Abstract

Vegetation health monitoring is key to identifying early signs of water stress, pollutant-induced toxicity, and plant diseases in green urban stormwater facilities. However, rigorous monitoring to collect accurate quantitative data is an expensive and time-consuming process. This paper examines the feasibility of using uninhabited aircraft systems (UAS), in comparison to standard ground-based methods, for monitoring biomass and primary production in two bioswale cells at an urban stormwater facility. Implementation of the UAS-based approach involved flight planning in an urban area to meet resolution requirements of bioswale imagery obtained from near-infrared and red-green-blue cameras. The resulting normalized difference vegetation index (NDVI) estimated from UAS data was tracked over a 2-month period during the transition from spring to summer, showing the spatial distribution of NDVI and the change in vegetation coverage areas over time. In comparison, ground-based measurements of the fraction of intercepted photosynthetically active radiation (PAR) presented multiple practical challenges during implementation in the field, leading to over- and underestimates of intercepted PAR. Overall, UAS-derived NDVI was found to be a valuable reflectance-based, vegetation health-monitoring methodology that can be used by utilities and cities for practical, cost-effective, and rapid assessment of vegetation stress and for long-term maintenance in green stormwater facilities.

HIGHLIGHTS

  • Ground-based monitoring of green infrastructure (GI) facilities for estimation of fraction of intercepted photosynthetically active radiation poses multiple operational challenges.

  • Uninhabited aircraft systems (UAS) were found to enable efficient acquisition of multi-temporal normalized difference vegetation index (NDVI) data for GI site monitoring, through the workflow developed in this study.

  • Time series of UAS-derived NDVI showed the expected downward trend over a two-month period starting on June 1, 2019.

  • Spatial variation in UAS-derived NDVI provides plant-specific health information that may assist in GI site management.

INTRODUCTION

Green infrastructure (GI) for stormwater management is an attractive solution for urban municipalities, serving to reduce urban flooding and treat contaminated stormwater, while simultaneously providing aesthetically pleasing spaces in urban areas. Plants play a vital role in GI facilities by helping reduce the flow velocity of stormwater, providing nutrients to the microbial communities in the soil, maintaining soil permeability, and taking up contaminants that infiltrate the soil in stormwater runoff (USEPA 1999; Akpor & Muchie 2010). To support function and performance of plants in GI, vegetation health-monitoring practices must be established. Monitoring the health of vegetation in GI can reveal signs of water stress, pollutant-induced toxicity, and disease. Hence, early detection of stress in vegetation is critical for long-term sustainability and effectiveness of GI, especially when there is a potential for buildup of sediments and contaminants in media and vegetation that may reduce treatment efficiency (Lucke & Nichols 2015). Methods for monitoring GI hydrology have been extensively investigated in the literature. However, methods for monitoring the health of urban GI vegetation are not well studied. Previous work on monitoring vegetation in GI in practice has primarily been limited to the identification, location, and description of vegetation density and height, with no detailed methodology included beyond a visual inspection of these plant-based elements at GI facilities such as bioswales, bioretention systems, raingardens, and green roofs (USEPA 2002, 2009).

Multiple potential methods exist that could be applied toward long-term monitoring of vegetation in GI facilities, including both ground-based and remote sensing methods. Remote sensing technology holds the potential for efficient, recurrent monitoring of large-scale green spaces for vegetation health and productivity. Recent studies have used vegetation indices and vegetation density metrics derived from aerial imagery acquired using aircraft and from satellites to explore canopy cover and urban greenness. These studies have contributed to planning for design and maintenance and valuing urban green areas (Franco & Macdonald 2018; Gadi et al. 2018). Studies have also explored the cooling effects of large urban green areas where various types of remote sensing data have been utilized including multispectral, hyperspectral, and light detection and ranging (LiDAR) data (Bartesaghi-Koc et al. 2017, 2019). Additionally, remote sensing imagery has been used in studies for relating urban GI and green spaces to effectiveness of GI in mitigating flood events (Kim et al. 2017; Lee 2018).

Despite the usefulness of remote sensing methods in urban green spaces, limited research exists on the application of uninhabited aircraft systems (UAS; also defined as ‘unmanned aircraft systems’ by Federal Aviation Administration; FAA 2016a) for monitoring and management of stormwater GI facilities and urban water infrastructure management (Hill & Babbar-Sebens 2019; McDonald 2019). For example, the use of UAS for monitoring of vegetation health in stormwater GI using vegetation indices, such as the normalized difference vegetation index (NDVI), has been limited to countable studies, such as those conducted recently by Dimitrov et al. (2018), in which NDVI derived from UAS waveband data was used to detect health and disease of vegetation in an urban green space of Bulgaria. However, this study did not explore the usefulness of their approach for temporal monitoring of vegetation classes and did not provide any insight on how the NDVI data could be compared with their in situ data for potential tradeoffs in monitoring approaches. In contrast, UAS-derived NDVI has been more popular in the agriculture industry for low-cost assessment of crop health and overall field monitoring (Sankaran et al. 2015; Shi et al. 2016).

In addition to remote monitoring, individual plant growth can also be monitored using ground-based methods. Destructive methods of measuring plant growth and health are labor intensive, involving harvesting plants at the ground level, drying and weighing them to determine biomass (Prabhakaraa et al. 2015). Various hand-held sensors can measure parameters of plant health, including chlorophyll meters (Muñoz-Huerta et al. 2013) that measure vegetation greenness, chlorophyll fluorometers (Mohammed et al. 1995) that measure chlorophyll fluorescence, leaf porometer instruments (Toro et al. 2019) that measure stomatal conductance, and ceptometer instruments (Bréda 2003) that measure the fraction of intercepted photosynthetically active radiation (fIPAR). Chlorophyll meters and porometers provide measures of vegetation system functions at the leaf level, requiring large samples due to variations between leaves (Bauerle et al. 2004; Möller et al. 2007; Cordon et al. 2016). However, ceptometer instruments measure intercepted radiation at the plant level. The fIPAR is the amount of solar radiation in the spectral range of 400–700 nm (i.e., the visible portion of the electromagnetic spectrum) that is intercepted by vegetation for the purpose of vegetation growth. The fIPAR has been shown to have a strong correlation with dry biomass production, or net primary productivity and is used in models of dry matter production efficiency and radiation use efficiency (Monteith 1972; Rosati & Dejong 2003). The fIPAR is also considered an essential plant parameter for defining functional groups used in process-based models. It has been previously used to define functional groups for wetlands in several representative wetland regions in the USA (Williams et al. 2017). Additionally, studies show that the fIPAR and NDVI have a linear relationship when applied to row crops and prior to the onset of senescence in sampled vegetation (Hatfield et al. 1984; Serrano et al. 2000). This indicates that NDVI may provide an estimate of fIPAR and biophysical parameters related to fIPAR.

While the use of NDVI and fIPAR for monitoring or detecting change in crops and natural systems is a well-studied topic, there is a lack of knowledge about the effectiveness of monitoring these parameters as indicators of stress and health of vegetation in engineered systems, specifically stormwater GI systems (McDonald 2019). The implementation of remote technology to monitor GI vegetation growth and health may provide insights into vegetation-dependent performance while informing targeted operational and policy decisions. Fast detection of physiological changes in vegetation may provide indications of exposure to new or altered contaminant loads, signs of senescence and disease, water stress, or other stressors. In turn, this information can guide decisions related to adaptive mitigation or removal/disposal and replacement of vegetation and possibly soil media in GI facilities.

The goal of this study is to evaluate methods for rapid and cost-effective monitoring of GI vegetation health, with a focus on UAS platforms for remotely monitoring bioswale facility vegetation health via the estimation of NDVI. UAS are of interest as for GI health monitoring for two primary reasons. First, the spatial extents of typical stormwater GI facilities in urban areas (generally less than a few km2) are well suited to UAS data collection. (Larger areas would be more suitable to data collection from conventional manned aircraft or satellites, whereas smaller sites could be effectively studied using solely ground-based methods.) Second, the relative low cost and logistical simplicity of UAS data collection compared with data acquisition from conventional aircraft makes repeat, high-resolution data collection possible. To compare the UAS-based approach with a ground-based vegetation health-monitoring method, this study collected data on fIPAR of the vegetation canopy, a measure that has previously been reported to be proportional to biomass production. These two measures (fIPAR and the NDVI) were then compared to explore potential relationships between each other and to determine if the NDVI can act as an estimate of fIPAR and the physiological parameters derived from fIPAR at the stormwater bioswales facility. Finally, temporal trend analysis using the UAS-derived NDVI was also evaluated for iterative monitoring of stress and vegetation health in the bioswale facility.

METHODOLOGY

Site description

The study was conducted at the OSU-Benton County Green Stormwater Infrastructure Research (OGSIR) Facility in Corvallis, Oregon. OGSIR is a field research facility for testing GI installations that captures runoff from approximately 9,300 m2 of catchment area in the county property, as shown in Figure 1. The county property is a maintenance and service facility that is representative of many industrial sites used for parking, storage, and refueling county vehicles and storage of construction materials and equipment. Stormwater generated in the catchment is routed via underground stormwater pipes into a 5,700 liters in-line underground storage tank, and then pumped into a sedimentation bay (or, sediment bay) via an effluent pump. The stormwater in the sedimentation bay then overflows via 45°, 4 in, V-notch weirs into each of the three bioswale cells of OGSIR facility for treatment. The three cells are each 28.4 m long, 3.2 m wide, and 0.9 m deep, except at the middle of each cell where there is an additional 0.3 m deep longitudinal trench containing a perforated underdrain running along the length of the cell. While, during the study period, Cell 1 contained no vegetation, bioswale Cell 2, and Cell 3 contained vegetation native to the local region and resilient to local climate. Cell 2 was originally planted primarily with sedges, rushes, and grasses, and Cell 3 was planted with a mixture of grass and broad leaf vegetation. In both cells, vegetation die-off occurred between the planting date in the summer of 2014 and the data collection dates in the summer of 2019. Furthermore, the two cells have heterogenous canopies due to the irregular plant locations and variations in plant health, species, and heights (<0.3–1.2 m), with bare areas dominated by weeds during the spring and dry grass during the summer. Hence, in this study, the measurement of vegetation parameters was focused on specific areas in the two bioswale cells that covered a range of originally planted species.

Figure 1

OSU-Benton County Green Stormwater Infrastructure Research (OGSIR) facility and its catchment at far end.

Figure 1

OSU-Benton County Green Stormwater Infrastructure Research (OGSIR) facility and its catchment at far end.

Ground-based monitoring approach

Ground-based measurements of fIPAR for estimating vegetation health were made using the Accupar LP-80 ceptometer. Measurements with the ceptometer were made within 2 h of local noon, as shown in Figure 2. Table 1 lists that dates when fIPAR data were collected, once in the middle of June and then again at the end of June, in 2019. The mid-June sampling event involved the collection of fIPAR data on June 15, 2019 in Cell 3. However, because of the extensive human labor involved in using ceptometer and the need for the data collection to occur within 2 hours of noon, the end-of-June sampling event was split over 2 days. Cell 2 fIPAR data were collected on June 29, 2019, and Cell 3 fIPAR data were collected on June 30, 2019. Since, both days (June 29 and 30) had similar weather, the data from two cells were combined to represent end-of-June (June 29, 2019) readings that coincided with UAS scanning on June 29, 2019 (discussed in detail in the next sub-section).

Table 1

Dates when data were collected from the different cells at OSU-Benton County Green Stormwater Infrastructure Research (OGSIR)

DatefIPAR collectedUAS imagery collected
June 1, 2019 – All cells 
June 15, 2019 Cell 3 All cells 
June 16, 2019 Cell 2 – 
June 29, 2019 Cell 2 All cells 
June 30, 2019 Cell 3 – 
July 27, 2019 – All cells 
DatefIPAR collectedUAS imagery collected
June 1, 2019 – All cells 
June 15, 2019 Cell 3 All cells 
June 16, 2019 Cell 2 – 
June 29, 2019 Cell 2 All cells 
June 30, 2019 Cell 3 – 
July 27, 2019 – All cells 
Figure 2

Accupar LP-80 ceptometer in use at the OSU-Benton County Green Stormwater Infrastructure Research (OGSIR) facility and the resulting AOI and wand orientation beneath a plant canopy.

Figure 2

Accupar LP-80 ceptometer in use at the OSU-Benton County Green Stormwater Infrastructure Research (OGSIR) facility and the resulting AOI and wand orientation beneath a plant canopy.

PAR readings from the 80 sensors in the probe were grouped into eight segments along the 81.3 cm (32 in) length of wand. A Leica GS14 Global Navigation Satellite System (GNSS) receiver, receiving master auxiliary concept (MAC) network corrections from the Oregon Real-time GNSS Network (ORGN), was additionally used to collect coordinates at the center of each area of interest (AOI), following procedures that previous studies have shown capable of yielding accuracies on the order of 2 cm horizontally and 3 cm vertically (Allahyari 2016; Allahyari et al. 2018). AOIs represented physical canopy space of individual plants that were randomly selected along the length of the bioswale cells and across plant groups (e.g., sedges, rushes, grasses, and broad leaf plants).

At each AOI, the GNSS antenna pole was centered over the point, and PAR measurements were taken above the canopy and below the canopy, carefully avoiding shading the sensors during measurements. Measurements were taken using a variation on the Plant Method described by Johnson et al. (2010). To reduce the effects of directionality and canopy heterogeneity on measurements, measurements were taken in four directions (N-S, E-W, NE-SW, and NW-SE) with each measurement centered in the middle of the AOI, as shown in Figure 2. A total of 32 measurements were acquired both above and below the canopy at each AOI. The average of above and below canopy measurements were used to calculate the fIPAR (Serrano et al. 2000), as shown in the following equation:
formula
(1)
where T is the PAR transmitted through the canopy (i.e., the below canopy PAR measurement), and S is the incident PAR at the canopy (i.e., or the above canopy PAR measurement).

UAS-based monitoring approach

Flight mission planning

All flights in this study were conducted under 14 CFR Part 107, with the UAS operated by a Remote Pilot in Command (RPIC) holding a Remote Pilot certificate with a Small UAS rating. When deemed necessary, a visual observer was used to assist the RPIC in maintaining safe flight operations and ensuring visual line of sight to the remote aircraft at all times.

While creation of orthomosaics is well understood for large agricultural operations, capture of small urban targets at high resolution presents some unique challenges. There are two primary regulatory constraints to operating UAS in urban areas. The first is that urban areas – particularly those served by major airports – often contain controlled airspace below 400 ft, sometimes extending down to the ground surface. Fortunately, the FAA's Low Altitude Authorization and Notification Capability (LAANC) has greatly facilitated the process of applying for and obtaining an airspace authorization. In some cases, the more restrictive regulation related to operating in urban areas is the stipulation that: ‘No person may operate a small unmanned aircraft over a human being unless that human being is: (a) Directly participating in the operation of the remote aircraft or (b) Located under a covered structure or inside a stationary vehicle …’ (FAA 2016b). Because urban areas tend to contain large numbers of pedestrians and moving (i.e., nonstationary) vehicles, small UAS (sUAS) operation under Part 107 is significantly more challenging than in rural areas. One option is to apply for a Part 107.39 waiver. Depending on the area and the flexibility in scheduling, it may also be possible to schedule sUAS operations for times when pedestrian and vehicle traffic is restricted, such as when an area is blocked off for construction. It is recommended that municipalities adopting UAS technology for stormwater management facility monitoring maintain a formal UAS program with documented operational procedures and responsibility for managing waivers and establishing safety plans.

Another challenge is that there may be obstacles (e.g., trees, buildings, and poles) within or near the scan area. The RPIC must identify these and determine their height to establish a safe altitude for operations. Once a safe flight altitude is established, it is necessary to verify that the desired ground sample distance (GSD) can be obtained from that altitude or higher. For GI site monitoring, the limiting factor in establishing the GSD is typically the need to be able to identify individual plants in the imagery. The required altitude to achieve this GSD is a function of the physical pixel size on the camera chip and the focal length of the camera lens. Because of the complex and diverse geometry of GI vegetation, it is important to capture with a large image overlap setting in order to maximize the number of tie points. For this research, 95% overlap (endlap and sidelap) was used. For flight at low altitude with tight turnarounds, a low flight speed (1.5–3 m/s) is appropriate. This limits motion blur and allows smoother turns.

Once programmed in the mission planning software, the UAS is generally able to automatically execute the pre-planned flight plan, with no direct input from the RPIC. However, the RPIC is required to ensure that the aircraft is ready for flight in all respects and to be prepared to handle any unexpected incidents aloft. In particular, it is important to ensure that the UAS compass is well calibrated on site. Failure to do so may lead to skewed images and unsteady flight. The RPIC must be prepared to immediately assume direct control if they judge that the vehicle is at risk – many UAS obstacle avoidance systems are disabled during autonomous flight.

Data acquisition and processing

Near-Infrared (NIR) and Red-Green-Blue (RGB) imagery were simultaneously collected with drones on three occasions (see Table 1), on June 1, 2019, June 29, 2019, and July 27, 2019. The flight on each day occurred within 2 hours of noon. The imagery collected on June 1, 2019 was acquired simultaneously with the built-in camera on a DJI Mavic Pro UAS and a MAPIR Survey3N NIR camera mounted externally on the Mavic Pro (DJI 2017a; MAPIR 2017). The image sets on June 29, 2019 and July 27, 2019 were acquired similarly with a DJI Phantom 4 Pro UAS with the Survey3N NIR camera mounted externally on the Phantom 4 Pro (DJI 2017b). In addition to the dates above, RGB imagery was also collected using the Mavic Pro on June 15, 2019, so that acquisition of ground-based fIPAR data coincided with dates and times of drone flights.

Imagery was acquired at an approximate altitude of 21 m above the rain garden at 1.8 m/s with 95% endlap and sidelap in a lawn mower pattern. These flight parameters enabled acquisition of high-resolution imagery with minimal motion blur and greater than the minimum amount of overlap required for structure from motion (SfM) photogrammetry, which is generally in the range of 75% front overlap (‘endlap’) and 60% side overlap (‘sidelap’) (Pix4D 2020). For safety purposes, before setting the operational flying height, the remote aircraft was first operated at a greater altitude to observe nearby obstacle heights. The highest relevant obstacle (in this case, a nearby tree) was found to be approximately 15.2 m tall, to which a safety factor of 1.4 was then applied to conclude a minimum operating altitude of 21 m. A low speed of 1.78 m/s was selected to minimize motion blur and to allow for clean turnarounds between the tightly spaced passes in the urban environment.

Tests were performed with passes aligned with the long and short dimensions of the survey area, with no observed difference in quality. The collected imagery was then processed using WebOpenDroneMap (Zecevic et al. 2017), an open-source SfM photogrammetry software package, to create orthomosaics with a GSD of 1 cm, with the resolution selected to enable individual plant-level analysis. The reason for investigating WebOpenDroneMap, as opposed to commercial SfM software, is that the methods of this study were deemed more likely to be implemented in practice, if the costs and relatively steep learning curves of commercial off the shelf (COTS) software could be avoided through the use of free, open-source, web-based options.

Radiometric calibration

In retrievals of NDVI from uncalibrated consumer-grade cameras on UAS, extraneous variables related to the imaging geometry and camera acquisition parameters (e.g., ISO, aperture, and shutter speed) can impact the results, if not appropriately accounted for (Berra et al. 2017). Although NDVI inherently performs a type of normalization (in fact, the ‘N’ in the acronym stands for ‘normalized’), it is insufficient to account for large changes in camera and environmental parameters. This is particularly true when the RGB (red, green, and blue) and NIR imagery are acquired from separate cameras, with independently modified acquisition settings, and when acquisition is performed on different days, as done in this study. By performing radiometric calibration, it is possible to reduce the impacts of these extraneous variables and obtain pixel values that are more closely related to true surface reflectance in different image bands and, hence, NDVI values that can be meaningfully compared across time and space. The next two sub-sections provide an overview of how radiometric calibration was performed in this study.

Spectral data acquisition

Reflectance spectra were acquired with the OceanOptics Flame-S spectrometer and the OceanView software (Ocean Optics 2013, 2015). The instrument covers the VIS and NIR spectral range (350–1,000 nm), has an optical resolution (full width at half maximum) of 1.34 nm, and a sampling interval of 0.37 nm. An optical fiber with a 25° field of view transfers light into the spectrometer.

Calibration targets were selected to meet the requirements described in Smith & Milton (1999) and Haghighattalab et al. (2016). Specifically, matte ethylene-vinyl acetate (EVA) foam panels were used. These are commonly used calibration targets, since they are inexpensive and act as diffuse reflectors with highly Lambertian behavior and horizontal homogeneity (Jeong et al. 2018; Padro et al. 2018).

The calibration procedure was performed within 2 hours of noon on a clear, sunny day at the study site using 24 cm × 24 cm white, gray, and black matte EVA foam calibration panels mounted on black canvas and separated by several centimeters to reduce adjacency and background effects (Padro et al. 2018). These panels are many times larger than the pixel sizes (1 cm/pixel) of the imagery used for calibration. Reflectance measurements were taken from 5 to 8 cm above each panel at a 45° angle (Milton et al. 2009; Pe'eri et al. 2013). Prior to measurements of each panel reflectance, dark measurements and diffuse white reflectance standard measurements were taken. Reflectance measurements were normalized with the white reference standard, or the light incident on the calibration panel surface.

Seven to eight measurements were collected from different angles above each panel between 11:10 am and 11:20 am local on a clear, sunny day on June 15, 2019. This reflectance data set was used to calibrate imagery from three dates June 1, June 29, and July 27, 2019 with the assumption that no changes to the calibration panels occurred through this time span that would affect reflectance. To shield the instrument from large temperature changes, which could potentially impact the acquired spectral measurements, the instrument was placed in an insulated housing during data collection. The temperature inside the insulated housing was recorded immediately preceding the start of spectra collection for the white calibration panel, between the collection of spectra for each panel, and after all spectra were collected. The panels were stored indoors in a dry, climate-controlled location between the field acquisition dates.

Empirical line method

To perform the empirical line method (ELM), the mean reflectance value corresponding to the spectral ranges of the UAS camera image bands was found and plotted for each calibration target (Wang & Myint 2015). Since manufacturer-provided spectral sensitivity curves for the built-in RGB cameras on the UAS were not available, typical spectral ranges for complementary metal–oxide–semiconductor (CMOS) image sensors were assumed: 580–670 nm for the red band, 480–610 nm for the green band, and 400–520 nm for the blue band (Hunt et al. 2013). For high-accuracy applications, future work could include a step of empirically determining the spectral sensitivity curves of the camera bands, but, for purposes of this study, the impacts of assuming default (typical) spectral ranges were deemed negligible; a difference of up to ±30 nm in the assumed spectral range of the red band was calculated to produce only a 0.52–3.29% difference in the reflectance output from the ELM. In contrast to the RGB camera, the center wavelength corresponding to the MAPIR Survey3N NIR camera is published: 850 nm (Ramseyer 2017).

The mean digital numbers (DNs) were extracted from the calibration panels in the orthophotos for each of the four bands. Only the central pixel values of each calibration target were extracted from the orthophotos to avoid boundary effects (Smith & Milton 1999; Wang & Myint 2015; Haghighattalab et al. 2016). The mean DN of each target in the orthomosaics was plotted against the mean reflectance value obtained from the spectrometer measurements for each camera waveband, where each point in the plot represents a different calibration target (Wang & Myint 2015). The resulting regression models were used to determine the calibration parameters (the y-intercepts) for each waveband (Wang & Myint 2015). The equations were then applied to the individual bands of the orthophotos.

Normalized difference vegetation index

The NIR and RGB orthomosaics generated in the previous step were georeferenced using control points at the four corners of the bioswale facility, and the bands were combined to create NDVI maps. The NDVI, Equation (2), is the normalized difference between the NIR and the visible red band. This equation can either be formulated in terms of the raw DNs in the red and NIR image bands, or in terms of surface reflectance, , computed from the raw DNs, as discussed in Section 3.2.1 (Gamon et al. 1995):
formula
(2)

Healthy leaves produce an abundance of chlorophyll pigments, which absorb highly in the red part of the spectrum (Knipling 1970). However, chlorophyll pigments have no effect on the absorption or reflectance in the NIR. Instead, the structure of the leaves, specifically the spongy mesophyll and air cavities, is responsible for the high reflectance in the NIR part of the spectrum, where NIR light is scattered and reflected (Knipling 1970).

NDVI values are floating point numbers ranging from −1 to 1, where negative values indicate the presence of water or possibly wet soil, values close to zero indicate soil or dead vegetation, values in the low positive range indicate stressed vegetation, and values in the high positive range represent healthy vegetation (Weler & Herring 2000).

Data analysis

Regression analysis

The relationship between the NDVI derived from the UAS imagery and the ground-based measurements of the fIPAR was explored by overlaying the orthomosaic and the corresponding fIPAR measurements acquired at the end of June. The mean of NDVI pixel values falling within each AOI was extracted, with each AOI of approximate area 0.5 m2 containing between 4,000 and 5,000 pixels in the three NDVI images. Figure 3 shows the location of 18 AOIs that were sampled on June 15 in bioswale Cell 3 and the location of 41 AOIs that were sampled on June 29–30 in the two bioswale cells – Cell 2 and Cell 3. Since NIR imagery was not collected on June 15, 2019, a regression analysis was performed on the data collected at AOI locations only on June 29–30, 2019 to explore the correlation between fIPAR and NDVI for the two cells in the bioswale facility. The data points were grouped by the cell they were located inside, as well as based on which sedge, rush, or broad leaf vegetation groups the data point represented. A regression analysis was then performed separately on each group.

Figure 3

Locations of data points and AOIs in bioswale Cells 2 and 3 overlaid on UAS-derived RGB imagery collected on June 15, 2019 (top) and June 29, 2019 (bottom).

Figure 3

Locations of data points and AOIs in bioswale Cells 2 and 3 overlaid on UAS-derived RGB imagery collected on June 15, 2019 (top) and June 29, 2019 (bottom).

Spatial analysis

NDVI and fIPAR values at different locations were examined for spatial trends related to proximity of locations to the bioswale cell inlets (overflow sediment bay in Figure 3), the primary source of water during the spring growing season. As water is conveyed down the length of the cells (toward the north direction), it infiltrates and enters the underdrain. Though drip irrigation is used during the summer as a vegetation maintenance intervention in these bioswale cells, the plants in the far north end of the cells generally receive a lower annual amount of stormwater in comparison to the end closer to inlets in sediment bay. To determine if the proximity to the inlet is related to the health of vegetation, the values of NDVI and fIPAR were related to the distance from the inlet. The relationship between fIPAR and the distance from the inlet was examined for all sampling locations.

Temporal trend analysis

The temporal trend analysis examines NDVI data from the summer of 2019, with images acquired monthly at the end of May, June, and July. NDVI data within set ranges of values were extracted from each of the three bioswale cells to examine how the area of coverage changes for each NDVI range over the course of the 2 months. The ranges of NDVI values explored were 0.6–1.0, representing healthy vegetation, 0.2–0.6, representing unhealthy or stressed vegetation, −0.2 to 0.2, representing soil and dead vegetation, and −1.0 to −0.2, representing a complete absence of vegetation. These ranges were empirically selected to represent vegetation health status and the presence of soil or lack of vegetation for this particular study site.

RESULTS AND DISCUSSION

Ground-based monitoring of fIPAR

Spatial analysis

Figure 4 shows the classified fIPAR values overlaid on RGB orthomosaics for one cell (Cell 3) of the bioswale facility on June 15, 2019 and two cells (Cells 2 and 3) on June 29–30, 2019. The GNSS data for Cell 2 gathered on June 15 were unreliable (post analysis revealed that the receiver was likely not initialized) and, therefore, were not included in the fIPAR analysis. The Cell 3 sampling locations were refined on June 29 to maximize the coverage of plant groups. Figure 5 graphs these fIPAR values in relation to the distance from the inlet to each cell. When examining the fIPAR in relation to the distance from the inlet, there is very little change as the distance between the sample points and the inlet increases. The range of fIPAR values was found to be above 0.50, or more than 50% of light is intercepted by vegetation, on both dates, with no values being lower than 0.60 on June 29–30, 2019 and no values lower than 0.52 on June 15, 2019. Furthermore, the fIPAR values were found to not be consistently high for vegetation that visibly appeared healthy or low for vegetation that visibly appeared unhealthy. Although vegetation visually appears unhealthy, many species of grasses are dormant in the summer months, indicating that these species may retain density during dormancy. Hence, without additional differentiation between unhealthy and dormant vegetation and without an understanding of how the dormancy of vegetation species present in these facilities can affect vegetation density, we found that fIPAR values cannot be reliably alone used as an indicator of vegetation health. Other studies have also reported challenges in using fIPAR for sampling mixed vegetation in various growth stages and can have very different values that are dependent on the plant species (Hatfield et al. 1984; Serrano et al. 2000).

Figure 4

Classified fIPAR at 18 AOIs for June 15, 2019 and 41 AOIs for June 29–30, 2019 overlaid on RGB orthomosaics from UAS-derived imagery obtained on June 15 (left) and June 29 (right).

Figure 4

Classified fIPAR at 18 AOIs for June 15, 2019 and 41 AOIs for June 29–30, 2019 overlaid on RGB orthomosaics from UAS-derived imagery obtained on June 15 (left) and June 29 (right).

Figure 5

fIPAR values related to the distance from the inlet, the primary water source.

Figure 5

fIPAR values related to the distance from the inlet, the primary water source.

Sedges and rushes are common species in GI and are present in the OGSIR facility. The mean fIPAR value obtained on June 29–30, 2019 for a sample of eight sedge plants (including slough and dense sedge species) in the OGSIR facility, as shown in Figure 6(a), was found to be 0.92. The mean fIPAR value for seven rush plants (including hardstem bulrush and spreading rush species) sampled in the facility at the same time, as shown in Figure 6(b), was found to be 0.85. Previously reported fIPAR values in the literature for slough sedge, hardstem bulrush and poverty rush include those that were sampled from natural stands in four different wetlands in the United States. These values for two wetlands in Texas, one in Delaware, and one in North Dakota were found to be 0.67, 0.32, and 0.47, respectively (Williams et al. 2017). Note that these values obtained in the natural wetlands are much lower than those obtained in OGSIR's engineered environment. This indicates that sedge and rush vegetation could be denser in engineered systems than in natural systems. Since fIPAR values of wetland species or species commonly used in urban GI are seldom reported in the literature, the use of values reported for natural wetlands may lead to underestimation of plant growth parameters.

Figure 6

Locations of sampled sedge vegetation shown circled in white in (a), and sampled rush vegetation shown circled in yellow in (b) in the OGSIR facility and associated fIPAR values collected on June 29–30, 2019. Please refer to the online version of this paper to see this figure in colour: http://dx.doi.org/10.2166/hydro.2020.195.

Figure 6

Locations of sampled sedge vegetation shown circled in white in (a), and sampled rush vegetation shown circled in yellow in (b) in the OGSIR facility and associated fIPAR values collected on June 29–30, 2019. Please refer to the online version of this paper to see this figure in colour: http://dx.doi.org/10.2166/hydro.2020.195.

Method evaluation

There are many practical challenges in obtaining reliable PAR results in bioswale cells, potentially resulting in biased fIPAR measurements. At the OGSIR facility, the cross-section geometry of the topography of bioswale cells is trapezoidal with side slopes of 4:1 (H:V). Obtaining level PAR measurements in four directions beneath the canopy of shrubs requires compensating for this side slope by raising the Accupar wand and obtaining measurements within instead of entirely beneath the canopy. This also occurs close to the bioswale cell walls, where the height of the walls requires raising the wand to acquire level measurements. This results in an underestimate of fIPAR, where the Accupar sensors are not able to detect the radiation intercepted by the entire canopy, but only part of the canopy. The opposite situation can occur when obtaining measurements next to wells and other obstructions in the bioswale cells. In this case, some AOI measurements include shadows from these obstructions, creating an overestimate of the fIPAR at the AOI.

Challenges also occur from structural differences between species and from the use of a GNSS unit to obtain coordinates at AOI center points. The rushes and sedges form dense clumps at the ground level instead of woody stems common below the canopies of shrubs. Pushing the Accupar wand through the clump at the base of sedges and rushes disturbs the plant, changing the structure of the plant and the resulting PAR measurements. Placing the GNSS antenna pole into the center of plants selected for AOIs also disturbs the plant canopy and acts as an obstruction, casting a shadow during measurement. While care was taken to minimize disruption to the vegetation when obtaining measurements, some areas of the facility have very dense vegetation where it is necessary to disturb vegetation in order to place the instruments, potentially resulting in biased measurements.

To collect data for 41 AOIs in this small GI facility on June 29 took approximately 4 hours of field labor with one additional hour of processing time. This leads to approximately 6 min of field labor per plant. The larger the GI facility, the more manpower and more instruments would be necessary to acquire sufficient data to use this method for monitoring GI vegetation health. The combined cost of purchasing the Accupar LP-80 and GNSS smart antenna and controller for a small facility, and the manpower and additional cost of more instruments required to scale this method to larger GI facilities reduces the feasibility of using this approach for routine maintenance and monitoring operations.

UAS-based monitoring approach

Figure 7 shows the orthomosaics generated with WebODM from the imagery acquired on June 1, 2019, June 29, 2019, and July 27, 2019. The June 1 and June 29 image sets were acquired in full sun, while the July 27 image set was acquired in overcast conditions. The images have a resolution of 1 cm/pixel, with minimal visible distortion. The camera settings were constant between the flights with the same UAS configuration, but the light conditions were different on each date resulting in the dissimilar brightness of orthomosaics at the sampling dates.

Figure 7

RGB and NIR orthomosaics generated from image sets acquired on three dates showing the difference in coloration and brightness between the two cameras (Mavic Pro RGB on June 1, 2019 and Phantom 4 Pro RGB on June 29 and July 27, 2019) and with the same NIR camera with different environmental conditions.

Figure 7

RGB and NIR orthomosaics generated from image sets acquired on three dates showing the difference in coloration and brightness between the two cameras (Mavic Pro RGB on June 1, 2019 and Phantom 4 Pro RGB on June 29 and July 27, 2019) and with the same NIR camera with different environmental conditions.

Radiometric calibration

Reflectance spectra

Eight reflectance spectra were recorded for each white, gray, and black calibration panel on June 15, 2019 at the OGSIR site. To reduce noise, the means of the eight observed reflectance spectra were computed for each panel, and the resultant mean reflectance spectra (Figure 8) were used in all subsequent steps. To evaluate the consistency between the observed spectra for each panel, the standard deviations were also computed and found to be low: ±3.0, ±1.2, and ±0.2 percentage points of reflectance for the white, gray, and black calibration panels, respectively.

Figure 8

Mean reflectance curves for the white, gray, and black calibration panels.

Figure 8

Mean reflectance curves for the white, gray, and black calibration panels.

Image DNs

The standard deviation of the DN values ranges from 0.36 to 3.77, or 0 to 1.48% of the 8-bit DN value range (0–255). Jeong et al. (2018) tested commercial calibration panels and found that the standard deviation of DN values ranged from 0 to 12.66, or 0 to 4.96% of the 8-bit DN value range. The DN values for these panels on all dates are well within this range. The low standard deviation of calibration panel reflectance data and DN values in the two image bands confirms that the panels exhibit Lambertian properties.

Empirical line method
Figure 9 shows the scatter plots and the two linear regression models examined for converting image DNs in the red and NIR wavebands to reflectance for the June 1, June 29, and July 27, 2019 data set. The relationship between surface reflectance and image DNs is not always linear, as shown by Wang & Myint (2015). However, higher-order polynomials can lead to overfitting, and with only three calibration targets in this study, a linear fit was deemed most appropriate. The relationship between the image DNs and the surface reflectance of the calibration panels is given in the following equation and illustrated in Figure 9:
formula
(3)
where Ri is the converted reflectance of band i pixels, DNi is the uncalibrated DN of band i, and a and b (the offset and scale factor, respectively) are the parameters determined through the linear regression (Wang & Myint 2015; Jeong et al. 2018).
Figure 9

The linear relationship between calibration panel surface reflectance and image DN for the three flight dates, June 1, June 29, and July 27, 2019.

Figure 9

The linear relationship between calibration panel surface reflectance and image DN for the three flight dates, June 1, June 29, and July 27, 2019.

Each of these calibration equations in Figure 9 has a negative y-intercept. To avoid nonsensical negative reflectance values, the red-band DNs below 54.61, 75.48, and 78.67 and the NIR-band DNs below 9.03, 6.85, and 4.45 for the June 1, 2019, June 29, 2019, and July 27, 2019 data sets, respectively, were all set to the lowest possible value of zero reflectance.

Method evaluation

Acquiring UAS imagery in the field involved minimal time and effort, with each flight mission for scanning all bioswales at the OGSIR facility being completed under 1 h. Each mission included setup time and break-down time, and acquisition of one to two backup data sets. The two UAS used for the three flights produced good quality data, with minimal distortion in the final orthomosaics obtained from both the DJI Mavic Pro and the DJI Phantom 4 Pro data sets. Data processing times with the structure-from-motion software WebODM varied by camera, with the NIR data sets processing requiring up to 1–2 h and the RGB data sets processing requiring 12–24 h. Data were processed with a 2.20 GHz AMD A8-7410 APU. Obtaining a validation set of EVA foam panels is inexpensive with EVA foam panels frequently sold in packages with multiple panels of various colors.

Normalized difference vegetation index

Figure 10 shows classified NDVI maps of the OGSIR bioswale facility for the three dates with the same 10 classes, where there are five classes from −1.0 to 0.0 and five from 0.0 to 1.0. The spatial distribution of each class is shown in the figure. The shades of green represent healthy vegetation, yellow represents senescent or fading vegetation, orange and brown represent soil and dead vegetation, and red represents no vegetation. The values close to −1 are in the gravel/rock on the side of the rain garden. Cell 1 contains only weeds, showing the effect they have on the NDVI in all cells in the data sets. Once these weeds die, the remaining green areas in Cells 2 and 3 represent the original vegetation planted in the facility in 2014. The largest concentration of green values, or healthy vegetation on June 29 and July 27, is close to the inlet on the south end of the facility, with the coverage of green values decreasing at one-third of the length of the facility and continuing to decrease moving away from the inlet.

Figure 10

Classified NDVI data for the three dates showing the spatial distribution of NDVI values in bioswale cells. Please refer to the online version of this paper to see this figure in colour: http://dx.doi.org/10.2166/hydro.2020.195.

Figure 10

Classified NDVI data for the three dates showing the spatial distribution of NDVI values in bioswale cells. Please refer to the online version of this paper to see this figure in colour: http://dx.doi.org/10.2166/hydro.2020.195.

Spatial analysis

Figure 11 shows the NDVI change in relation to the distance from the inlet. The drop in the NDVI values for the entire area of Cell 1 represents the change in weed coverage as water availability and the temperature changed from June 1 to later in summer (June 29 and July 27). In addition to Cell 1, the high NDVI values in Cell 2 and Cell 3 on June 1, 2019 also coincide with the extensive coverage of weeds. Once the weeds died in Cell 2 and Cell 3 later in the summer, the NDVI values in these cells also experienced a downward trend, especially at locations further away from the inlet. NDVI values in Cell 2 stayed between 0.45 and 0.75 within 12 m from the Cell 2 inlet, with values falling below 0.4 at distances larger than 15 m from the inlet. NDVI values stayed between 0.5 and 0.8 within 9 m from the Cell 3 inlet, with values falling below 0.4 at distances larger than 18 m from the inlet.

Figure 11

NDVI related to distance from the inlet for Cells 1, 2, and 3.

Figure 11

NDVI related to distance from the inlet for Cells 1, 2, and 3.

Note that eight of the 41 AOIs that were used for ground-based monitoring were at sedge locations, while seven were at rush locations (Figure 6(a)). The mean NDVI of the eight sedges estimated using the UAS-based monitoring approach was found to be 0.82, 0.78, and 0.82 on June 1, June 29, and July 27, respectively, while the mean NDVI for the rush locations was 0.67, 0.79, and 0.69 on the same three dates. Though we recommend additional soil and plant-tissue sampling in future studies to explore this relationship for GI systems, strong relationships between NDVI, biomass, soil, and water have been reported for natural wetlands (e.g., Boelman et al. 2003; Goswami et al. 2015). Considering these previous studies, the high NDVI values indicate dense vegetation and a correspondingly high biomass in the OGSIR facility, especially close to the inlet. Changes in NDVI-based biomass estimates over time can indicate the presence of stressors and spatial variations in these estimates can indicate stress in vegetation caused by spatial variability in the plant's environment. Because NDVI is a function of not only environmental stressors (e.g., drought) but also plant type, it is important to monitor NDVI for multiple species within a stormwater facility over time in order to best inform management decisions.

Temporal trend analysis

Figure 10 also indicates a visible change in healthy vegetation coverage from the beginning of June through the end of July. To examine temporal trends in vegetation health, we calculated the area covered by healthy vegetation, unhealthy or stressed vegetation, soil and dead vegetation, and a complete absence of vegetation. NDVI values ranging from 0.6 to 1.0 in Cell 2 covered 60 m2 of the facility on June 1 and decreased by 44% by June 29. NDVI values close to zero and below zero covered 6 m2 on June 1 and increased by 83% by June 29. The areas covered by these NDVI values in Cell 2 remained close to constant during July with changes in the area of 10% or less in all classes. NDVI values ranging from 0.6 to 1.0 in Cell 3 covered close to 74 m2 on June 1 and declined by 48% by June 29. NDVI values indicating soil, dead vegetation, and no vegetation in Cell 3 covered 6 m2 on June 1 and increased by 84% by June 29. Changes in the area of 11% or less occurred in Cell 3 during July in all classes. NDVI values ranging from 0.6 to 1.0 in Cell 1, which contained only weeds, covered an area of 32 m2 on June 1 and decreased by 80% by June 29. NDVI values close to zero and below zero covered an area of 14 m2 on June 1 and increased by 81% by June 29. The area covered by NDVI classes in Cell 1 changed by 21–28% during July, with classes representing stressed vegetation or no vegetation showing an increase in area and classes representing healthy vegetation and dying vegetation showing a decrease in area.

Method evaluation

The NDVI data generated from the calibrated UAS image sets proved useful in detecting changes in the area covered by vegetation classes over time and in exploring the spatial distribution of NDVI values. However, an important limitation to be aware of is that NDVI generated from different sensors cannot be directly compared without examining the spectral sensitivity regions and band peaks included in the bandpass filters of each CMOS sensor (Miura et al. 2006). The CMOS sensors in the Mavic Pro and Phantom 4 Pro likely have differing spectral ranges in the red spectral response curve and differing red peaks. Unless the differences are minor, they will affect the resulting NDVI values. It is important to be aware that many manufacturers, including DJI, do not make the spectral response curves available to the public. Finally, while we attempted to account for this limitation in this study by using the radiometric calibration panels and the ELM to convert to surface reflectance, additional studies may provide insight on the impacts of using different cameras without the radiometric calibration step.

Zhang et al. (2015) found that shadows cause an overestimate of NDVI, but when computing vegetation parameters from NDVI, such as vegetation fractional coverage, the results were not strongly affected by the presence of shadows. A side-by-side comparison of the June 29 RGB orthomosaic and the June 29 classified NDVI map indicated that part of the area covered by the highest NDVI values is shaded. While some of the areas may correspond to healthy vegetation cast in shadows and some represent shadows cast by obstructions in the facility in areas of dry or dead vegetation, these areas have very high NDVI values indicating an overestimate of NDVI.

There are many methods for detecting shadows in imagery, but histogram thresholding is the simplest method and assumes that shadow and sunlit areas have different histogram levels (Adeline et al. 2013). Increased accuracy can be reached by applying a shadow-detection methodology to the image bands prior to generating the NDVI data. However, UAS flights should be planned to reduce the presence of shadows by performing flights close to noon and considering the effect of seasonality on shadow size. Even without removing shaded areas from the analysis, this method allows for tracking changes in facility vegetation once a baseline of health is established. Deviations from this baseline should trigger a maintenance plan of action.

This is a promising method that could be used toward long-term periodic monitoring of facilities. With access to image analysis software, like ESRI ArcMap, the NDVI is easy to compute and analysis provides useful information on the health of vegetation in GI. The method can be improved using higher quality sensors designed for NDVI generation and using shadow-detection methods. Consistently using the same sensors for each flight ensures that resulting NDVI data will be comparable across a temporal data set, eliminating the need to generate cross-sensor models to translate reflectance from one sensor to match reflectance from another (Miura et al. 2006). Using a sensor that has both red and NIR bands also reduces the need for radiometric calibration using the ELM. However, the NDVI results obtained from inexpensive sensors, calibration methodology, and NDVI classification performed in this study provide a good idea of vegetation health with the potential to inform long-term maintenance decisions.

Comparison of ground-based and UAS-based monitoring of bioswales

Figure 12 shows the scatter plots and regression models to determine the NDVI relationship to fIPAR for sedge (Figure 12(a)), rush (Figure 12(b)), and broad leaf (Figure 12(c)) vegetation groups in the bioswale cells, using the data collected on June 29, 2019. It appears that the broad leaf and sedge NDVI-fIPAR relationship is stronger with coefficients of determination (R2) of 0.64 and 0.55, with a very weak relationship for the rush group (R2 = 0.12). Since the vegetation groups have small sample sizes, there is a potential for outliers to affect the regression model. For example, in Figure 12(c), the low point (bottom left of the figure) appears to have both high influence and leverage on the regression analysis. This outlier may be a function of the small sample size. Future studies are recommended to take this issue into consideration when collecting additional fIPAR samples of each vegetation group, so that the relationship for different types of stormwater GI vegetation species can be determined more accurately.

Figure 12

Scatter plots and regression models of the NDVI relationship to fIPAR of sedge, rush, and broad leaf vegetation groups including vegetation from both cells of the OGSIR bioswale facility for data collected on June 29–30, 2019.

Figure 12

Scatter plots and regression models of the NDVI relationship to fIPAR of sedge, rush, and broad leaf vegetation groups including vegetation from both cells of the OGSIR bioswale facility for data collected on June 29–30, 2019.

Previous studies have shown a strong relationship between radiometer-derived NDVI and ground-based measurements of fIPAR in wheat crops during the growth phase with coefficients of determination of 0.81 and 0.97 (Hatfield et al. 1984; Serrano et al. 2000). Strong relationships were also found between UAS-derived NDVI and ground-based measurements of fIPAR in peach and citrus orchards with coefficients of determination of 0.88 and 0.85, respectively (Guillen-Clement et al. 2012). However, these relationships were found for agricultural study sites with the relationship examined for a single crop. In contrast, the bioswale facility at the OGSIR facility contains mixed vegetation and presents unique challenges related to accessing the sampling point and taking consistent measurements in densely planted areas and varying canopy types.

CONCLUSIONS AND FUTURE WORK

This study examined the feasibility of using UAS for efficient, multitemporal monitoring of vegetation in urban bioswales. UAS imagery data sets were acquired with two remote aircrafts, and the data were processed in an open-access, web-based structure from motion (SfM) software utility to generate 1-cm resolution orthomosaics. These orthomosaics were then radiometrically calibrated to surface reflectance measurements to reduce any inaccuracies caused by the camera parameters and illumination conditions and to ensure that the multiple sensors represent the actual reflectance. This is especially important when comparing and combining data from multiple sensors and when comparing data from multiple dates. Calibration was performed via the ELM using spectral reflectance data of three calibration panels using a field spectrometer.

NDVI maps were generated from the calibrated red and NIR bands of the orthomosaics. Three NDVI maps were examined for spatial distribution and temporal changes. Cell 1 contained only weeds during the study time period, while Cells 2 and 3 contained vegetation originally planted in the facility and seasonal weeds. High NDVI values dominated all three cells on the early June NDVI map, representing the abundance of healthy weeds at the end of spring. At the end of June and July, the seasonal weeds had died and the high NDVI values corresponded to only the healthy bioswale vegetation originally planted in the facility.

Spatially, mean NDVI values in Cells 2 and 3 were found to be highest in the first 9–12 m measured down the length of the facility from the inlet, or the third of the facility closest to the inlet. Cell 3 maintained higher NDVI values in this area of the facility than Cell 2, indicating that mixed vegetation remains healthier when close to the primary water source than when vegetation is limited to the sedges, rushes, and grasses of Cell 2. Most of the mean NDVI values in Cell 3 in the 9 m closest to the inlet were above 0.6, representing very healthy vegetation in this region. When monitoring the NDVI over time and spatially, knowledge of the facility growth patterns increases the ability to establish a baseline of vegetation health. Deviations in the baseline can trigger site visits by GI facility managers and technicians, further data acquisition, maintenance, and analysis of potential causes of the deviation. Overall, NDVI maps can prove valuable for long-term monitoring of GI vegetation, especially in large GI facilities. Furthermore, this study concluded that UAS represents an efficient and cost-effective method of obtaining the NDVI data.

A second component to the study involved comparing the UAS-derived NDVI data against fIPAR measurements collected using a hand-held ceptometer to see if the former could be used to predict the latter. fIPAR measurements were obtained at multiple sampling locations in Cells 2 and 3 of the OGSIR bioswale facility. Results from this portion of the study revealed no statistically strong relationship between fIPAR and UAS-based NDVI. The practical challenges in obtaining large samples of consistent measurements using the fIPAR methods likely influenced the exploration of a relationship between the two parameters. The bioswale facility is inherently different from the agricultural study sites investigated in the previous literature, with an irregular planting arrangement, mixed species in the canopy make-up, and unregulated growth and health. Agricultural sites, where fIPAR and NDVI have been previously explored, generally have regular rows of a single crop maintained for consistent growth and health. The structure of agricultural sites promotes consistency in data collection methods with fairly controlled conditions. A lack of a relationship between NDVI and fIPAR precludes the use of NDVI results as a proxy for fIPAR in biomass productivity models of bioswale and similar urban GI facilities. However, the NDVI derived from calibrated imagery is a valuable reflectance-based vegetation health parameter and as a stand-alone indicator of stress in vegetation.

In future studies, it may be of interest to test different methods of field measurement of fIPAR (specifically, methods that enable more consistent measurements across varying species, plant densities, and canopy types) and to then reassess the relationship between field-measured fIPAR and UAS-derived NDVI. More consistent biomass measurements can be obtained using destructive methods where vegetation samples cut at ground level are dried, weighed, and used to represent the biomass of all facility vegetation (Prabhakaraa et al. 2015). Additionally, allometric equations can be used to determine above-ground biomass. Allometric equations for bioswale vegetation can be developed from biometric variables unique to vegetation species including plant height, diameter, crown area, and crown shape (Ali et al. 2015).

Another recommended extension of the methods of this study is to further investigate UAS mission planning to reduce shadows and the associated impacts on NDVI. With flights planned for the same time of day using the same flight path for individual facilities, the effect of shadows becomes dependent on seasonal changes in the sun angle. Future research is recommended to investigate how integrated UAS with additional cameras and machine learning could assist with adaptive scanning of areas of interest in bioswales. For large-scale implementation of UAS-based monitoring of urban GI facilities, additional work could also focus on operational efficiency and effectiveness, especially improving cycle times, reducing pilot workloads, and managing evolving regulations. A final recommendation for future work is to utilize additional calibration panels in the radiometric calibration procedure, to improve the regression models.

ACKNOWLEDGEMENTS

We would like to convey gratitude to students in the Hydroinformatics Research Group, with special thanks to Mr Azad Dazaea, for their assistance with collection of data. Part of this research was financially supported by Award ER18-C3-1230 funded by Strategic Environmental Research and Development Program (SERDP) at U.S. Department of Defense.

DATA AVAILABILITY STATEMENT

All relevant data are included in the paper or its Supplementary Information.

REFERENCES

Adeline
K. R. M.
,
Chen
M.
,
Briottet
X.
,
Pang
S. K.
&
Paparoditis
N.
2013
Shadow detection in very high spatial resolution aerial imagers: a comparative study
.
ISPRS Journal of Photogrammetry and Remote Sensing
80
,
21
38
.
doi:10.1016/j.isprsjprs.2013.02.003
.
Akpor
O. B.
&
Muchie
M.
2010
Remediation of heavy metals in drinking water and wastewater treatment systems: Processes and applications
.
International Journal of the Physical Sciences
5
(
12
),
1807
1817
.
Ali
A.
,
Xu
M. S.
,
Zhao
Y. T.
,
Zhang
Q. Q.
,
Zhou
L. L.
,
Yang
X. D.
&
Yan
E. R.
2015
Allometric biomass equations for shrub and small tree species in subtropical China
.
Silva Fennica
49
(
4
),
10
.
Article id 1275
.
doi:10.14214/sf.1275
.
Allahyari
M.
2016
Accuracy Evaluation of Real-Time GNSS Survey Observations
.
M.S. Thesis
,
Department of Civil and Construction Engineering, Oregon State University
,
Corvallis, OR
.
Allahyari
M.
,
Olsen
M. J.
,
Gillins
D. T.
&
Dennis
M. L.
2018
Tale of two RTNs: rigorous evaluation of real-time network GNSS observations
.
Journal of Surveying Engineering
144
(
2
),
05018001
.
Bartesaghi-Koc
C.
,
Osmond
P.
,
Peters
A.
&
Matthias
I.
2017
A methodological framework to assess the thermal performance of green infrastructure through airborne remote sensing
.
Procedia Engineering
180
,
1306
1315
.
doi:10.1016/j.proeng.2017.04.293
.
Bartesaghi-Koc
C.
,
Osmond
P.
&
Peters
A.
2019
Mapping and classifying green infrastructure typologies for climate-related studies based on remote sensing data
.
Urban Forestry & Urban Greening
37
,
154
167
.
doi:10.1016/j.ufug.2018.11.008
.
Bauerle
W. L.
,
Weston
D. J.
,
Bowden
J. D.
,
Dudley
J. B.
&
Toler
J. E.
2004
Leaf absorptance of photosynthetically active radiation in relation to chlorophyll meter estimates among woody plant species
.
Scientia Horticulturae
101
,
169
178
.
doi:10.1016/j.scienta.2003.09.010
.
Berra
E. F.
,
Gaulton
R.
&
Barr
S.
2017
Commercial off-the-shelf digital cameras on unmanned aerial vehicles for multitemporal monitoring of vegetation reflectance and NDVI
.
IEEE Transactions on Geoscience and Remote Sensing
55
(
9
),
4878
4886
.
doi:10.1109/TGRS.2017.2655365
.
Boelman
N. T.
,
Stieglitz
M.
,
Rueth
H. M.
,
Sommerkorn
M.
,
Griffin
K. L.
,
Shaver
G. R.
&
Gamon
J. A.
2003
Response of NDVI, biomass, and ecosystem gas exchange to long-term warming and fertilization in wet sedge tundra
.
Oecologia
135
(
3
),
414
421
.
doi:10.1007/s00442-003-1198-3
.
Bréda
N. J. J.
2003
Ground-based measurements of leaf area index: a review of methods, instruments and current controversies
.
Journal of Experimental Botany
54
(
392
),
2403
2417
.
https://doi.org/10.1093/jxb/erg263
.
Cordon
G.
,
Lagorio
M. G.
&
Paruelo
J. M.
2016
Chlorophyll fluorescence, photochemical reflective index andnormalized difference vegetative index during plant senescence
.
Journal of Plant Physiology
199
,
100
110
.
doi:10.1016/j.jplph.2016.05.010
.
Dimitrov
S.
,
Georgiev
G.
,
Georgieva
M.
,
Gluschkova
M.
,
Chepisheva
V.
,
Mirchev
P.
&
Zhiyanski
M.
2018
Integrated assessment of urban green infrastructure condition in Karvolo urban area by in-situ observations and remote sensing
.
One Ecosystem
3
,
e21610
.
doi:10.3897/oneeco.3.e21610
.
DJI
.
2017a
Mavic Pro User Manual (V2)
.
DJI
. .
DJI
.
2017b
Phantom 4 Pro/Pro+ User Manual
.
DJI
. .
Federal Aviation Administration
.
2016a
14 CFR Parts 21, 43, 61, 91, 101, 107, 119, 133, and 183: Operation and Certification of Small Unmanned Aircraft Systems
.
U.S. Department of Transportation
.
Federal Aviation Administration
.
2016b
Summary of Small Unmanned Aircraft Rule (Part 107)
.
Available from: https://www.faa.gov/uas/media/Part_107_Summary.pdf (accessed 25 May 2020)
.
Franco
S. F.
&
Macdonald
J. L.
2018
Measurement and valuation of urban greenness: remote sensing and hedonic application to Lisbon, Portugal
.
Regional Science and Urban Economics
72
,
156
180
.
doi:10.1016/j.regsciurbeco.2017.03.002
.
Gadi
V. K.
,
Garg
A.
,
Prakash
S.
,
Wei
L.
&
Andriyas
S.
2018
A non-intrusive image analysis technique for measurement of heterogeneity in grass species around tree vicinity in a green infrastructure
.
Measurement
114
,
132
143
.
doi:10.1016/j.measurement.2017.09.010
.
Gamon
J. A.
,
Field
C. B.
,
Goulden
M. L.
,
Griffin
K. L.
,
Hartley
A. E.
,
Joel
G.
,
Penuelas
J.
&
Valentini
R.
1995
Relationships between NDVI, canopy structure, and photosynthesis in three California vegetation types
.
Ecological Applications
5
(
1
),
28
41
.
Goswami
S.
,
Gamon
J. A.
,
Vargas
S.
&
Tweedie
C. E.
2015
Relationships of NDVI, biomass, and leaf area index (LAI) for six key plant species in Barrow, Alaska
.
PeerJ PrePrints
.
doi:10.7287/peerj.preprints.913v1
.
Guillen-Clement
M. L.
,
Zarco-Tejada
P. J.
,
Berni
J. A. J.
,
North
P. R. J.
&
Villalobos
F. J.
2012
Mapping radiation interception in row-structured orchards using 3D simulation and high-resolution airborne imagery acquired from a UAV
.
Precision Agriculture
13
,
473
500
.
doi:10.1007/s11119-012-9263-8
.
Haghighattalab
A.
,
Perez
L. G.
,
Mondal
S.
,
Singh
D.
,
Schinstock
D.
,
Rutoski
J.
,
Ortiz-Monasterio
I.
,
Singh
R. P.
,
Goodin
D.
&
Poland
J.
2016
Application of unmanned aerial systems for high throughput phenotyping of large wheat breeding nurseries
.
Plant Methods
12
(
35
),
1
15
.
doi:10.1186/s13007-016-0134-6
.
Hatfield
J. L.
,
Asrar
G.
&
Kanemasu
E. T.
1984
Intercepted photosynthetically active radiation estimated by spectral reflectance
.
Remote Sensing of Environment
14
,
65
75
.
doi:10.1016/0034-4257(84)90008-7
.
Hill
D. J.
&
Babbar-Sebens
M.
2019
Promise of UAV-assisted adaptive management of water resources systems
.
Journal of Water Resources Planning and Management
145
,
2017
2020
.
doi:10.1061/(asce)wr.1943-5452.0001081
.
Hunt
E. R.
,
Doraiswamy
P. C.
,
McMurtrey
J. E.
,
Daughtry
C. S. T.
,
Perry
E. M.
&
Akhmedov
B.
2013
A visible band index for remote sensing leaf chlorophyll content at the canopy scale
.
International Journal of Applied Earth Observation and Geoinformation
21
,
103
112
.
doi:10.1016/j.jag.2012.07.020
.
Jeong
Y.
,
Yu
J.
,
Wang
L.
,
Shin
H.
,
Koh
S.-M.
&
Park
G.
2018
Cost-effective reflectance calibration method for small UAV images
.
International Journal of Remote Sensing
39
(
21
),
7225
7250
.
doi:10.1080/01431161.2018.1516307
.
Johnson
M.-V. V.
,
Kiniri
J. R.
&
Burson
B. L.
2010
Ceptometer deployment method affects measurement of fraction of intercepted photosynthetically active radiation
.
Agronomy Journal
102
(
4
),
1132
1137
.
doi:10.2134/agronj2009.0478
.
Kim
H. W.
,
Kim
J.-H.
,
Li
W.
,
Yang
P.
&
Cao
Y.
2017
Exploring the impact of green space health on runoff reduction using NDVI
.
Urban Forestry & Urban Greening
28
,
81
87
.
doi:10.1016/j.ufug.2017.10.010
.
Knipling
E. B.
1970
Physical and physiological basis for the reflectance of visible and near-infrared radiation from vegetation
.
Remote Sensing of the Environment
1
,
155
159
.
doi:10.1016/S0034-4257(70)80021-9
.
Lee
H. K.
2018
A panel data analysis of a spatial measurement of green infrastructure and its potential effectiveness on peak streamflow
.
Environment, Development and Sustainability
.
doi:10.1007/s10668-018-0210-3
.
Lucke
T.
&
Nichols
R. W. B.
2015
The pollution removal and stormwater reduction performance of street-side bioretention basins after ten years in operation
.
Science of the Total Environment
536
,
784
792
.
doi:10.1016/j.scitotenv.2015.07.142
.
MAPIR
.
2017
MAPIR Survey3 Quick Start Guide
. .
McDonald
W.
2019
Drones in urban stormwater management: a review and future perspectives
.
Urban Water Journal
.
https://doi.org/10.1080/1573062X.2019.1687745
.
Milton
E. J.
,
Schaepman
M. E.
,
Anderson
K.
,
Kneubühler
M.
&
Fox
N.
2009
Progress in field spectroscopy
.
Remote Sensing of Environment
113
,
S92
S109
.
Miura
T.
,
Huete
A.
&
Yoshioka
H.
2006
An empirical investigation of cross-sensor relationships of NDVI and red/near-infrared reflectance using EO-1 Hyperion data
.
Remote Sensing of Environment
100
,
223
236
.
doi:10.1016/j.rse.2005.10.010
.
Mohammed
G. H.
,
Binder
W. D.
&
Gillies
S. L.
1995
Chlorophyll fluorescence: a review of its practical forestry applications and instrumentation
.
Scandinavian Journal of Forest Research
10
(
1–4
),
383
410
.
doi:10.1080/02827589509382904
.
Monteith
J. L.
1972
Solar radiation and productivity in tropical ecosystems
.
Journal of Applied Ecology
9
,
747
766
.
doi:10.2307/2401901
.
Muñoz-Huerta
R. F.
,
Guevara-Gonzalez
R. G.
,
Contreras-Medina
L. M.
,
Torres-Pacheco
I.
,
Prado-Olivarez
J.
&
Ocampo-Velazquez
R. V.
2013
A review of methods for sensing the nitrogen status in plants: advantages, disadvantages and recent advances
.
Sensors
13
,
10823
10843
.
Möller
M.
,
Alchanatis
V.
,
Cohen
Y.
,
Meron
M.
,
Tsipris
J.
,
Naor
A.
,
Ostrovsky
V.
,
Sprintsin
M.
&
Cohen
S.
2007
Use of thermal and visible imagery for estimating crop water status of irrigated grapevine
.
Journal of Experimental Botany
58
(
4
),
827
838
.
doi:10.1093/jxb/erl115
.
OceanOptics
.
2013
OceanView Installation and Operation Manual
.
Ocean Optics, Inc.
Available from: https://oceanoptics.com/product/oceanview-downloads/ (accessed 25 May 2020)
.
OceanOptics
.
2015
Flame Miniature Spectrometer User Manual
.
Ocean Optics, Inc.
Available from: https://oceanoptics.com/product/flame-spectrometer/ (accessed 25 May 2020)
.
Padro
J.-C.
,
Munoz
F.-J.
,
Avila
L. A.
,
Pequer
L.
&
Pons
X.
2018
Radiometric correction of Landsat-8 and Sentinel-2A scenes using drone imagery in synergy with field spectroradiometry
.
Remote Sensing
10
,
1687
.
doi:10.3390/rs10111687
.
Pe'eri
S.
,
McLeod
A.
,
Lavoie
P.
,
Ackerman
S.
,
Gardner
J.
&
Parrish
C.
2013
Field calibration and validation of remote sensing surveys
.
International Journal of Remote Sensing
34
(
18
),
6423
6436
.
Prabhakaraa
K.
,
Hively
W. D.
&
McCarty
G. W.
2015
Evaluating the relationship between biomass, percent groundcover and remote sensing indices across six winter cover crop fields in Maryland, United States
.
International Journal of Applied Earth Observation and Geoinformation
39
,
88
102
.
doi:10.1016/j.jag.2015.03.002
.
Ramseyer
N.
2017
Pre-Process Survey3 Images in MAPIR Camera Control (MCC). MAPIR Camera. Available from: https://www.mapir.camera/blogs/guide/pre-process-survey3-images-in-mapir-camera-control-mcc (accessed 25 May 2020)
Rosati
A.
&
Dejong
T. M.
2003
Estimating photosynthetic radiation use efficiency using incident light and photosynthesis of individual leaves
.
Annals of Botany
91
,
869
877
.
doi:10.1093/aob/mcg094
.
Sankaran
S.
,
Khot
L. R.
,
Espinoza
C. Z.
,
Jarolmasjed
S.
,
Sathuvalli
V. R.
,
Vandemark
G. J.
,
Miklas
P. N.
,
Carter
A. H.
,
Pumphrey
M. O.
,
Knowles
N. R.
&
Pavek
M. J.
2015
Low-altitude, high-resolution aerial imaging systems for row and field crop phenotyping: a review
.
European Journal of Agronomy
70
,
112
123
.
https://doi.org/10.1016/j.eja.2015.07.004
.
Serrano
L.
,
Filella
I.
&
Penuelas
J.
2000
Remote sensing of biomass and yield of winter wheat under different nitrogren supplies
.
Crop Science
40
,
723
731
.
doi:10.2135/cropsci2000.403723x
.
Shi
Y.
,
Thomasson
J. A.
,
Murray
S. C.
,
Pugh
N. A.
,
Rooney
W. L.
,
Shafian
S.
,
Rajan
N.
,
Rouze
G.
,
Morgan
C. L. S.
,
Neely
H. L.
,
Rana
A.
,
Bagavathiannan
M. V.
,
Henrickson
J.
,
Bowden
E.
,
Valasek
J.
,
Olsenholler
J.
,
Bishop
M. P.
,
Sheridan
R.
,
Putman
E. B.
,
Popescu
S.
,
Burks
T.
,
Cope
D.
,
Ibrahim
A.
,
McCutchen
B. F.
,
Baltensperger
D. D.
,
Avant
R. V.
Jr
,
Vidrine
M.
&
Yang
C.
2016
Unmanned aerial vehicles for high-throughput phenotyping and agronomic research
.
PLoS One
.
https://doi.org/10.1371/journal.pone.0159781
.
Smith
G.
&
Milton
E.
1999
The use of the empirical line method to calibrate remotely sensed data to reflectance
.
International Journal of Remote Sensing
20
(
13
),
2653
2662
.
doi:10.1080/014311699211994
.
Toro
G.
,
Flexas
J.
&
Escalona
J. M.
2019
Contrasting leaf porometer and infra-red gas analyser methodologies: an old paradigm about the stomatal conductance measurement
.
Theoretical and Experimental Plant Physiology
31
,
483
492
.
https://doi.org/10.1007/s40626-019-00161-x
.
United States Environmental Protection Agency (USEPA), Office of Water
.
1999
Preliminary Data Summary of Urban Storm Water Best Management Practices
.
United States Environmental Protection Agency, Office of Water
,
Washington, D.C
United States Environmental Protection Agency (USEPA), Office of Water
.
2002
Urban Stormwater BMP Performance Monitoring: A Guidance Manual for Meeting the National Stormwater BMP Database Requirements
.
Available from: https://nepis.epa.gov
.
United States Environmental Protection Agency (USEPA), Office of Water
.
2009
Urban Stormwater BMP Performance Monitoring Manual
. .
Wang
C.
&
Myint
S.
2015
A simplified empirical line method of radiometric calibration for small unmanned aircraft systems-Based remote sensing
.
IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
8
(
5
),
1876
1885
.
doi:10.1109/JSTARS.2015.2422716
.
Weler
J.
&
Herring
D.
2000
Measuring Vegetation (NDVI & EVI). NASA: Earth Observatory. Available from: https://www.earthobservatory.nasa.gov/features/MeasuringVegetation
.
Williams
A. S.
,
Kiniri
J. R.
,
Mushet
D.
,
Smith
L. M.
,
McMurry
S.
,
Attebury
K.
,
Land
M.
,
McCarty
G. W.
,
Shaffer
J. A.
,
Effland
W. R.
&
Johnson
M.-V. V.
2017
Model parameters for representative wetland plant functional groups
.
Ecosphere
8
(
10
),
e01958
.
doi:10.1002/ecs2.1958
.
Zecevic
Z.
,
Popovic
T.
&
Krstajic
B.
2017
Cloud based solution for automatic image mosaicking and georeferencing
. In:
Proceedings of 22nd International Scientific-Professional Conference Information Technology
. .
Zhang
L.
,
Sun
X.
,
Wu
T.
&
Zhang
H.
2015
An analysis of shadow effects on spectral vegetation indexes using a ground-based imaging spectrometer
.
IEEE Geoscience and Remote Sensing Letters
12
(
11
),
2188
2192
.
doi:10.1109/LGRS.2015.2450218
.
This is an Open Access article distributed under the terms of the Creative Commons Attribution Licence (CC BY 4.0), which permits copying, adaptation and redistribution, provided the original work is properly cited (http://creativecommons.org/licenses/by/4.0/).