Australian UAV was engaged by the City of Hobart to provide an colour map orthomosaic and Digital Surface Model (DSM) of the Hobart coastline. They also requested a proof of concept fine-scale dataset of an indicative section of coastline vulnerable to erosion. The Hobart Coastal Drone Survey was undertaken in two phases, the first in September 2016, and the second in January 2017.
Project Outcomes and Purpose
Drone Survey Process
Undertaking drone surveys involves three main steps. These are summarised in Figure 1.
Figure 1: The Drone Survey Process
Orthomosaic Creation Fixed-Wing Drone Survey
Aircraft: The most efficient drone survey method for mapping areas greater than a few hectares is a fixed-wing drone survey. The eBee RTK (real-time kinematic) drone was chosen for the fixed-wing survey as it is the best available aircraft based on its versatility and high degree of accuracy. The 17km of coastline was captured over two days with this aircraft.
Ground control: The eBee RTK requires reduced ground control as it receives real-time correction to the high accuracy on-board GNSS (Global Navigation Satellite System, AKA GPS) receiver to improve aircraft positional uncertainty from 2-10 m (standard navigation-grade GPS accuracy) down to a few centimetres. Along with the onboard RTK GNSS, there were additional ground control points captured using a Geodetic RTK GNSS. A denser ground control network would have been required had we been using standard GNSS positioning on a less advanced aircraft.
Mosaic creation: The image processing workflow combines the images and the ground control to generate two mosaics from 1952 images covering approximately 17 km of coastline, one in the southern section and one for the northern section. We flew at 120m altitude Above Ground Level (AGL) giving an average orthomosaic resolution of 3.62 cm/pixel. We are limited by the weight of the camera we can carry (which limits camera resolution and quality) and to this flying height as the Civil Aviation Safety Authority (CASA) regulations prohibit drones from flying about 120m without additional permits. This resolution compares to 5-10 cm resolution usually possible with traditional aerial imagery.
DSM Creation: In addition to creating the orthomosaic, the processing system produces a Digital Surface Model (DSM). This 3D terrain model is accurate to approximately +/- 9 cm in areas where the terrain is not obstructed by vegetation or water. The imagery was captured vertically (ie looking straight down, AKA nadir imagery).
Accuracy Assessment: To assess absolute accuracy, twelve verification points were measured across the mosaic (and height values were extracted from the DSM) at locations marked on the ground and compared to coordinates obtained with RTK Differential GNSS. The assessed accuracy of the mosaic is indicated by RMS (root mean squared) error statistics. These are:
RMS error X: 68 mm
RMS error Y: 54 mm
RMS error Z: 90 mm (for the derived DSM)
The absolute accuracy can be improved with more ground control, smaller ground sensing distance (flying lower) and, if required a more accurate ground control survey. The examples below highlight the improved clarity of features when captured from low altitude UAV as compared to traditional manned aircraft aerial imagery.
Figure 2: Standard manned aerial imagery of Sandy Bay Sailing Club (SBSC)
Figure 3: Fixed-wing drone imagery of Sandy Bay Sailing Club (SBSC)
Orthomosaic and DSM Creation Multi Rotor Drone Survey
As noted above an additional high density, low-level survey was undertaken in areas identified as potentially vulnerable to sea-level rise. This demonstration area provides improved detail but due to the aircraft type is really only suitable for use in small area capture. The data was used to further verify the data collected by the fixed-wing.
Aircraft: To provide an example of the level of detail possible with multi-rotor aircraft we flew a drone carrying a 20 megapixel camera on a gimbal over the Short Beach Reserve at Marieville Esplanade, Hobart, Tasmania at and approximate 30 m flying height.
Ground Control: Eleven Ground control measured with differential GNSS along with seven verification points.
Orthomosaic and DSM creation: This survey was flown at approximately 30 m AGL and both nadir and oblique imagery were captured. A total of 537 images were captured and processed to produce an orthomosaic with an average resolution of 6.1 mm/pixel.
Accuracy Assessment: The seven verification points were measured across the mosaic (and height values were extracted from the DSM) and compared to coordinates obtained with RTK DGPS. The assessed accuracy of the mosaic is indicated by RMS error statistics. These are:
RMS error X: 13.8 mm
RMS error Y: 11.9 mm
RMS error Z: 21.3 mm
As these statistic indicate, the low altitude multi rotor survey is much more accurate in all three dimensions with the very high density of ground control and the images below provide an indication of the clarity of the data. The time taken to collect the data however, and therefore cost, demonstrate the importance of understanding the required output accuracy requirements and tailoring the methods to the budget to achieve the best possible result that is fit for purpose.
Figure 4: Standard aerial imagery of the mouth of the Sandy Bay Rivulet at Short Beach
Figure 5: Fixed-wing drone imagery of the mouth of the Sandy Bay Rivulet at Short Beach
Figure 6: Multi rotor drone imagery of the mouth of the Sandy Bay Rivulet at Short Beach
Viewing these datasets online using a three-dimensional interface can provide valuable insight into the data which are not obvious when viewing as a top-down map. They provide measurement tools for calculating lengths, area and volumes as well as volume change between time series snapshots.
Figure 7: Top down view of Multi rotor drone derived 3D Textured Mesh (the mouth of the Sandy Bay Rivulet at Short Beach)
As well as viewing as mapping and 3D model deliverables, it is also possible to present the high-resolution oblique imagery of in an inspection interface, allowing you to view the source imagery from the drones perspective, and even place notes into the images for collaborative viewing.
Figure 8: Inspection view of the jetty steps near Short Beach Reserve
Key Considerations for Survey Design and Change Monitoring
Drone surveys that use imagery to generate orthomosaics and DSMs require careful planning. As explained in Figure 1 the flight planning and ground control density and distribution are key considerations in relation to the required accuracy metrics, along with the reliance on good weather conditions, and good take off, landing and aircraft monitoring sites. The final stage of data processing is dependent on the application in mind and the requirements of each dataset with respect to their accuracy, detail and scale. For change analysis the data capture must be co-registered, repeatable and comparable.
- For co-registration the data captured in one survey must be in as close to the same absolute location, once georeferenced, as any previous and subsequent surveys, sometimes this requires permanent co-registration features visible in each survey. Manual identification of targets in the photography is a key step and requires careful marker placement for each control point. Clear, identifiable targets are needed to ensure accurate registration and co-registration.
- Repeatability is difficult when variability at the pixel level (each pixel is 5-15 mm for multi rotor surveys from 20-50 m AGL and 25-40 mm for fixed-wing surveys from 80-120 m) is impacted by changes in lighting, vegetation cover and changes in the target surface such as erosion.
- This in turn dictates how comparable the dataset produced is because any difference is either due to actual change or error (false change). To see change at the centimetre level the dataset needs to be accurate enough in each survey to allow for reliable change detection, realistically this might require sub centimetre datasets and careful cleaning of processing artefacts (discussed later).
Data capture frequency: Some sites will need more frequent/fine scale data capture than others. The frequency of data capture is dictated by the range of applications for which the data will be used.
- Event-based monitoring provides insight into the impact of storm events, high tides, etc.
- Flying at regular intervals enables change monitoring, however it is difficult to ascertain the cause of changes. The change detected may be an accumulation of multiple large scale event-based changes or it may be a gradual change that is only detectable over time (once the change exceeds the threshold of the change detection possible for that data resolution).
Deciding on scale: When deciding on scale, the key question to answer is ‘what scale of change are you expecting to monitor?’
- Ultra-fine scale monitoring (2-5 cm) is expensive but allows for more detailed change analysis. Ultra-fine scale datasets require much more intensive mapping missions and produce detailed snapshots of the terrain (and vegetation surface). In areas where small changes can have large long term impacts these higher cost surveys can provide critical insight into the cause and impact of changes such as erosion.
- Courser scale monitoring is suitable where changes greater than 5-1o cm are anticipated. They are more cost-effective and provide valuable insight into whole of coast change.
- Combining the two approaches can allow ongoing monitoring of the entire coastline that is supplemented by fine scale mapping of focus sites to give insight into changes along remaining shorelines that portray similar characteristics.
Ground Control: As mentioned, ground control is a key consideration, the more control the better (ideally 8-15 per flight for the ultra-fine data, unless using RTK on the aircraft), but the cost increases. There comes a point where adding more control does not improve the result.
- Permanent control can save money over time when ongoing monitoring is required as it only needs to be surveyed once.
- Temporary control needs to be laid out and surveyed on the day of the flights and flying is very weather dependant, this adds a significant planning hurdle.
Another aspect of temporary control is ground control point target disturbance. There is a significant risk of markers being disturbed by passers-by or animals who can nudge, flip or pick up the markers. It is therefore important that they are securely placed and and signposted to not be disturbed.
Safety and Privacy: The City of Hobart coastline poses a number of challenges common to urban drone surveys including safe take-off (launch) and landing sites, public safety and maintaining line of sight. In some cases imagery will include people and private property. With respect to safety and privacy, it is important to understand that the laws that govern drone safety are defined by CASA and the laws that govern privacy are defined at both state and federal level and in some cases at the local council level. Drone surveys require image capture and sometimes that capture is from overhead (nadir) and at other times oblique imagery is captured. The goal is to capture sufficient imagery to accurately map the terrain and generate an orthomosaic that covers the entire area of interest. Inadvertent capture of people and their houses is a by-product and is not a premeditated attempt to invade their privacy. When undertaking a drone survey near the public it is important to brief them if possible/feasible, but the client will need to decide on the level notification required.
One option is to notify everyone of the planned survey. This may be a very large undertaking and because drone surveys are very weather dependant it would be impossible to predict exactly when the drone flights would occur (often the decision to fly is only made minutes before the first flight). Fixed-wing surveys flying at 100-120 m AGL are certainly less invasive and it is unlikely that the public would not even be aware of the high flying drone. Multi rotor surveys are much more noticeable, particularly as they are usually flown at 30‑50 m where the drone is very visible and within earshot. These close range surveys are perhaps the most in need of a public notification process.
It is possible to mask the images or edit the datasets produced to remove any portions that raise privacy concerns. This may be necessary when the drone survey outputs are publicly accessible.
Data processing and clean up: Another key consideration is the resourcing for data processing and clean up. The datasets produced from the drone survey data are not perfect representations of reality. The mosaicking can create artefacts along the mosaic boundary. Similarly, the trees and shadows can create false terrain due to the way the processing matches them. This is unavoidable and to clean it up the data needs to be edited to remove these sections. When the false terrain is in the water, the decision of where the water stops and the shore begins is critical if further analysis is planned and careful definition of the criteria for this shoreline delineation is necessary.
The flying component of the drone survey is only the first stage and the cost and implications of asking the drone survey provider to define the shoreline must be carefully considered as it may be better for a geomorphologist or in-house staff to make the decision. This may require the use of 3D editing software as traditional GIS software such as ArcGIS do not provide true 3 dimensional data portrayal (in most GIS an X,Y location can only have one Z (height) value, this is called 2.5D and prevents the display of overhangs and complex 3D shapes). 3D software can be used to post process the data and clean up water artefacts and other artefacts resulting from shadows, reflections, homogeneous surfaces and vegetation (photography based 3D surface model creation algorithms struggle to penetrate through vegetation and so the surface model is the “top” of the vegetation, not the ground beneath it). Australian UAV are proficient in editing this data and liaising with the client to ensure the best possible result is achieved.
- The drone surveys carried out for the Hobart Drone Survey 2016 provide valuable benchmark datasets at two scales. The fixed-wing survey can allow changes greater than 5-10 cm to be detected and the multi rotor survey can allow change detection at the sub 3 cm level (for unvegetated terrain). The orthomosaic and 3D terrain DSM produced are a useful snapshot of the coast and the fine scale survey of Short Beach Reserve shows the power of close range drone photogrammetry.
- The density and distribution of ground control impact accuracy and cost, there is a fine balance between the two.
- Similarly, fixed-wing surveys are a faster, cheaper way to map, but the flying height and subsequent resolution result in datasets that may not be suitable for fine scale change monitoring. CASA Regulations make flying lower that 30 m in areas where the public frequent difficult, but these safety precautions are critical if drone mapping is to become ubiquitous. Similarly, flying above 120 m could increase coverage, however drones cannot usually share airspace with manned aircraft in Australia and so over very large areas (>10,000ha) there comes a point where manned mapping flights become more cost effective.
- In seeking to understand what drone surveying can do for you it is critical that you evaluate each potential project and decide on the most appropriate data scale/resolution for each application as cost increases significantly when a project requires ultra-fine scale mapping (sub 2 cm).
- Accurately mapping infrastructure and the shoreline is certainly possible and the datasets produced here demonstrate this.
- The cost and time saved in using drones to map when compared to traditional surveying techniques is significant.
- The data is not a perfect representation of reality and there are occasional artefacts from the photogrammetric processing, particularly in areas of dense vegetation and water. The clean-up and data extraction phase are time consuming and rely on expertise, you need to assess your in‑house capabilities in this regard or use an external provider with a demonstrated technical background to ensure that you can make the most of the data whilst keeping costs down.
- LiDAR (laser scanning) technology is progressing fast and may be a more viable option in future, but at this stage the photogrammetric approach provides the most accurate models of unvegetated terrain (this is discussed further on our website, see Drone Data vs LIDAR: Clarifying Misconceptions, http://www.auav.com.au/articles/drone-data-vs-lidar/).
Change detection relies on accurate and timely data capture, drones are set to become the standard mapping tools for councils and governments the world over. It is fantastic that City of Hobart has lead the way in evaluating this technology and AUAV are confident that we can provide the drone services needed, these impressive first datasets are hopefully only the beginning.
13th February 2017
Harwin, S., Lucieer, A., and Osborn, J. (2015). The Impact of Calibration Method on The Accuracy of Points Clouds Derived Using Unmanned Aerial Vehicle Structure-from-Motion (UAV-SfM). Remote Sensing, 7(9), 11933-11953, http://www.mdpi.com/2072-4292/7/9/11933.
Turner, D., Lucieer, A., and de Jong, S.M. (2015). Time series analysis of landslide dynamics using an unmanned aerial vehicle (UAV). Remote Sensing. 7(2): 1736 – 1757, http://www.mdpi.com/2072-4292/7/2/1736.
Harwin, S., & Lucieer, A. (2012). Assessing the Accuracy of Georeferenced Point Clouds Produced via Multi-View Stereopsis from Unmanned Aerial Vehicle (UAV) Imagery. Remote Sensing (UAV special issue), 4(6), 1573-1599, http://www.mdpi.com/2072-4292/4/6/1573/pdf.