Evaluating the accuracy of unmanned aerial systems to quantify glacial ice habitats of harbor seals in Alaska

Abstract

Long-term monitoring programs to evaluate climate-driven changes to tidewater glaciers, an important habitat for harbor seals (Phoca vitulina) in Alaska, are primarily carried out by costly, weather-dependent aerial surveys from fixed-winged aircraft. Unmanned aerial systems (UAS) can be an alternative cost-effective solution for gathering image data to quantify, monitor, and manage these habitats. However, there is a paucity of information related to the accuracy of using imagery collected by UAS for purposes of measuring floating icebergs. We evaluated the accuracy of using a UAS with a built-in 20 megapixel (MP) camera as well as a consumer-grade digital 16 MP camera to capture images of floating and stationary icebergs for the purpose of collecting vertical height measurements. Images (n=869) were captured of simulated icebergs (Cuboidal foam boxes) “Cb” (n=5) and real icebergs (n=5) that were either grounded or floating. The mean error ratios obtained were less than 10% and derived by comparing the mean calculated measurements of heights of Cb obtained from images captured by UAS with the physically measured heights of these Cb. The mean error ratio for height measurements of grounded icebergs (n=4) and one floating iceberg was also less than 10%. Within an object-image distance range of 6–25 m, the cameras captured images that were suitable to accurately calculate the heights of floating and grounded objects, and drift or uncontrolled movement of the UAS caused by wind or temporary loss of GPS did not appear to have any significant effects on measurement error. Our study provides substantial evidence of the high accuracy associated with using images captured by UAS for measuring dimensions of structures positioned on water and land surfaces. Ultimately, accurate surveys of glacial ice used by harbor seals will improve our understanding of the role of decreasing habitat in explaining population variability between different tidewater glaciers.

Introduction

Tidewater glaciers are valley glaciers that terminate and calve icebergs into nearshore marine environments from the terminal boundary where the glacier meets the ocean (McNabb et al., 2015). Warming sea temperatures have altered rates of frontal ablation of many tidewater glaciers which drives the transfer of solid ice from the glacier to nearshore marine environments in the form of glacial ice pieces that includes icebergs (Holmes et al., 2019). Many tidewater glaciers in Alaska have retreated over the last 50 years coinciding with warming ambient air and ocean temperatures (Hall et al., 2015; Luckman et al., 2015). The melt rate of floating icebergs in seawater that is above freezing, is proportional to the water temperature raised to an exponential of 1.5 (Russell-Head, 1980). Therefore, even small increases in ambient temperatures can result in a significantly shorter time frame for icebergs to exist in a solid form.

Tidewater glaciers in Alaska provide habitat for some of the largest seasonal aggregations of harbor seals in the world (Jansen et al., 2010, 2015; Womble et al., 2021). Harbor seals select floating glacial ice pieces as haulout to escape the marine environment and these platforms can be used for pupping, resting, and molting (Womble et al., 2021). However, very little is known about the dimensional characteristics of glacial ice pieces that are selected as platforms for haulout. Concern has been raised about the impacts of climate-driven changes to ice-dominated environments and the impacts on the population dynamics of marine mammals that utilize these habitats (Laidre et al., 2008; Moore & Huntington, 2008, Mathews & Pendleton, 2006). For example, Calambokidis et al. (1987) noted that between 1973–1986 floating icebergs that were utilized by harbor seals as haulout substrate in Muir Inlet, Glacier Bay in southeast Alaska shifted 10 km upstream following the retreat of the Muir Glacier. Prior to this retreat, seal populations in upper Muir Inlet in the East Arm of Glacier Bay exceeded 1,300 seals (Streveler, 1979); however, grounding of the Muir Glacier resulted in termination of floating icebergs, loss of haulout substrate (Mathews, 1995) and by 2008, no seals were pupping in nearshore region of upper Muir Glacier (Womble et al., 2010).

Climate-driven changes in the spatial distribution of ice in the nearshore of tidewater glaciers influence the structuring variability of the icescape layout and can therefore impact the spatial location of floating ice haulouts and other possible microhabitats used by harbor seals (Womble et al., 2021). Two-dimensional (2D) vertical measurements are commonly used to classify glacial ice pieces that are calved from the terminus of tidewater glaciers and float in the nearshore (Armstrong et al. 1966). Long (1992) classified glacial ice pieces in Alaska based on the 2D vertical heights above sea level to include, brash ice” 0–0.5 m, “growler “≤ 1.0 m, “bergy bits”, 1.0–4.5 m and “iceberg”, ≥4.5 m. Quantifying glacial ice haulouts that can be accessed by animals to escape the marine environment can be examined by the 2D vertical heights of these glacial ice pieces because pinnipeds have limited climbing abilities, and steepness of the haulout prevents or curtails access to haulouts (Schneider and Payne, 1983). Measuring the 2D vertical heights in addition to horizontal measurements of access points along the edges of floating ice pieces can be used to estimate gradients of the access points on haulouts. Additionally, harbor seals might use different microhabitats within the nearshore for different purposes (i.e., densely packed ice and widely disbursed icebergs) and the established classification of glacial ice could also be applied to quantify these microhabitats.

At present, analysis of glacial ice in tidewater glaciers are typically completed using high resolution aerial images obtained from harbor seal surveys (Jansen et al. 2010; 2015). Images are captured by cameras mounted on fixed-wing aircraft, that are flown at relatively high altitudes (approximately 305 m) along established transects and surveys are can be hindered by adverse weather conditions (McNabb et al., 2016, Womble et al., 2021). Advancements in computer imaging and technological innovations in remote sensing software, as well as reduced operational costs and increased capabilities of low-altitude (<20 m) unmanned aerial systems (UAS), have greatly advanced our abilities to map and evaluate environments and can be an alternative method for capturing aerial images (Watts et al., 2012). The latest UAS models are equipped to fly in adverse weather conditions and some models can house high resolution cameras. However, little is known about the accuracy of using a UAS with a built in camera to capture images of floating icebergs for the purpose of measuring dimensional features of floating glacial ice pieces.

The overall goal of the present study was to develop and test the accuracy of a cost-effective technique to estimate the dimensional properties of floating glacial ice that occur in the nearshore of tidewater glaciers in Alaska. The objectives of the study were to: 1) use images captured by UAS to evaluate the accuracy of measuring the heights of test objects having a defined cuboidal shape that were either floating or grounded (i.e., simulated icebergs made of foam boxes, “Cb”), and 2) evaluate the accuracy of measuring the heights of irregularly shaped grounded and floating icebergs from images captured by consumer grade digital cameras. The technique described in this study could be considered as a foundational method for efficiently measuring and associating glacial ice pieces that may be used by harbor seals with the established 2D classification of floating ice pieces. The technological framework provided by this study will provide a method to quantify physical attributes of glacial ice based on 2D vertical height of icebergs as a means of investigating fine-scale use of microhabitats by pinnipeds.

Methods

The UAS used for the present study was manufactured by Da Jiang Innovations Sciences and Technologies Ltd.(DJI) and the model tested was the Phantom 4 Pro (P4P), a consumer grade model. The P4P is equipped with a built-in camera attached to the body of the UAS that is housed on a 3-axis electronic gimbal cradle. The P4P camera contains a 20 megapixel (MP), 1-inch (15.86 mm) Sony Exmor R CMOS (complementary metal-oxide semiconductor) sensor (model: IMX 163 sensor) with the following dimensions: height 8.8 mm (3648 effective vertical pixels); length 13.2 mm (5472 effective horizontal pixels); diagonal length 15.86 mm, pixel size 2.40 μm, and pixel density of 17.18 MP/cm2. This UAS model was equipped with a vision positioning module that tracks the ground below to keep the craft steady when a GPS connection can’t be obtained and uses both visual data and sonar to read ground patterns as well as current altitude. The guaranteed hover accuracy of the P4P is vertical: ±0.1 m (with vision positioning) and ±0.5 m (with GPS positioning); horizontal: ±0.3 m (with vision positioning); ±1.5 m (with GPS positioning) (DJI specifications). The altitude drift in the P4P was minimized by manually controlling the craft so the base of the built-in camera remained aligned with the top of a vertically erected 1.5 m survey rod or in some cases the propellers were deactivated and the entire body of the UAS was physically held in a fixed position by placing the UAS on a stationary platform. All flying activities were carried out by a UAS pilot licensed by the Federal Aviation Administration (Part 107). The flight time of the P4P UAS used in the present study was 25 minutes with approximately 50% of the total flight time spent in positioning the UAS at the appropriate position to capture images of the object, take-off, and landing. Manufacturers recommend that when flying over water, to commence the return flight with at least 20% of the battery life which significantly reduces the time dedicated to image capture.

Objective 1

To satisfy the requirements for the first objective we investigated measurement error that resulted when the heights of five known regular cuboidal shaped objects were measured from images of those objects that were captured by the UAS’s camera at different object-image distances. We tested six different experimental treatments (ET) of UAS position and Cb location that include the following: (ET 1) Cb on land, UAS on land; (ET 2) Cb on water, UAS on water; (ET 3) Cb on land, UAS in air; (ET 4) Cb on water, UAS in air; (ET 5) Cb on water, UAS on land, and (ET 6) Cb on land, UAS on water (Figure 1). The UAS was either positioned airborne and allowed to hover at an altitude of approximately 1.1–2.0 m above the ground, placed on a solid land surface, or floated on a water platform while the objects were located on land and floated on a water surface. The land-based trials (ET 1, 3, and 6) were conducted on a solid surface parking lot while the water-based trials (ET 2, 4, and 5) were conducted in Aurora Harbor along the jetty of a sheltered dock in Juneau, Alaska.

Figure 1.

Design for testing the accuracy of images captured by a 20 MP camera housed on an electronic gimbal system attached to a UAS. Six experimental treatments (ET) were used on 5 different sized Cuboidal boxes (Cb); ET 1, UAS on land Cb on land; ET 2, UAS on water Cb on water; ET 3, UAS in air Cb on land; ET 4, UAS in air Cb on water; ET 5, UAS on land Cb on water; and ET 6, UAS on water Cb on land. Images were taken at 10 different object-image distances that ranged from 3.04 – 30.48 m (10 – 100 ft).

We investigated the error ratio associated with changes in object height and object-image distance for each of the six ET (Table 1). Images were captured for each of five Cb heights with defined edges and known heights of 1.2 m, 0.9 m, 0.8 m, 0.5 m, and 0.3 m at distance intervals ranging between 3.1–30.5 m (i.e., 10–100 ft.). Object-image distances were systematically increased in intervals of 3.1 m (i.e., 10 ft.) within this range after three replicate images of the Cb were captured. A metric ruler was permanently affixed to the face of each Cb as a reference. Anecdotal observations indicated that there was unintentional movement (drift) in the positioning of the UAS when it was airborne that appeared to increase in magnitude during windy conditions. Unwanted movements during image capture can potentially be a source of error because the drift can alter the object-image distance. To evaluate the effect of drift, the UAS camera was activated but lift rotors deactivated and the UAS held in a stationary position at a height of approximately 1 m above the ground or sea surface. The UAS was placed on a fixed stand when on land or held by hand over the water surface to allow for the three replicate images to be captured of the targeted object. This provided data to compare measurement errors from images that were captured from the airborne UAS with images of the activated UAS but kept in an exact position.

Table 1.

Experimental design for testing the accuracy of length measurements extracted from images captured at 10 unique object-image distances.

Treatment ID	UAS position	Cb Location	Cb Height (m) Investigated for each Object-Image Distance	Object-Image Distance (m) Investigated
1	Land	Land	1.2, 0.9, 0.8, 0.5, 0.3	3.0, 6.1, 9.1, 12.2, 15.2, 18.3, 21.3, 24.4, 27.4, 30.5
2	Water	Water	1.2, 0.9, 0.8, 0.5, 0.3	3.0, 6.1, 9.1, 12.2, 15.2, 18.3, 21.3, 24.4, 27.4, 30.5
3	Air	Land	1.2, 0.9, 0.8, 0.5, 0.3	3.0, 6.1, 9.1, 12.2, 15.2, 18.3, 21.3, 24.4, 27.4, 30.5
4	Air	Water	1.2, 0.9, 0.8, 0.5, 0.3	3.0, 6.1, 9.1, 12.2, 15.2, 18.3, 21.3, 24.4, 27.4, 30.5
5	Land	Water	1.2, 0.9, 0.8, 0.5, 0.3	3.0, 6.1, 9.1, 12.2, 15.2, 18.3, 21.3, 24.4, 27.4, 30.5
6	Water	Land	1.2, 0.9, 0.8, 0.5, 0.3	3.0, 6.1, 9.1, 12.2, 15.2, 18.3, 21.3, 24.4, 27.4, 30.5

Note: A 20 mp camera affixed to a UAS captured images of cuboidal boxes (Cb) that were placed on either water or land surfaces. The following variables: UAS Position, Cb Location, Cb Height (m), Object-Image distance (distance between Cb and UAS) (m).

To investigate the error associated with floating Cb, a raceway was built directly in front of a boat dock using a 3 mm nylon rope that extended to 31 m in front of the boat dock. The nylon rope was pulled taught to define the boundaries of a rectangular track to which a small floating platform that supported the UAS was confined but allowed forward and backward movement from the boat dock by means of a pulley system. This prevented unwanted sideways movement of the floating UAS from wind, tide, or currents and allowed the platform containing the UAS to be accurately positioned at the desired object-image distances from the boat dock. The Cb were floated but securely attached to the boat dock for one set of ET (Figure 1, ET 2, 4, and 5) and placed on the boat dock for ET 6. The same Cb were used as objects for the six ET of UAS positions and Cb locations and the same procedures for data capture were followed for each treatment.

Objective 2

Surveys that included either floating or grounded icebergs were conducted at the Mendenhall Glacier in Mendenhall valley, Juneau Alaska (58.4958° N, 134.5322° W), a former tidewater glacier that has receded from the coast and is now currently landlocked undergoing continued recession, thinning, and calving icebergs into Mendenhall Lake (Motyka et al., 2003). We evaluated the accuracy of measuring irregularly shaped icebergs that were either grounded or floating by comparing mean error ratios of the true or physically measured heights of the icebergs with the calculated heights obtained from images of the icebergs.

Irregular shaped objects can create several challenges that are related to identifying the object height in the image due to the inherent loss of information arising from projective transformation. It is well known that information is lost in depth associated with portrayal of a 3D object on a 2D surface (Su, et al., 2014). To minimize this error we selected four grounded icebergs that were beached on a rocky moraine at the terminus of the Mendenhall Glacier. The grounded icebergs allowed accurate measurements to be made of the object as well as gave the opportunity to establish a fixed apex point that was as best as possible in the same plane as the base of the object for purposes of measuring the 2D height.

A power analysis indicated that a sample size of three icebergs would be adequate to evaluate statistical trends and as such we decided to use a conservative sample (n=4). Accordingly, four irregular-shaped pieces of ice of varying sizes (measured heights: 1.8 m, 2.1 m, 0.9 m, and 0.7 m) that were grounded on an small rocky beach formed at the base of the tidewater glacier were randomly selected. The four icebergs chosen for the study contained numerous peaks and a peak that was most convenient to set a measuring scaffolding was selected. The scaffolding was constructed to support a vertically erected 10 m measuring survey rod or 1 m ruler in the case of smaller ice pieces and a horizontally-levelled 1 m ruler that extended from the peak of the iceberg to the vertical survey rod. The horizontal ruler was kept in place at the peak of the iceberg by attaching it to an ice-screw that was screwed into the peak. The intersection of the horizontal ruler and the vertical survey rod was considered to be the measured height of the iceberg. A 20 m transect was constructed using a standard non-stretch flexible measuring tape that was fixed to the base of the iceberg where the survey rod contacted the moraine. Three replicate images of each iceberg were captured at each of the object-image distances selected which ranged between 2–20 m and object-image distances were increased in intervals of 1 m within this range. The digital camera used for these irregular shaped icebergs was Nikon® model TG-4, having a 16 MP sensor (pixel density 56.06 MP/cm2). The camera was positioned by hand at designated distance intervals keeping the height of the camera at eye level (approximately 1 m above the ground) and rapidly depressing the shutter knob three times so that images were captured in triplicate at each of the desired object-image distances.

One floating piece of ice was identified in Mendenhall Lake that could easily be approached by wading through shallow water and selected for the study. A Cb (height 0.3 m) was secured and allowed to float next to the floating ice that was measured using a survey rod (height 0.48 m). The Cb was positioned to allow for capture of both objects (iceberg and Cb) in each image. A standard survey rod that was aligned with the peak of the iceberg was secured to the Cb that floated next to this iceberg and used as a reference marker. A 25 m transect was extended from the iceberg and the same procedures for image capture repeated as described in Objective 2.

Equations and Statistical Analyses

Dimensional measurements captured in images were made using a mathematical relationship between the ratio of the measured height of the object and the distance between the object and the lens of the camera (Kendal, 2007). Specifically, the ratios of any of the dimensional lengths of an object and the image of that object that is formed on a camera sensor are proportional to the distance of separation between the actual object, camera lens, and the digital sensor (i.e., thin-lens formula) (Chakravarti & Siegel, 2001). Heights of objects were calculated from images using the following procedures: images were first imported into AGIsoft photo scan editing software and the metadata associated with images obtained (i.e., focal length of lens that captured image and pixel dimensions of image). Images were then imported into Microsoft Windows® Paint for further dimensional analysis of object heights based on the pixel size of the image. The Microsoft Windows® Paint program divides the picture into pixels and height of objects in the image could be calculated using the following equation:

$$\left[Hr\right]\left(mm\right)=\frac{\left[Do\right]\left(mm\right)\times \left[Ho\right]\left(\mathit{\text{pixels}}\right)\times \left[Hs\right]\left(mm\right)}{\left[F\right]\left(mm\right)\times \left[Hi\right]\left(\mathit{\text{pixels}}\right)}$$ Equation 1

where height of real object [Hr] (mm) = (distance to object [Do] (mm) × height of object in image [Ho] (pixels) × sensor height [Hs] (mm)) / (focal length [F] (mm) × height of image [Hi] (pixels). Error ratios (Er) were calculated for each data point and a mean calculated for each ET and for each of the ten object-image distances used. Error ratios between the measured height (MH) and the heights from image analysis (Hr) of icebergs were determined using the formula

$$\left[Er\right]=\frac{\left[MH\right]\left(mm\right)-\left[Hr\right]\left(mm\right)}{\left[MH\right]\left(mm\right)}$$ Equation 2

The following variables: object-image distance, UAS position, object location, measured height, as well as interactions between these variables, were modeled to evaluate effects on the overall error ratio. To determine differences in the mean of Hr and MH a regression analysis and frequency histogram of the mean error ratios were done to evaluate bias in ET 1–6. The one-way multivariate analysis of variance (one-way MANOVA) was used to determine whether there were any differences between the independent groups examined on more than one continuous dependent variable. A MANOVA was used because we investigated several dependent variables that were being compared to determine the contribution each had on the error ratio observed. For example, size of object, distance between object and image, type of platform that objects were placed (i.e., land and water), and position of UAS (i.e., land, air, and water). A MANOVA was the best statistical solution to compare multiple dependent variables simultaneously and interactions between variables. The MANOVA considered the effects of the 6 ET for the 5 Cb (considered the “object”) and was used to investigate the effects of the following variables on the measuring error: object-image distance (dist), object size (obj.size), UAS position (UAS.pos), object location (obj.loc), distance and object size (dist : obj.size), distance and UAS position (dist : UAS.pos), distance and object location (dist : obj.loc), object size and UAS position (obj.size : UAS.pos), object size and object location (obj.size : obj.pos), and UAS position and object location (UAS.pos : obj.loc) (Zar, 1999). The error ratios associated with measurements that were taken in the horizontal direction were also measured for at least one treatment and the results statistically compared with measurements taken in the vertical direction using a T-Test.

Results

Objective 1

The mean error ratios for ET 1–6 were normally distributed (Figure 2). The mean error ratios for ET 1–6 were −0.01±0.02, 0.02±0.02, 0.00±0.02, 0.01±0.03, −0.02±0.03, and 0.01±0.02, respectively. Of 869 useable images, 16.46% of all the calculated heights of the Cb in ET 1–6 had error ratios of 0.0, and 92.6 % had error ratios that were within the range of −0.05 to +0.05 (Figure 2). The mean error ratio for the 869 images of Cb was 0.02±0.01. Results from the MANOVA indicated that all variables and interactions were significant contributors to measurement error (Appendices S1 and S2). Accordingly, we used a least squares linear regression to analyze data for ET 1–6. In general, the mean error ratios derived by comparing the calculated heights of Cb with the measured heights from each of the ET ranged from −0.10 to 0.11 and 70% - 96% of these values occurred in the range −0.03 to 0.03 (Figure 3).

Figure 2:

Frequency distribution of mean error ratios for 869 images used in the study. Percentage of error ratios obtained from images captured from the six experimental treatments where Cuboidal boxes (Cb) were used as objects for image capture. Images were taken at 10 different object-image distances that ranged from 3.04 – 30.48 m (10 – 100 ft).

Figure 3:

Mean error ratios for six experimental treatments (ET) measuring 5 different sized Cuboidal boxes (Cb), (a) 1.2 m, (b) 0.9 m, (c) 0.8 m, (d) 0.5 m, and (e) 0.3 m. Images were taken at 10 different object-image distances that ranged from 3.04 – 30.48 m (10 – 100 ft). ET 1, UAS on land, Cb on land (open square); ET 2, UAS in water Cb on water (open triangle); ET 3, UAS on air Cb on land (open diamond); ET 4, UAS in air Cb on water (black diamond); ET 5, UAS on land Cb on water, (black triangle); ET 6, UAS in water Cb on land (black square).

The calculated heights under-estimated the measured heights for all sizes of Cb and for all experimental combinations in 39% of the ET and over-estimated the measured heights in 42% of trials, with 18% estimated without error. ET 3 yielded the highest number of trials (23.3 %) and ET 2 yielded the lowest number of trials (10.2 %) where the calculated and the measured heights were identical and had an error ratio of 0. In both or these treatments ET 3 (73.3 %) and ET 1 (61.9 %) respectively, the majority of the calculated heights had error ratios in the range of −0.03 to +0.03. Over a distance range of 6 – 25 m, approximately 86.1% of data in: ET 1, 2, 3, 5, and 6 had error ratios that were ≤ 0.03. In comparison, ET 4 had larger ratio errors (−0.07–0.11) throughout the entire range of distances tested (3 m - 30 m) with 68% of the values occurring in the range of −0.03 to +0.03.

Objective 2

The mean error ratios of the 4 grounded icebergs surveyed (measured heights: 1.8 m, 2.1 m, 0.9 m, and 0.7 m) were 0.04±0.01, −0.03±0.02, 0.01±0.01, and 0.01±0.05, respectively and the minimum error ratios for all treatments occurred within an object-image distance range of 5 −15 m (Figure 4). The mean error ratio for height measurements of a floating iceberg (measured height 0.48 m) from images that were captured by the UAS that was activated but stationary was −0.06±0.08 and −0.06+0.06 for the images collected from the airborne UAS (Figure 5). In comparison, the mean error ratio for the floating iceberg from images that were captured by the UAS that was activated and airborne was 0.08±0.13. There were no significant differences (p>0.05) between horizontal and vertical measurements error ratios (t-value, 1.59326; p-value 0.06 when compared with UAS airborne and object on land (Appendix S3).

Figure 4.

Mean error ratios from four grounded icebergs. Images were taken at different object-image distances that ranged from 1–20 m and distances were increased in increments of 1 m. Iceberg heights were: 1.80 m (black circle), 2.10 m (black triangle), 0.88 m (black diamond), 0.70 m (black square).

Figure 5.

Comparison of mean error ratios for airborne (square) and stationary (triangle) UAS to evaluate the effect of drift when the UAS is in flight. Images were taken at 8 different object-image distances that ranged from 3.04 – 25.00 m.

Discussion

The major findings from the present study were that (1) consumer grade UAS packages containing cameras having a sensor pixel density of at least 4.39 MP/cm² were suitable for capturing image data at a distance of 6 – 25 m from the object, with a mean error ratio of 0.02 ± 0.01. Images captured by UAS could be used to measure dimensional properties of floating objects with a mean error ratio of 0.01±0.03 and approximately 50% of calculated measurements had an error ratio in the range of −0.03 to +0.03. (2) The combination of large object-image distances and limited image resolution were the major sources of error, especially when object-images distances exceeded 25 m, which resulted in pixilation in images. Drift caused from the UAS being airborne had minimal effect on the accuracy of measurements.

Researchers have used cameras (handheld and attached to UAS) to capture images of target objects and computer software packages to create three-dimensional models from these images for purposes of measuring dimensional characteristics (Birdal et al., 2017; Miller et al., 2015). It was expected that using a consumer grade UAS to capture images of floating icebergs in tidewater glaciers to measure the dimensional properties of these objects would yield error ratios that were similar in magnitude and distribution to the observed errors in ET 4 (Cb on water, UAS in air). Even in this case, the mean error ratio based on object-image distance as well as object size did not exceed 0.05 and therefore images captured would allow relatively accurate estimations of the dimensional features of these icebergs. Additionally, the largest range in the distribution of error ratios occurred at object-image distances that were less than 6.0 m, which would not be a practical distance for image capture of larger-sized icebergs. Thus, short distances (<6.0 m) that are associated with relatively large errors should be excluded.

Our findings were within the range of error reported by Ridolfi & Manciola, (2018) in a similar study that used images captured by UAS to estimate the height of a water column at a dam site. These authors reported that the mean error between observed and estimated values were relatively small (0.045±0.02, −0.01±0.02, and 0.031±0.02 cm) and likewise the absolute errors in height in our study were 0.02±0.01 for the six ET investigated. These authors also reported that the sources of uncertainty in their study were related to the perspective and pixilation in images, a phenomenon that was also observed in our study. For example, Ridolfi & Manciola, (2018) reported that detection of the exact edges of the waterline created a source of uncertainty because of waves, angle and intensity of the incoming light source, as well as abnormalities in the camera (i.e., image focus, image resolution, perspective and lens distortion). Similarly in our study, the exact edges of Cb became blended into the background at object-image distances that exceeded 25 m and at least 20% of images that were captured at distances >20 m did not have adequate detail due to either pixilation or poor resolution.

Cameras having digital sensors with higher pixel density is a method of reducing error caused from pixilation when images are captured at relatively long distances (Jackson & Jackson, 2015). Cameras with larger sensor sizes can potentially support more pixels per square inch and allow images to be captured from further distances with reduced pixilation. For example, point-and-shoot cameras used in the mid-range consumer grade UAS (e.g., P4P) have relatively small sized sensors in the range of 12.80 × 9.60 mm (12.4 MP) with total camera weight of approximately 300 g. In comparison, high-end cameras developed for UAS such as the Hasselblad A6D-100c (Hasselblad Inc.) has a CMOS sensor of 100 MP (11600 × 8700 pixels, 4.6 × 4.6 μm) but weighs approximately 1633 g, without the stabilizing gimbal and attachment hardware. Larger sized sensors usually equate to heavier cameras and would require the UAS to house a heavier payload which significantly reduces the total flight time of many consumer grade models (Choi & Schonfeld, 2017).

Another potential error source for any consumer grade UAS (typically lacking an advanced GPS system) stems from the inaccuracy associated with its exact position in time and space caused by drift, changes in altitude and loss of satellite signals (Sorbelli et al., 2018). The error associated with drift and exact positioning could be reduced by employing Real-Time Kinematic (RTK) positioning to provide real-time corrections to location data when the UAS is in flight (Barry & Coakley, 2013). Installing an RTK system in a UAS is relatively costly because the UAS has to be specially designed to accept software and hardware for RTK. The RTK system has a ground receiver that emits a real time correction directly to the UAS which greatly improves the stability of positioning of the UAS in space and time. There are, however, some drawbacks to using an RTK system in a fjord ecosystem because of the difficulty in finding suitable non-mobile locations to place a base station due to the steep topography of the terrain. Our study demonstrated that the mean error related to drift for even relatively small sized Cb was negligible. Therefore, it might not be necessary to invest in extra RTK equipment and the commercially available consumer grade UAS appears to be suitable for use to capture images of small sized objects, particularly if they are within a 6 – 25 m object-image distance.

In general, there are inherent losses of information when three-dimensional objects are projected onto a 2D image plane such as the sensor in a digital camera (i.e., projective transformation) (Cao et al., 2011). This results from the sensor in a digital camera being restrictive in collecting the array of a complete set of light rays that occur in a three-dimensional space (Zhou & Nayar, 2011). Accordingly, for this study, 2D vertical heights of icebergs that were measured were selected to allow the apex and the base of the iceberg to be in the same plane. Although not covered by this study, future studies might include the use of offset stereo cameras to accurately measure the heights of oblique parts of icebergs.

Three-dimensional imagery created from photogrammetry can produce images that incorporate more information leading to higher accuracy in measuring dimensional features of real objects from images (Chudley et al., 2019; Miller et al., 2015; Ozden et al., 2010). However, this technique requires a series of camera positions around the object, and for each position, the camera is firmly fixed and set to capture an image of the stationary object. Images are captured in a pattern to allow approximately 40% overlap between consecutive images. The reconstructive image software (e.g., structure from motion) uses an algorithm to identify pixels that appears in overlapping images that represent the same point on the real object that was captured at different angles by the camera. This information triangulates the predicted camera positions and is used by the program to create a point cloud of the object that occupies a 3-dimensional virtual space (Westoby et al., 2012). Although this method produces relatively accurate measurements in ideal conditions (i.e. indoor controlled conditions), the algorithm is not particularly suited for floating icebergs that are in constant motion. Further advancements in this field can also lead to improvements in the algorithms used to compensate for oblique image capture and situations where both camera and object are in motion. At present, even small movements of the target objects can cause significant errors in rendering the 3D models.

Nearshore regions of tidewater glaciers are very dynamic and influenced by currents generated from tidewater melt, winds, and ocean currents (Brinkerhoff et al., 2017). Our method avoids several complications that can be problematic in other techniques because of the short time (<1 second) to capture a single image of a glacial ice piece. The image captured can be used to accurately measure the 2D vertical height above sea-level using the lens equation if the object-image distance is measured. This method can be used to gather 2D vertical height information that can be used to form a basic quantification of microhabitats in the nearshore. Our technique does not require much time compared to photogrammetry methods and can be deployed over extended periods to evaluate spatial changes in the presence, absence, or changes in the types of ice as classified by the 2D vertical heights above the sea level. Future studies can also investigate how these changes in ice might be related to changes in animal distribution that are obtained from population survey efforts. Future works in the technique presented can also investigate errors associated with images that are captured at oblique angles and improve on the efficiency of the technique used in this study. However, investigating the accuracy associated with image capture at oblique angles was beyond the scope of our study.

The present study demonstrates that many variables can impact the error of measuring both floating and grounded irregular objects. However, even the largest numerical error values obtained were well within a relatively acceptable margin for purposes of evaluating the range in size of icebergs that harbor seals may access as haulout substrate. We intentionally chose the base model of the consumer grade UAS for our study because it was expected that higher grade models with higher resolution camera systems and more advanced GPS devices would yield greater accuracy with spatial positioning of the UAS and images could be captured from further distances. For purposes of developing a practical technique to measure dimensional properties of floating glacial ice, our study demonstrated that the consumer grade UAS and cameras could provide accurate information of the 2D profiles of floating ice pieces. Although beyond the scope of this study, 2D profiles could be used to examine temporal changes in the spatial distribution of animals based on changes in glacial ice microhabitats.

The use of UAS to examine the relationship between iceberg haulouts and distribution of animals within the nearshore of a tidewater glacier will depend on the particular research question being addressed. Evaluating the relationship between animal ecology and habitat structuring in marine environments is directly associated with analyzing dimensional features of important structures that exist within the environment that animals use (Connell, 1961; Underwood and Chapman, 1989). For example, loosely packed brash ice structures or large icebergs (> 5.0 m) having steep access points preventing entry, may present challenges for harbor seals of varying age classes or reproductive states. Using 2D vertical heights of icebergs as one metric can enable the calculation of oblique gradients of access points on the iceberg, and improve our abilities to quantify the nearshore icescapes and help researchers better understand accessibility of certain haulouts and microhabitats to animals. Information of this nature might not be easily identified by using overhead images of floating ice pieces. Our study demonstrated that measurements obtained from images captured with just a 20-megapixel camera were suitable for enlargement and examination of fine details in the image when the object-image distances were in the range of 10 – 25 m. The study also indicated that floating ice can be measured accurately based on the 2D vertical heights of the ice above water. Since the 2D vertical heights of ice is already an established way used to classify floating ice pieces, the methodologies described in this study can be used to classify the spatial outlay of the nearshore of a tidewater glacier, including dimensional ranges of sizes of icebergs that pinnipeds use for haulout and pupping.

The present study opens up avenues for future research to quantify climate-driven changes in attributes of ice conditions over various time scales. Ultimately icescapes that occur within the nearshore of tidewater glaciers can be structurally complex and highly influenced by the calving rates of the glacier (Womble et al., 2021). The positioning of ice deposited into the marine environment creates structures of hard surfaces that are utilized by pinnipeds (Womble et al., 2021). The entire structuring of nearshore icescapes can be influenced by several factors that drive variability of the positioning, absence, and presence of ice over spatial and temporal scales (Cowan 1992). Mapping and quantifying these habitats have been problematic in the past because of the remoteness of tidewater glaciers and the cost associated with aerial surveys from occupied aircraft (Moreland et al., 2015), which regularly encounter dangerous flying conditions from stormy or rapidly changing weather. Our study demonstrated a very practical, cost effective and accurate method to measure ice structures in these dynamic areas using a camera mounted to a UAS to capture a single image of the target ice that takes less than a second to capture. The relatively short time taken for a single image capture reduces the effects of image blur due to movement of the floating object. Additionally, the relatively small size and quietness of the UAS allows surveys to continue in adverse weather conditions and at lower altitudes posing little if any disturbances to animals.

Warming climates are rapidly modifying many nearshore tidewater habitats. Quantifying microhabitats by the 2D vertical heights of ice can be an efficient means of developing indices that will allow comparisons between different tidewater glaciers. Two dimensional vertical heights in association with other dimensional measurements of glacial ice could be used to identify suitable haulouts that harbor seals might use for pupping, an important feature of tidewater glaciers that are pertinent to the ecology of these animals. The percentage of icebergs that are suitable and accessible as haulouts could be evaluated over different time scales and in so doing will allow researchers to evaluate changes in habitat availability. Several researchers have suggested that reduced quantities of suitable ice for hauling out could explain changes in populations of harbor seals (Mathews 1995). Follow up research could use the technique that is presented in this study to document a range of measurements that could be used to identify suitable icebergs that harbor seals select for haulout including 2D vertical heights which will also allow estimations of the gradients of haulouts. Studies of this nature might yield even more accurate measurements using full frame format mirrorless cameras having sensor sizes that are several magnitudes greater than the ones tested in the present study. The potential for imaging from farther distances should be sought as it will allow researchers to study protected marine mammals without disturbance.

Data Availability Statement

Data (Pegus et al. 2021) are available from ScholarWorks@UA: http://hdl.handle.net/11122/12493