We declare that we received research funding from two sources: AMMC and Woodside Energy Inc. We also declare that Insitu Pacific Pty Ltd provided, as in kind support, the ScanEagle UAV and associated equipment (ground control station, launch and recovery equipment), management and support personnel, and flying time. This does not alter our adherence to all the PLOS ONE policies on sharing data and materials.
Conceived and designed the experiments: AH. Performed the experiments: AH. Analyzed the data: NK DP. Contributed reagents/materials/analysis tools: AH NK DP. Wrote the manuscript: AH NK DP.
Aerial surveys of marine mammals are routinely conducted to assess and monitor species’ habitat use and population status. In Australia, dugongs (
The conservation and management of many marine mammal species is dependent on monitoring their population status by conducting aerial surveys. For example, in the US, the
In Australia, dugong (
Field and analysis techniques for marine mammal surveys have continually been improved to provide robust abundance estimates [e.g. 28,29,30,31,32]. They are generally flown in a small aircraft at a set altitude and speed along transect-style flight paths designed to minimise biases in sampling. Usually, two to four observers call or record sightings in real time. The surveys can either use (a) strip transect sampling, where all animals are counted within a marked strip width on either side of the aircraft and all are considered equally detectable [e.g. 18], or (b) line transect (distance) sampling, where observers record all sightings, and the density of the animals is calculated based on the probability of seeing an animal according to its distance from the trackline [see 28,33]. For the latter, observers record angles and bearings to each sighting so that the distances from the trackline can be calculated. For both techniques, the GPS locations of the sightings are generally calculated according to the aircraft’s GPS location at the time the sighting was called.
Recent developments in the technical capacity and civilian use of Unmanned Aerial Vehicles (UAVs, defined as vehicles flying without a human pilot on board) have led to some investigations into the potential use of these systems for aerial surveys of marine mammals [
Human risk – Manned aerial surveys pose a risk to the observers with at least five aircraft crashes having killed 11 marine mammal researchers during aerial surveys [
Costs – There are large costs involved in chartering aircraft, hiring and training observers, and paying for accommodation. Aerial surveys are usually limited to Beaufort sea states of 3 or less [e.g. 45], and therefore the costs include keeping the aircraft and observers on standby waiting for the appropriate weather conditions. As UAVs continue to be developed for civilian use, we hope that the cost of aerial surveys can be reduced through cheaper operating costs, fewer personnel, flying longer hours per day (i.e., there is no need for resting observers and UAVs can fly longer before refueling), and the potential for surveying in a wider range of weather conditions.
Missed sightings and/or misidentification of animals – Manned aerial surveys depend on (usually four) experienced observers who are capable of correctly identifying marine mammals to species level [
Low resolution of location data – The location of each animal group sighted during manned surveys has error associated with time lag between observers seeing animals and then calling/marking their location, as well as measurement errors inherent in determining the location of the animals that are some distance from the aircraft (trackline). UAV imaging systems will provide an immediate snapshot of the sighting and an accurate GPS location of the image. There is also the potential to take this further and use the detailed flight logs recorded by the UAV system to obtain the GPS location of each animal within the image.
Ability to survey isolated or otherwise inaccessible habitat areas – Some marine mammal species occur in areas that are inaccessible or too dangerous for manned surveys because of their distance from the nearest airstrip or from shore. UAVs do not require an airstrip and can be operated from a ship so may overcome these limitations depending on their range and endurance.
Considering the high demand for dugong aerial surveys in Australia, and the potential for UAVs to eliminate risk to human safety and improve the data collected from these surveys, we conducted a trial in Shark Bay, Western Australia, with the aim of establishing methods for UAV dugong surveys.
The specific objectives of this trial were:
Determine the effectiveness of a UAV with a customised imaging system for detecting and identifying dugongs.
Test the capabilities of the UAV system for surveying dugongs in a range of environmental conditions.
Determine the ideal resolution: area coverage ratio according to altitude (given our imaging system parameters).
Either of the two sampling techniques outlined above (strip and line transect sampling) could conceivably be applied to UAV surveys, but for this research we have focused on the strip sampling technique because this is the standard method for surveying dugongs [
This research was conducted under authorisation by The Murdoch University Animal Ethics Committee (permit R2365/10) and Department of Environment and Conservation, Western Australia, permits CE002918 and SF007592.
The trial was conducted in Shark Bay, which is situated midway along the coast of Western Australia (25°30’S, 113°30’E). The bay is 13,000 km2 in area and divided into two embayments separated by the Peron Peninsula. Shark Bay is afforded a high level of protection as both a Marine Park and a World Heritage Area (WHA). One of the values for which the Bay was nominated as a WHA is its large expanses of seagrass meadows (4,000 km2) and diversity of seagrass species (12 species) [
Insitu Inc., a subsidiary of The Boeing Company, are one of the largest UAV developers and operators worldwide. Their UAVs have been tested by a number of other researchers investigating the use of UAVs for wildlife monitoring [e.g. 36,37,38,54]. Insitu Pacific Pty Ltd, (Insitu’s Australia-based subsidiary) equipped their
Wing Span | 10.2 ft | 3.11 m | |
Length | 4.5 ft | 1.37 m | |
Empty Structure Weight | 28.8 lb | 13.1 kg | |
Max Takeoff Weight | 44.0 lb | 20.0 kg | |
Max Horizontal Speed | 80 knots | 41 m/s | |
Cruise Speed | 48 knots | 25 m/s | |
Ceiling | 19500 ft | 5944 m | |
Endurance | 24+ hours | ||
Propulsion | 1.9 hp (1.4 kw), 2-stroke engine | ||
Fuel | Gasoline (100 octane unleaded non-oxygenated gas) | ||
Navigation | GPS / Inertial | ||
Launch | Pneumatic catapult (“Superwedge UAV Launcher”) | ||
Recovery | SkyHook wingtip capture (“Skyhook”) |
The imaging system payload for this trial contained a Nikon® D90 12 megapixel digital SLR camera, together with a fixed video camera in the nose. Imagery from the latter camera was viewed in real-time from the Ground Control Station, providing improved situational awareness. The SLR camera was mounted within the airframe pointing directly downwards using a number of shock-absorbing mounts to reduce vibrations. Each image capture was tagged in real-time with GPS information from a dedicated receiver. Image capture was controlled (including start, stop and capture rate) via the Ground Control Station and the capture rate could be scheduled to achieve a prescribed proportion of overlap ensuring complete coverage along the transect lines flown. All images were stored on the camera’s memory card and downloaded post flight. A standard 50 mm lens was used throughout the trial. A polarising filter was fitted to the lens for all flights, and the direction of the polarisation was kept constant. A yellow filter was tested during one flight to determine whether this would lessen the effect of sun glitter (refer to section Image analysis).
The UAV Ground Control Station, Superwedge UAV Launcher and Skyhook retrieval system (
Photos of (A) the Ground Control Station and storage (inside shipping containers), (B) the Superwedge UAV Launcher, and (C,D) the Skyhook capturing the UAV, set up in Shark Bay Western Australia.
Map showing the location of our trials at Shark Bay (the Ground Control station was situated at Redcliff), and as an example, the locations of all images captured from all three surveys during Flight 3. Images containing dugong sightings are highlighted
Our permit from the Australian Civil Aviation Safety Authority (CASA) restricted us to flying the
We programmed the UAV to fly a series of parallel line transects over seagrass banks known to be frequented by dugong herds on a daily basis. Our small ‘survey’ consisted of 10 transects, each 1.8 km in length. The 10 transects were spaced at 72 m intervals, which was the width of view of the water surface within the images when the
We flew the UAV on seven missions (flights), and during each flight we aimed to conduct three surveys, increasing the altitude from 500, to 750 and then 1000 ft (see
500 | 48 | 72 | 10% | 0 (edge to edge) |
750 | 72 | 108 | 10% | 25% |
1000 | 96 | 144 | 10% | 50% |
a This is the area of the sea surface visible within each image, where length is along the transect, and width is perpendicular to the transect. Width is equivalent to transect strip width.
The purpose of the trial was to test the capabilities of the imaging system rather than conduct an unbiased survey. Therefore, for each survey we wanted to ensure that the number of images containing dugongs could be maximised. To achieve this goal, the exact location of the survey area, although always on the same seagrass bank, was determined immediately prior to each trial flight according to real time boat-based observations of the dugongs. While the
Post flight, one experienced aerial survey observer (lead author) manually reviewed all images captured whilst transects were being flown. This review consisted of searching for identifiable animals (i.e., including fauna other than dugongs) and scoring the environmental conditions in each image. The Raw images were viewed in ViewNX 2™ (Ver 2.0.1, Nikon®, 2010). Sighting data recorded for each image included:
Species, or where not possible, general taxa.
Number of individuals of given species / taxa.
Number of calves.
Number classified (subjectively) as ‘certain’ and number of ‘uncertain’ individuals; uncertain sightings were either clearly fauna but of unclear taxa, or a ‘dugong shape’ that could not confidently be distinguished within the image.
Number of double counts (i.e., the same individual animal occurring within the 10% overlap of successive images).
Calves were distinguished because it is important to monitor the proportion of calves in a population when assessing population status. During manned surveys, dugong calves are generally discernible due to their small relative size and their close proximity to their mother. We wanted to ensure that calves are similarly discernible in UAV surveys.
Individual animals resighted (double counted) in successive images along each transect could be identified and were subtracted from the count of individuals for that image. For surveys flown at 500 ft, it was assumed there was no overlap of images between transects because the width of area covered in each image was the same as the distance between the transects. However, at 750 and 1000 ft, there was overlap of images between transects (
All images containing dugong sightings were then re-checked by the same reviewer to ensure dugong counts and associated sighting data were accurate.
Three environmental variables were also scored for each image: sea state, turbidity and sun glitter. Sea state was scored for each image according to the Beaufort sea state scale. Turbidity (which incorporated a measure of depth) was subjectively scored for each image captured according to the following categories:
Shallow with the bottom clearly visible
Shallow with the bottom visible but obscured by turbidity
Deep with the bottom not visible, but clear water
Deep with the bottom not visible and turbid water
Sun glitter is the sparkling reflection of sun on the water. Each point of light is a specula reflection of the sun, called a sun glint, and glitter is a combination of many sun glints reflecting off wavelets [
The effect of altitude on the certainty of identifying dugongs was tested within a mixed-effect generalized linear model (implemented with R package
In order to determine the effects of four covariates—altitude, sun glitter, turbidity and Beaufort sea state—on the dugong sighting rate, we made the following assumptions:
The number of dugongs available to be photographed during the survey is equal relative to each of the covariates. (We acknowledge that turbidity, which incorporates a course measure of water depth, is likely correlated with dugong distribution, however in the absence of data to adequately model this relationship, we are assuming turbidity does not influence dugong distribution.)
The number of dugongs available to be photographed during a single survey flight remains constant (i.e., throughout that flight, but can vary between flights).
Groups of dugongs are distributed randomly throughout the survey area, and there are no systematic trends in the values of the nominated environmental covariates throughout the survey area.
We then tallied the total number of dugongs sighted in each image (the sample unit), which was accompanied by the associated four covariates. Exploratory analysis found no significant co-linearity between these covariates, and so these were treated as independent during model selection.
We fit a generalized linear mixed-effects model (GLMM) to the number of dugongs detected in each of the images for analysis of the relationship between the ability to see dugongs and our covariates. The response variable—the number of dugongs per image—is considered to be Tweedie distributed in order to account for dugongs forming groups [
To account for temporal autocorrelation in the number of dugongs in the survey area at any one time, each flight was treated as a random effect; to account for spatial autocorrelation within each flight, each transect was also treated as a random effect, which was nested within flight. Where there were multiple altitudes flown within each flight, the same set of transects were re-used; this is reflected in the modeling by treating each transect as a replicate within a flight. Each random effect was fitted as an intercept only (i.e., we assumed no interaction between the random effect and any of the fixed effects). We assumed compound symmetry in correlations within each level of the random effects, i.e., we considered each image to be similarly correlated with every other image in that level.
The covariates altitude, sun glitter, turbidity and Beaufort sea state were treated as fixed effects. Altitude (500, 750 and 1000 ft) and Beaufort sea state (0-5) entered the models as continuous values. Sun glitter estimates of 0, <25%, 25-50% and >50% entered the models as an ordinal variable 1, 2, 3 and 4. In testing for the effect of turbidity, this variable was specified in two ways to capture both the traditional way in which it is described (i.e., the 1-4 classification outlined above) and a construct designed to describe the interaction between depth and water clarity. For the water depth × clarity specification, the turbidity variable was split into two separate variables, requiring three parameter estimates. These were: a two-level factor that described depth, either shallow (base level of factor) or deep; a two-level factor that described how clear the water was, either clear (base level of factor) or murky; and an interaction term linking the depth and water clarity variables back to the original 1-4 classification. A constraint was applied during model-selection to ensure that depth, clarity and their interaction remained together in the model when testing that specification of turbidity.
Because the data are counts, we used a logarithmic link function, and we used an offset term to account for the area covered by images (note, the area per image increases with increasing altitudes, as shown in Results).
We used a backwards selection process to choose the model of best fit for the fixed-effects component. Given the data were non-Normal and overdispersed, a
We conducted seven flights between 16 and 21 September 2010. Details of each flight are provided in
1 | 16/09/2010 | 11:31 | 500 | 476 | 35 | 416 | 25 | 0 | 1 | |
12:00 | 750 | 220 | 30 | 178 | 12 | 0 | 1 | |||
2 | 18/09/2010 | 8:24 | 500 | 702 | 0 | 648 | 51 | 3 | 5 | |
8:56 | 750 | 316 | 0 | 138 | 178 | 0 | 4 | |||
9:49 | 1000 | 235 | 0 | 44 | 191 | 0 | 4-5 | |||
3 | 18/09/2010 | 12:43 | 500 | 494 | 0 | 0 | 20 | 474 | 3 | |
13:15 | 750 | 316 | 0 | 0 | 10 | 306 | 3 | |||
13:56 | 1000 | 234 | 0 | 1 | 117 | 116 | 2 | |||
4 |
19/09/2010 | 14:24 | 750 | 222 | 0 | 4 | 106 | 112 | 2 | |
14:50 | 1000 | 236 | 0 | 74 | 107 | 55 | 2-3 | |||
5 | 20/09/2010 | 14:27 | 500 | 475 | 0 | 407 | 30 | 38 | 2-3 | |
14:52 | 750 | 314 | 0 | 305 | 9 | 0 | 2-3 | |||
15:24 | 1000 | 231 | 0 | 231 | 0 | 0 | 2-3 | |||
6 | 20/09/2010 | 15:51 | 500 | 484 | 4 | 480 | 0 | 0 | 2-3 | |
16:17 | 1000 | 240 | 24 | 216 | 0 | 0 | 2-3 | |||
7 | 21/09/2010 | 12:24 | 500 | 494 | 0 | 10 | 98 | 386 | 0 | |
12:48 | 750 | 318 | 0 | 24 | 96 | 198 | 0 | |||
13:17 | 1000 | 236 | 3 | 130 | 81 | 22 | 0 |
a This was the only flight where the yellow filter was used on the SLR camera lens.
The
It was possible to identify a range of fauna within the images, including dugongs, dolphins, turtles, sharks, rays, seasnakes, fish schools and birds on the water surface. We note that dolphins and turtles could be identified to species level in many cases. However, the research reported here is focused on using dugongs as a case study so further analyses of other animal sighting data were beyond the scope this work.
Of all images captured along predefined transect lines, across all surveys, a total of 626 images contained sightings of dugongs. The total count, after eliminating all double counts from overlap of images along the transect line, was 1036 dugong sightings. Of these, 974 dugong sightings were identified with certainty, including 110 calves. Example images containing dugong sightings at each altitude are shown in
Example images containing dugong sightings (outlined in red), where (A) was captured at 500 ft, (B) was captured at 1000 ft, (C) is an example of where a dugong visible at the bottom of this image was not visible under the sun glitter in the successive image (750 ft), and (D) was captured during the worst wind conditions (750 ft).
The expected proportion of dugongs identified as “certain” (as opposed to “possible but uncertain/unclear”) was 0.95 (95% CI = 0.90, 0.98). This included all sightings, whether at the surface, near the surface or on the seafloor. These proportions did not differ significantly among the three altitudes (χ2 test statistic of ≈ 0;
Sun glitter was worst during the early afternoon, as shown in
Number of images within each sun glitter category (a subjective estimate of the percentage area of the image affected) according to the time at the beginning of the surveys, and the corresponding sun elevation angle (calculated using Solar Angle Calculate freeware [
When reviewing the images, the overlap between successive images helped overcome sun glitter issues (see Figure
The yellow filter did not appear to reduce the effect of sun glitter, with over half the images captured during the flight on 19 September having 50% or more of the image affected. The polarising filter would only have reduced glare in those images captured at the right angle relative to the sun (see Discussion). There was no obviously discernible effect of the polarising filter on any of the images captured.
The GLMM that best accounted for the effects of altitude, sun glitter, turbidity and Beaufort sea state on the dugong sighting rate included turbidity as the only covariate (this was with turbidity specified both as an integer, ordinal value (i.e., 1-4) and split into a ‘depth’ and ‘water clarity’ variable, with an associated interaction, as described above). Therefore, turbidity variables (both specifications) were the only ones found to significantly affect the dugong sighting rate within the images (
Intercept | -9.4191 | 0.2310 | -40.77 | ||
Turbidity: depth | -3.5806 | 0.2544 | -14.08 | ||
Turbidity: clearness | -1.2481 | 0.1273 | -9.81 | ||
Turbidity: interaction | 4.1117 | 0.4645 | 8.85 | ||
Flight | 0.39777 | 0.091 | |||
Transect | 0.27494 | 0.064 | |||
3.97985 |
From
Sighting rates of dugongs for each UAV flight (adjusted to dugongs per 1.0 km2 for ease of comparison) according to the four nominated turbidity levels (1 = shallow with the bottom clearly visible, 2 = shallow with the bottom visible but obscured by turbidity, 3 = deep with the bottom not visible, but clear water, and 4 = deep with the bottom not visible and turbid water); model estimates derived from fixed effects of GLMM and using depth × water clarity specification of turbidity. A comparison of model fit is shown where ○ is the model estimate and + indicates the upper and lower 95% confidence interval on the model estimate.
This initial trial of a basic payload system (digital SLR camera) has successfully demonstrated that the
Neither our dugong sighting rate, nor our ability to identify dugongs with certainty, were affected by the altitude at which the survey was flown. The
When conducting flights at 1000 ft the transect strip width we achieved (144 m) is narrower than is used during manned surveys, where usually, observers record sightings from 200 m strips on both sides on the aircraft (i.e., the total strip width is 400 m [
Turbidity was expected to influence sighting rates because this variable affects the proportion of time dugongs are available to be sighted, i.e., Pollock et al [
We were also expecting that sea state might influence dugong sighting rates. Sea state is known to affect the ability of human observers to sight dugongs during aerial surveys [
Sun glitter was visible in a high proportion of images, particularly during the early afternoon, but did not affect our dugong sighting rates. During manned aerial surveys, sun glitter also causes problems for observers and survey flights are typically scheduled for early morning and late afternoon to avoid intense midday glitter. In our UAV surveys, however, it appeared that overlap between successive images (along the transect line) overcame the problem of dugongs being masked by sun glitter. This strategy is used in vertical aerial photography for shallow water benthic habitat mapping, where it is suggested that 60% overlap between images compensates for sun glitter [
The disadvantage of having large amounts of overlap between images is the large number of images captured and therefore the large amount of memory storage needed, in our case, on board the
Using a polarising filter can also reduce reflections of light off the water (both glare and glitter). We recognised that one needs to orient a polarising lens in an appropriate direction relative to the sun in order to filter the reflected light. However, we oriented the lens in a constant direction to determine whether the filter would have an effect in at least some images. We did not detect any obvious reduction in glare or glitter in any particular sets of images. The greatest benefit from the polarising filter occurs when both the sun elevation angle and the camera angle are at 37 degrees from the surface of the water (Brewster’s angle [
In our introduction we listed a number of limitations of manned surveys that we suggest UAV surveys may overcome. The first – eliminating human risk – is clearly achieved, because no observers were needed in light aircraft. The second – reducing costs – is difficult to quantify as the commercial company we used sets rates according to the specific requirements of each job and therefore it is not appropriate to quote exact costs. However, during this early stage in the development of UAV methods, Insitu Pacific aim to make their costs at least competitive with manned surveys. Considering this, our trial suggested that UAVs could reduce survey costs by reducing the time needed to complete a survey through (a) the flexibility in acceptable wind conditions, (b) the option to overcome masking of sightings due to sun glitter by overlapping successive images, and (c) the ability to fly longer hours within a day without having to refuel or rest observers. A shorter survey time frame would provide a more robust ‘snapshot’ of animal distribution and abundance across the entire survey area because there is less opportunity for significant animal movement.
The advantage of having a permanent record of each sighting is that images can be re-checked to ensure sighting data are accurate, and strict criteria can be applied for eliminating uncertainties. Reviewing images in consultation with other experts can also increase the accuracy of species identification. Future assessments of UAV sighting data could include an effort to assess observer bias by having multiple experienced people review a subset of images. This idea is conceptually the same as having multiple observers in a manned aircraft; the aim is to maximise detection probability and calculate observer bias using a capture-recapture approach [
Manual review of the images is time consuming, however, and the efficacy of UAV surveys will depend on the development of image analysis algorithms to automate the detection of animals (or possible sightings) within the images. There are currently no published algorithms that can detect marine fauna in aerial images, although a number of research groups are attempting to develop software for this purpose. This problem is particularly challenging relative to other areas of image processing automation because (a) marine fauna often occur in images that are made complex by the sea floor, white caps and sun glitter, and (b) the size, colour and shape of the animal can change according to its position in the water column. Although we have suggested here that UAV surveys may not be limited by the same sea state and sun glitter conditions that limit manned surveys, our results apply only to images that have been manually reviewed. The development of software that can automate the processing of images from UAVs could allow us to quantify detection error of the algorithm as a proxy for observer bias, and eliminate the error caused by human fatigue (from reviewing images or observing during manned surveys). However, whether such an algorithm could overcome the challenges of high sea state and high glare conditions remains untested.
The next step in establishing the efficacy of replacing manned aerial surveys with UAVs is to determine whether the proportion of dugongs available to be sighted from the air at any one moment is equivalent for both methods. A great deal of work has been conducted to estimate availability corrections for manned dugong surveys [
Acquiring the GPS location of every image provides greater accuracy in location data than for manned surveys where observers are calling sightings. In strip transect sampling, observers typically do not call the position of the animal relative to the trajectory of the aircraft (i.e., how far forward or aft of abeam they sighted the animal). This sampling method is generally used for animals that surface briefly or occur in high numbers, meaning there is no time to measure the position of each sighting – it is either inside the strip or not. Therefore, if there is a 5 second window of opportunity to see an animal as the aircraft passes any one location, the animal could be located anywhere within a 250 m length of space. Even using distance sampling methods, where horizontal and vertical angles are measured, a few seconds delay in taking the measurements will affect the accuracy of the calculated location. During our UAV surveys, detailed data are recorded about the
During this trial we were not able to test the ability of the
Additional benefits of the UAV are the reductions in fuel consumption and potential noise disturbance. The
There is currently greater capacity to fly UAVs in civilian airspace in Australia than in some other countries such as the US, because of the regulations set by aviation safety authorities. However, the development of this technique provides alternatives in countries where UAVs are permitted and where accessibility to suitable aircraft and runways are limited. For such applications, a more mobile UAV system may be needed that can be easily transported internationally, but as stated above, the range and endurance of the UAV are also important considerations and can influence the size and transportability of the overall system. Camera systems can also be added to manned-aircraft surveys to gain some of the benefits listed. The trial and analysis presented in this paper, and many of the questions raised, are also relevant to that context.
This UAV trial showed that dugongs could be readily detected within images captured using the
The ScanEagle UAV system was provided and operated by insitu Pacific Pty Ltd. We specifically thank Neil Smith (insitu Pacific project manager), Rich Clifford, Carl Brown, Marty Evans, Peter Cassimatis and Lennon Cork for their various roles in preparing for, and assisting with, these field trials. We are grateful to David Holley from the Department of Conservation and Land Management, western Australia, for providing support for this project, including the use of the boat, which was skippered by Wayne Moroney. Thanks also to: Ken Pollock for his support, advice, and comments on this manuscript; Eric Kniest for his adaptation of his Vadar software for images collected from the UAV, and his comments on the manuscript; and to David Clifford for comments on a later version of this manuscript.