the river rather than within the entire image. Filamentous algae and macrophyte percent cover estimates within the UAV imagery were then compared to the in-situ quadrant class cover estimates per site using a linear regression [69] with the ggplot2 package [70] in R version 3.6.3 software [71].

Results
This section details this study's findings pertaining to image classification outcomes in relation to depth and solar angle, as well as percent cover estimates of benthic filamentous algae and rooted macrophytes calculated throughout study sites via aerial imagery classifications and in-situ surveys.

Random Trees Classification Results
Overall accuracy of all classes (algae, macrophytes, water, land, shadows, sun glint) from the Random Trees supervised classification of 32 sites of the Klamath River was 82% (Table 1). Commission error was lowest for the water class (5%) and highest for the macrophyte class (41%), while omission error was lowest for the shadows class (5%) and highest for the land class (32%). The algae class had lower commission error (27%) but higher omission error (17%) than the macrophyte class (41% commission error and 12% omission error). For individual sites, overall accuracy was highest at Weitchpec site 2 (WE 2) (95%) and lowest at Rocky Point site 4 (RP 4) (66%) ( Table A1). Weitchpec site 2 (WE 2) and Happy Camp site 1 (HC 1) had the lowest commission error of all 32 images' algae classes at 0%, while Above Interstate 5 site 5 (I5 5) had the lowest commission error of all images' macrophyte classes at 10%. For groups of sites, Weitchpec (WE) reach sites (N = 4) had the highest overall accuracy (89%) of all 32 sites, and Brown Bear (BB) reach sites (N = 3) had the lowest (73%) overall accuracy. Results demonstrated that depth and solar angle impacted the classification accuracies of the UAV imagery. Generally, overall accuracy increased with average depth per transect up to about 2.75 m (in-situ depth measurements ranged from 0 to 9 m) ( Figure 4). While the adjusted R 2 value was very low (adj R 2 = 0.115) between average depth per transect and overall accuracy, these results were statistically significant (p-value ≤ 0.05). To further understand the effects of solar angle and depth on accuracy, images were organized into three overall accuracy groups using Jenks natural breaks categorization: 1 (greater than 84% accuracy, N = 14), 2 (greater than 74% and less than or equal to 84% accuracy, N = 10), and 3 (less than or equal to 74% accuracy, N = 8). The majority of images in group 1 (highest overall accuracy) had solar angles between 47.5 and 58.  The spectral signals of this study's six classes can be described by the digital numbers captured by the native Phantom 4 Pro UAV RGB camera (Table A2). The algae class had the highest reflectance in the green band and second highest in the blue band, while the macrophyte class had highest values in the green band, followed by the red and then blue bands. The water class had the highest reflectance values in the blue band, and the land class had equal digital number values in the red and green bands. The shadows and sun glint classes each had equal digital numbers in their respective red, green, and blue band values.

Study Site Characteristics and Percent Cover Estimates
Rooted macrophytes dominated a majority of upstream reaches (I5 through Happy Camp), while filamentous algae dominated downstream reaches closer to the Pacific Ocean (Old Man River through Weitchpec). The I5 site 1 (I5 1) location had the highest macrophyte percent cover (31%), and the Old Man River site 2 (OMR 2) had the highest algae percent cover (39%) ( Figure 5). While filamentous algae were present at upstream sites (I5 to HC), no macrophytes (0%) were detected in the aerial imagery at Happy Camp site 4 (HC 4) or at any sites downstream of this location. The spectral signals of this study's six classes can be described by the digital numbers captured by the native Phantom 4 Pro UAV RGB camera (Table A2). The algae class had the highest reflectance in the green band and second highest in the blue band, while the macrophyte class had highest values in the green band, followed by the red and then blue bands. The water class had the highest reflectance values in the blue band, and the land class had equal digital number values in the red and green bands. The shadows and sun glint classes each had equal digital numbers in their respective red, green, and blue band values.

Study Site Characteristics and Percent Cover Estimates
Rooted macrophytes dominated a majority of upstream reaches (I5 through Happy Camp), while filamentous algae dominated downstream reaches closer to the Pacific Ocean (Old Man River through Weitchpec). The I5 site 1 (I5 1) location had the highest macrophyte percent cover (31%), and the Old Man River site 2 (OMR 2) had the highest algae percent cover (39%) ( Figure 5). While filamentous algae were present at upstream sites (I5 to HC), no macrophytes (0%) were detected in the aerial imagery at Happy Camp site 4 (HC 4) or at any sites downstream of this location.
Remote Sens. 2020, 12, x FOR PEER REVIEW 10 of 25 with both the most illumination of the substrate and least solar glare on the surface (which is heightened around solar noon). The spectral signals of this study's six classes can be described by the digital numbers captured by the native Phantom 4 Pro UAV RGB camera (Table A2). The algae class had the highest reflectance in the green band and second highest in the blue band, while the macrophyte class had highest values in the green band, followed by the red and then blue bands. The water class had the highest reflectance values in the blue band, and the land class had equal digital number values in the red and green bands. The shadows and sun glint classes each had equal digital numbers in their respective red, green, and blue band values.

Study Site Characteristics and Percent Cover Estimates
Rooted macrophytes dominated a majority of upstream reaches (I5 through Happy Camp), while filamentous algae dominated downstream reaches closer to the Pacific Ocean (Old Man River through Weitchpec). The I5 site 1 (I5 1) location had the highest macrophyte percent cover (31%), and the Old Man River site 2 (OMR 2) had the highest algae percent cover (39%) ( Figure 5). While filamentous algae were present at upstream sites (I5 to HC), no macrophytes (0%) were detected in the aerial imagery at Happy Camp site 4 (HC 4) or at any sites downstream of this location. To illustrate the changes in percent cover of submerged filamentous algae and rooted macrophytes moving downstream along the river, three examples that depicted low, medium, and high percent cover of UAV images were selected and their respective Random Trees classification results were displayed ( Figure 6). Tree of Heaven site 5 (TH 5) had low filamentous algae percent cover (0% algae, 7% macrophytes), as was commonly observed in upstream sites within close proximity to Iron Gate Dam. Brown Bear site 6 (BB 6) displayed equal percentages of filamentous algae and macrophytes (13% algae, 13% macrophytes) and represented a medium concentration of percent cover for the 32 sites in this study. Finally, Old Man River site 2 (OMR 2) was selected to display a high amount of primary producer percent cover because aerial imagery over this location detected the highest amount of filamentous algae of all sites (39% algae, 0% macrophytes). This site is representative of downstream sites below Happy Camp site 4, below which no rooted macrophytes were detected. Figure 6 also helps demonstrate the increase in channel width as the Klamath River approaches the Pacific Ocean. To illustrate the changes in percent cover of submerged filamentous algae and rooted macrophytes moving downstream along the river, three examples that depicted low, medium, and high percent cover of UAV images were selected and their respective Random Trees classification results were displayed ( Figure 6). Tree of Heaven site 5 (TH 5) had low filamentous algae percent cover (0% algae, 7% macrophytes), as was commonly observed in upstream sites within close proximity to Iron Gate Dam. Brown Bear site 6 (BB 6) displayed equal percentages of filamentous algae and macrophytes (13% algae, 13% macrophytes) and represented a medium concentration of percent cover for the 32 sites in this study. Finally, Old Man River site 2 (OMR 2) was selected to display a high amount of primary producer percent cover because aerial imagery over this location detected the highest amount of filamentous algae of all sites (39% algae, 0% macrophytes). This site is representative of downstream sites below Happy Camp site 4, below which no rooted macrophytes were detected. Figure 6 also helps demonstrate the increase in channel width as the Klamath River approaches the Pacific Ocean. The relationship between in-situ and UAV-based percent cover estimates of submerged filamentous algae and rooted macrophytes was assessed within the 32 sites to understand how aerial imagery results compared to river survey results. UAV percent cover estimates were moderately correlated to in-situ quadrant percent cover estimates for both the filamentous algae and macrophyte classes (algae: Pearson's r = 0.51, macrophytes: Pearson's r = 0.72), indicating a moderately strong linear relationship between each variable per dataset. Although the filamentous algae and macrophyte classes for UAV imagery and in-situ data had low and moderate adjusted R 2 values (algae: adj R 2 = 0.23, macrophytes: adj R 2 = 0.50), this result was statistically significant (p-value ≤ 0.05) (Figure 7). Thus, the null hypothesis that there is no relationship between the UAV and in-situ percent cover measurements for both filamentous algae and rooted macrophyte observations was rejected. For algae measurements, the residual standard error, or the average amount the response variable (algae) diverged from the regression line [72] was 9.214 with 30° of freedom. The residual standard error for macrophyte measurements is 7.208 with 30° of freedom. The F-statistic, an indicator of The relationship between in-situ and UAV-based percent cover estimates of submerged filamentous algae and rooted macrophytes was assessed within the 32 sites to understand how aerial imagery results compared to river survey results. UAV percent cover estimates were moderately correlated to in-situ quadrant percent cover estimates for both the filamentous algae and macrophyte classes (algae: Pearson's r = 0.51, macrophytes: Pearson's r = 0.72), indicating a moderately strong linear relationship between each variable per dataset. Although the filamentous algae and macrophyte classes for UAV imagery and in-situ data had low and moderate adjusted R 2 values (algae: adj R 2 = 0.23, macrophytes: adj R 2 = 0.50), this result was statistically significant (p-value ≤ 0.05) (Figure 7). Thus, the null hypothesis that there is no relationship between the UAV and in-situ percent cover measurements for both filamentous algae and rooted macrophyte observations was rejected. For algae measurements, the residual standard error, or the average amount the response variable (algae) diverged from the regression line [72] was 9.214 with 30 • of freedom. The residual standard error for macrophyte measurements is 7.208 with 30 • of freedom. The F-statistic, an indicator of significance between the groups of observations, was relatively low for algae (10.49) and moderate for macrophyte (32.01) measurements.  Results from this study indicated that filamentous algae were more easily detected and quantified in deeper regions and often inhabited areas farther from the riverbanks than did macrophytes below Iron Gate Dam. The filamentous algae class had high overall accuracy (overall accuracy of 80% or more and in-situ percent cover of 80% or greater per quadrant) down to a depth of 1.9 m, while rooted macrophytes (overall accuracy of 80% or more and in-situ percent cover of 80% or greater per quadrant) were accurately classified down to 1.2 m below the water's surface at discrete locations. Within the 32 sites, in-situ quadrant observations revealed that submerged rooted macrophytes generally grew at shallower depths than filamentous algae (0.83 m mean for rooted macrophyte in-situ percent cover > 0% per quadrant, 1.2 m mean for filamentous algae in-situ percent cover > 0% per quadrant). On average, rooted macrophytes grew within 7.7 m from either side of the bank within the 32 sites, and filamentous algae were most commonly found within 8.7 m from either bank. In the 32 sites analyzed, rooted macrophytes and filamentous algae both most frequently inhabited straight sections of the river that had lower river velocity (< 1 m sec −1 ) or in relatively shallow areas along banks.

Discussion
This study demonstrates that an inexpensive UAV equipped with an RGB camera is a useful tool for detecting and quantifying multiple types of submerged primary producers in a non-wadeable river. Filamentous algae and rooted macrophyte coverage had > 70% overall accuracy at 28 of 32 sites, and water surface solar reflection was the primary factor limiting the ability to assess submerged vegetation coverage. Along the 32 sites distributed among 10 reaches beginning below a series of large dams, there was greater dominance of rooted macrophytes upstream, transitioning to a greater dominance of filamentous algae at downstream sites. The larger spatial scale documented in UAV image capture can supplement current in-situ survey methods or provide a novel and useful method to assess primary producer coverage in rivers where in-situ surveys are not possible due to excessive depth or velocity.

UAV Monitoring of Benthic Primary Producers
This study's UAV method resulted in the highest classification accuracy in river conditions with high water clarity and homogenous primary producer assemblages. In contrast, other researchers have found that classification accuracy decreased with river depth, particularly in water deeper than 3 m [42,73,74]. Our results were slightly different from these studies in that filamentous algae were more easily detected and quantified in deeper water (up to 2.75 m). In our system, water depth covaried with water clarity (Figure 4), largely as a result of the tributaries entering the Klamath River below Iron Gate Dam which tend to contribute relatively clearer water to the main stem, creating Results from this study indicated that filamentous algae were more easily detected and quantified in deeper regions and often inhabited areas farther from the riverbanks than did macrophytes below Iron Gate Dam. The filamentous algae class had high overall accuracy (overall accuracy of 80% or more and in-situ percent cover of 80% or greater per quadrant) down to a depth of 1.9 m, while rooted macrophytes (overall accuracy of 80% or more and in-situ percent cover of 80% or greater per quadrant) were accurately classified down to 1.2 m below the water's surface at discrete locations. Within the 32 sites, in-situ quadrant observations revealed that submerged rooted macrophytes generally grew at shallower depths than filamentous algae (0.83 m mean for rooted macrophyte in-situ percent cover > 0% per quadrant, 1.2 m mean for filamentous algae in-situ percent cover > 0% per quadrant). On average, rooted macrophytes grew within 7.7 m from either side of the bank within the 32 sites, and filamentous algae were most commonly found within 8.7 m from either bank. In the 32 sites analyzed, rooted macrophytes and filamentous algae both most frequently inhabited straight sections of the river that had lower river velocity (< 1 m sec −1 ) or in relatively shallow areas along banks.

Discussion
This study demonstrates that an inexpensive UAV equipped with an RGB camera is a useful tool for detecting and quantifying multiple types of submerged primary producers in a non-wadeable river. Filamentous algae and rooted macrophyte coverage had > 70% overall accuracy at 28 of 32 sites, and water surface solar reflection was the primary factor limiting the ability to assess submerged vegetation coverage. Along the 32 sites distributed among 10 reaches beginning below a series of large dams, there was greater dominance of rooted macrophytes upstream, transitioning to a greater dominance of filamentous algae at downstream sites. The larger spatial scale documented in UAV image capture can supplement current in-situ survey methods or provide a novel and useful method to assess primary producer coverage in rivers where in-situ surveys are not possible due to excessive depth or velocity.

UAV Monitoring of Benthic Primary Producers
This study's UAV method resulted in the highest classification accuracy in river conditions with high water clarity and homogenous primary producer assemblages. In contrast, other researchers have found that classification accuracy decreased with river depth, particularly in water deeper than 3 m [42,73,74]. Our results were slightly different from these studies in that filamentous algae were more easily detected and quantified in deeper water (up to 2.75 m). In our system, water depth covaried with water clarity (Figure 4), largely as a result of the tributaries entering the Klamath River below Iron Gate Dam which tend to contribute relatively clearer water to the main stem, creating higher water clarity with distance downstream from the dam [75]. Deepening river channels created by high winter flows follow a similar longitudinal trend associated with these tributary inputs [76].
At the deepest sites with high overall accuracy, (Old Man River site 3: 2.09 m average depth, 90.38% overall accuracy; Orleans site 1: 2.75 m average depth, 92% overall accuracy; Weitchpec site 1: 2.69 m average depth, 91.3% overall average), conditions were optimal for capturing high-quality aerial imagery. At these sites, lower turbidity (Secchi depth between 2.1 and 4.4 m) (Table A1) was measured, and only filamentous algae (no rooted macrophytes) ( Figure 5) was detected. Having only one type of submerged primary producer to classify in each of these sites contributed to increased accuracy of the supervised classification [77,78] because it diminished the risk of conflated categorization. The counterintuitive finding that deeper river sites had higher accuracy when these deeper locations co-occur with lower turbidity reinforces previous findings that water clarity is a major predictor for how well imagery can be used to successfully detect and quantify submerged primary producers [43,79,80]. This demonstrates that UAV methods are not limited to shallow (< 1 m) rivers when turbidity is low and primary producer assemblages are homogenous.
This study was novel in that it attempted to classify both filamentous algae and rooted macrophyte groups using an RGB camera throughout a non-wadeable river [20,34]. However, there were instances of class confusion between filamentous algae and rooted macrophytes. In some dense areas of aquatic vegetation, algae grew on rooted macrophytes, making class separation challenging. In these cases, spectral unmixing may be a more suitable classification approach [81,82]. Furthermore, in-situ surveys along a single transect left several unknown or unidentified points of interest in the UAV imagery, making class discrimination challenging. Also, both algae and macrophyte classes had the highest spectral signals in the green band (Table A2), which further exacerbated confusion between these classes [83]. Some algae genera (Cladophora, Oedogonium) and macrophyte species (Zannichellia palustris, Potamogeton crispus) had both light and dark green reflectance values throughout the river, resulting in too much overlap of reflectance values to create a useful spectral library to identify filamentous algae and rooted macrophyte classes across all 32 images. Each image was classified individually, and the classifications were reduced to two distinct groups because creating spectral libraries for each image requires in-situ sampling and is time consuming. Due to the individual classification per image and lack of automation in this process, large sample sizes might not be well suited under these specific methods.
While OBIA is commonly used for high-resolution underwater classification of vegetation and wetland species [84,85], this study utilized a pixel-based supervised classification. Pixel-based methods facilitate the identification of large patches of rooted macroalgae or filamentous algae, as well as speckled and interspersed filamentous algae, both of which were common at study sites. Random Forest algorithms have proven successful in previous studies in complex, heterogeneous landscapes [19,85,86], demonstrating that this method can effectively reduce speckle and noise within imagery. To build on current work, future image analysis techniques that can be applied to benthic surveys in heterogeneous environments include aquatic vegetation indices, OBIA, structure from motion photogrammetry, and hyperspectral image capture [20,40,41,[86][87][88][89].
One of the most prominent challenges to achieving high classification accuracy in this study involved solar glare, which is affected by solar angle, wind, and ripples or surface tension on the water [90][91][92][93]. In this study, the sun's reflection produced large white spots in photos that were taken close to solar noon (around 71 • solar angle in this study), thus making detection below the surface impossible. Several images captured around solar noon had some of the lowest classification accuracy results (between 66% and 72%) due to issues related to solar angle and resulting glare. Wind and higher velocity water (> 1 m s −1 ), such as in areas with ripples, further exacerbated the prominence of solar glare and created distortions within the UAV imagery. These concerns affected how the 32 out of 60 total UAV images were selected for analysis, and several selected photos (n = 5) had sizable portions (between 8% and 12% of the image) of solar glare that prohibited visibility below the water's surface. Flying on a uniformly cloudy day would also reduce solar glare, although detection capabilities may suffer as light penetration through the water column would be reduced [94].
Methods from this study can be applied to a variety of environments, including both freshwater and marine [88,95], and will perform best in clear and relatively shallow (< 2 m) waters. Given that classification accuracy was highest in regions with reduced water surface tension, such as in a pool or a slow-moving river run, techniques from this study would also be well-suited for lake environments [19]. Additionally, the Random Trees classifier most easily identified large, homogenous patches of algae and macrophytes, and would similarly perform well with other large groups of macroalgae and submerged aquatic vegetation, such as kelp or seagrass [96][97][98]. Despite these applications, this method is not optimal for distinguishing multiple aquatic species, as RGB imagery is spectrally sparse. Therefore, aerial hyperspectral sensors would be more effective in these scenarios [99,100].
Detection and quantification techniques in future studies might be complicated in water with increasing depth, turbidity, and confounding primary producer types [43,101], and accuracy will likely decrease as these conditions converge. Thus, times of higher river flows that increase water depth and turbidity, including spring snowmelt or autumn rains, are expected to result in lower accuracy of primary producer identification and quantification. Changing water levels associated with water year variation, dam management, and season are expected to influence primary producer assemblages [45,102,103] in addition to methodological accuracy, making complementary in-situ surveys important during variable conditions. Examples of expected changes to primary producer assemblages include lower overall coverage of rooted macrophytes during high-flow years, and seasonal progression of rooted macrophyte and filamentous algae coverage, resulting in lower coverage as macrophytes and algae slough in the fall [104,105].

Recommendations
Analysis of UAV photos using a single-image, supervised classification method for 32 sites proved to be time intensive due to the need to create a novel spectral library and separate classification for each high-resolution image. Ways to address this in future studies involve assessing fewer sites and spectrally correcting the imagery in order to apply a uniform spectral library to all photos. While this study represented a static glimpse of conditions on the Klamath River at 32 individual locations, this method can successfully be applied to smaller-scale river studies with greater temporal scale, such as time series or change analyses of benthic communities in small sections of a river. For example, managers can adapt this technique to capture aerial images of the same river location each month to help answer how specific filamentous algal communities in one region of a river change in response to seasonal dam-altered flow regimes. While this small-scale study is best suited for small rotary-wing UAVs, pilots who are designing future projects may consider employing a larger rotary-wing UAV with additional power capacity when surveying larger portions of a river [106].
For monitoring greater river extents (> 100 m river length per site), it is helpful to consider flight altitudes, platforms, and study site characteristics before conducting a UAV-based benthic primary producer study. A flight plan at high altitudes (above 100 m) is recommended to both avoid obstacles (such as trees, power lines, and bridges) and assist in the image stitching process in order to create orthomosaic maps of large swaths of the river. When generating orthomosaics, researchers should use real-time kinematic GPS, or have ground control points, secured markers or landmarks with known GPS locations, or slivers of land in the images to facilitate image stitching [40,107,108]. When target algal or macrophyte patches are identifiable at large spatial scales, or when higher spectral resolution is required to discern among taxa, high-resolution satellite imagery can be an alternative to UAV imagery [109]. Fixed-wing UAVs may be difficult to fly in meandering river environments unless there are open riverbanks or roads adjacent to the study sites that can be used for take-off and landing [110]. As river survey extent increases, UAV data increases, thereby posing additional constraints to computational processing time. The sampling and image capture methods described in this study are broadly applicable to benthic primary producer studies, but caution should be taken when attempting to apply this technique to large spatial extents due to computational and logistical constraints. In summary, this study's methods are best suited for smaller-scale studies across time, rather than across space.
Another recommendation includes changing the order of the in-situ sampling protocol that was employed in this study. Instead of simultaneously conducting an in-situ transect survey and flying a drone overhead, it is more effective to capture aerial imagery first, identify several target sampling areas in the UAV photo, and then collect samples and observations in discrete locations of the river to reflect target areas within the photo. Also, including in-situ biomass calculations of filamentous algae and rooted macrophytes to the sampling protocol would augment percent cover estimates. Finally, incorporating a spectral calibration target and a corrective optical equation [36] would allow managers to apply a spectral library to the entire UAV image dataset, which would reduce time spent on individual classifications and increase the ability to discriminate among different submerged taxa.

Management Applications of UAV Monitoring in Non-Wadeable Rivers
UAV imagery is useful in detecting and quantifying submerged primary producer assemblages, and can be integrated into monitoring programs where current monitoring is limited or non-existent. Algae and aquatic plants are sensitive to ecosystem change, making them useful indicators of ecosystem health [111,112]. The type, distribution, and condition of algae and aquatic plants can drive water quality (pH, dissolved oxygen, suspended sediment, water temperature) and fisheries health (food resources, habitat), making quantification of these taxa useful in identifying mechanisms responsible for ecosystem change [7,111,113,114]. In non-wadeable rivers such as the Klamath River, documentation of aquatic plants and algae is generally limited to wadeable areas easily accessible from shore, where the primary producer assemblage may not be representative of the river reach [45]. In this study, UAV photos expanded the spatial scale of the surveyed area, and included areas generally ignored in established survey methods, including areas > 1 m deep, areas of swift current, and zones not adjacent to shoreline river access points. Although aerial images cannot capture the taxonomic detail with the same accuracy as in-situ surveys or samples collected for laboratory analysis of species composition, the larger spatial scale can help address other research questions and monitoring goals. Ultimately, combining in-situ and UAV methods will likely provide the most thorough monitoring practice.
UAV images can be used to track ecosystem change associated with management actions and restoration expected to influence primary producer assemblages. The Klamath River and many larger rivers are considered impaired due to high aquatic plant and algae biomass, spurring management actions aimed at reducing these proliferations [65,115]. In the case of the Klamath, nutrient criteria are set by water quality management agencies, but challenges with monitoring aquatic plant and algae growth limit the ability to monitor the effectiveness of regulatory nutrient reductions and restoration actions [116,117]. In rivers where confounding factors are expected to influence benthic algae growth, including nutrients, flow, and water clarity, alterations including dams and associated flow management can further influence growth patterns of algae and aquatic plants [118,119]. Where restoration has the potential to alter algae and aquatic plant assemblages, as is the case in the proposed removal of the four large hydroelectric dams ( Figure A1) on the Klamath River [120], monitoring these assemblages with UAV-based techniques provides a promising opportunity to learn about the effects of large-scale restoration on non-target taxa that may otherwise be challenging to monitor. In cases in which analytical expertise may not be available, UAV imagery can still be collected to document conditions before and after management actions so that data can later be analyzed.

Conclusions
Employing a pixel-based Random Trees supervised classification on RGB images from a low-cost UAV is an effective technique to classify and quantify both benthic filamentous algae and rooted macrophytes in a non-wadeable river. Although increases in water-column turbidity, water depth, and complex species assemblages can decrease the accuracy of rooted macrophyte and filamentous algae quantification, overall accuracy was 82% at the 32 assessed sites. This study fits within the level of accuracy of similar studies analyzing submerged aquatic vegetation coverage in clear water at shallow depths (< 1 m) [19,85,86,121], and results verified that these methods can be applied in deeper bodies of water, both freshwater and marine, that have high water clarity. These findings also expand on existing research by suggesting that classifying vegetation categories and assessing images in mixed vegetation assemblages, in addition to single taxa-dominant ecosystems, is possible with an RGB camera. It is recommended that future studies collect aerial imagery between about 45 to 60 • of solar elevation angle in the morning or afternoon to reduce solar glare. Photos taken around solar noon should generally be avoided to diminish the effect of solar reflection. Future analysis involves incorporating in-situ biomass calculations to augment percent cover estimates, and further investigating the utility of multispectral imagery detection at depths deeper than 1 m below the surface. Methods from this study are not suited for a purely white-water river caused by wind and rapids due to visibility issues related to solar glare and water surface tension. Instead, this study's methods are particularly useful in inaccessible study regions that host large, homogenous patches of submerged filamentous algae or macrophytes in relatively shallow (< 2 m) and clear water.
As UAVs become a common tool used by research and monitoring groups, the potential to increase ecosystem monitoring in environments not previously accessible and on larger spatial scales is expanding. Capturing aerial imagery is relatively efficient and can be combined with other sample collection efforts as part of regular water quality monitoring programs. These images can be cataloged for later assessment, which may prove incredibly valuable in the cases of retrospective monitoring of unplanned management actions, natural variation in ecosystem condition, or other changes. The collection of UAV images from established long-term study sites will be comparable through time to assess temporal changes with limited in-situ data, and aerial images collected in conjunction with in-situ samples will be informative in aquatic system management and decision-making.  of accuracy of similar studies analyzing submerged aquatic vegetation coverage in clear water at shallow depths (< 1 m) [19,85,86,121], and results verified that these methods can be applied in deeper bodies of water, both freshwater and marine, that have high water clarity. These findings also expand on existing research by suggesting that classifying vegetation categories and assessing images in mixed vegetation assemblages, in addition to single taxa-dominant ecosystems, is possible with an RGB camera. It is recommended that future studies collect aerial imagery between about 45 to 60° of solar elevation angle in the morning or afternoon to reduce solar glare. Photos taken around solar noon should generally be avoided to diminish the effect of solar reflection. Future analysis involves incorporating in-situ biomass calculations to augment percent cover estimates, and further investigating the utility of multispectral imagery detection at depths deeper than 1 m below the surface. Methods from this study are not suited for a purely white-water river caused by wind and rapids due to visibility issues related to solar glare and water surface tension. Instead, this study's methods are particularly useful in inaccessible study regions that host large, homogenous patches of submerged filamentous algae or macrophytes in relatively shallow (< 2 m) and clear water.
As UAVs become a common tool used by research and monitoring groups, the potential to increase ecosystem monitoring in environments not previously accessible and on larger spatial scales is expanding. Capturing aerial imagery is relatively efficient and can be combined with other sample collection efforts as part of regular water quality monitoring programs. These images can be cataloged for later assessment, which may prove incredibly valuable in the cases of retrospective monitoring of unplanned management actions, natural variation in ecosystem condition, or other changes. The collection of UAV images from established long-term study sites will be comparable through time to assess temporal changes with limited in-situ data, and aerial images collected in conjunction with in-situ samples will be informative in aquatic system management and decisionmaking.    Table A1. Results from the 32 sites: overall accuracy, algae percent cover, macrophyte percent cover, estimated ground sampling distance (GSD), Secchi depth, habitat, average depth per site, solar angle, flight latitude (center point), and flight longitude (center point).

Reach Name Site
Overall Accuracy (%)  Table A2. Digital numbers (spectral signals) for UAV image classes in this study.