Conference PaperPDF Available

Exploiting digital imagery for snow surface retrieval on sea



We show how imagery from uncalibrated airborne cameras can be used to reconstruct the snow/air interface on Antarctic sea ice, using data collected on the SIPEX-II research voyage during austral spring 2012. Imagery collected by an airborne surveying package was used to develop a 3D surface model using a structure-from-motion approach. This model was validated using coincident airborne LiDAR and in situ observation of total freeboard. Our study demonstrates that equivalent surveys may be obtained using unmanned vehicles (drones) carrying only a camera and basic navigation equipment. Using this method, detailed floe-to-multifloe scale models of snow topography may be derived without logistically intensive airborne surveying programs. In turn, this allows for quick repeat surveys - simplifying the capture of a surface topography time series at any given field research site. It also allows for highly detailed analysis of relationships between surface features and how the evolve over time. Finally, we show how different surveying scenarios affect data quality and the ability to easily co-register surface models with other coincident datasets. We discuss how future surveys should be planned, which data need to be collected alongside the imagery used to generate 3D models, and where future development should be aimed at in terms of uncertainty computations and data quality assessment.
This talk goes all the way back to a project executed between 2007 and 2012, where the Australian Antarctic Division deployed an
airbore LiDAR and camera in a modified helicopter (aka close range piloted drone). It was called the RAPPLS package see:
The RAPPLS instrument package surveyed a lot of sea ice. Green polygons show flights undertaken in 2007; pink polygons show 2012
flights. Three additional flying seasons (2008, 2009, 2010) took place in or near Prydz Bay.
Digital aerial photos have many uses. Here, locally georeferenced imagery is used to show an interpretive map of an ‘ice stationa
survey plot, a transect line, and the trajectory of an under-ice UAV.
Images can also be used to learn about the ice being flown over. This shows an early attempt at object-based image analysis (OBIA)
which segmented images using homogeneity ciriteria to split and merge groups of pixels (objects); and assigned classes based on the
mean spectral and textural properties of those objects.
Aerial imagery also does not need a huge, capital intensive platform. The APPLS package was amazing, but you could buy a new ‘out of
the box’ for every couple hours of flight time (with attendant increase in level of attention required for data quality)
…and since 2006, we’ve been able to reconstruct sea ice in 3D from imagery! Here, we see LIDAR elevations in green/blue at about 1.5m
point spacing and reconstructed-from-photographs elevation in blue-red at about 10cm point spacing. The line of orange dots is a drill
hole transect.
Why do we want 10cm point spacing? To see greater detail in surface topography. This slide shows we make a lot of mistakes in the
shape of topography, especially trying to model ice draft from elevation. On the left is airborne LiDAR with the top panel showing
topography, the middle panel showing ice draft modelled from that topography, and the bottom panel showing the actual under ice
topography from upward looking SONAR. The right side shows the same, using topography reconstructed from imagery.
The point here is that if we have high resolution data, we can try to make better guesses about relationships between classes of surface
features (dunes, ridges, other things) and what happens underneath. If we get the shape underneath the ice wrong, we
also get things like turbulent heat flux to the ice wrong (see the paper given in the slide).
This slide shows that we can actually reproduce topography well drill holes, LiDAR and imagery-derived terrain all match up. A constant
offset was required for the photogrammetric elevation data because the height of reference points used was not measured well in the
If we go beyond 2D, we are also OK. This slide also hints at a future for in situ drilling programs. If we are observing the top and the
bottom of the ice floe using airborne and under-ice mapping sensors, we can adjust our drilling strategy to provide better ‘control point’
geometries, and focus on ‘interesting features’ rather than wear ourselves out drilling in a straight line. Slide 7 also shows a general
phenomenon we only drill the flattest part of a floe!
In the RAPPLS programme, we collected imagery for different purposes and found that we could use it to reconstruct ice and snow
topography. More recent programs fly small UAVs specifically for the purpose of mapping snow and sea ice topography.
Here, we see that a dedicated imagery collection program for topographic reconstruction is comparable to terrestrial laser scanning
and with more complete coverage (no holes where the scanner beam is obstructed by ridges)
Focusing now on snow reiterating the point made in slide 7 about the benefit of increased detail. Even 1.5m point spacing LIDAR
misses important detail and variability in snow cover!
Adding detail in surface topography lets us estimate snow depth in more detail -here we got the best match with sonar-derived ice draft
from highly detailed (10cm) photogrammetric topography
We also want to know about patterns of snow distribution. Here we can see an estimate ofsmooth topography, which we might class
as snow dunes and ‘rough’ topography, which we might categorise as deformation ridges. Each has different snow depth distributions;
and different elevation characteristics (for example, total freeboard of ‘smooth’ regions was lower than ‘rough’ regions).
We also want to know how our different snow and topographic regimes map to what happens below the ice!
In turn, this detail will help us make better snow maps and focus field effort on specific snow and surface topography types instead of
trying to sample massive grids.
A text slide. Notes not required!
…and this detail makes for awesome outreach tools. The URL is now live, please check it out!
As for slide 18, from under the ice.
A few future notes.
Ver y key slide! For targeted image ry collec tio n c ampaigns , create very visib le ground co ntrol poi nts. The re is a GP S a ntenna in here
representing the origin of a floe-local coordinate system. But where is it? Next time, I’ll paint a huge target on it (and characterise the
effect of the paint on GPS signal prior to field deployment).
Use all the tools we can find be diverse, not dogmatic. Use both machine learning and physics-based approaches!
Reiterating points made earlier coordinate ground and airborne sampling for efficient use of resources
Also mine existing data!
…and this!
Thanks. The online talk is underway!
ResearchGate has not been able to resolve any citations for this publication.
ResearchGate has not been able to resolve any references for this publication.