I recently published an article in Open Access Government about the potential for machine learning technologies to revolutionise Polar science, with focus on optical remote sensing data from drones and satellites. You can read it online or download it from OAGov_Oct18
Camping on the ice sheet in September/October was a new experience – I’d never seen darkness on the ice before! The lack of light pollution and cloud-free skies made for a truly spectacular display of the Northern Lights. It was -25C and 35 knot winds pretty much constantly, so it was a constant battle between wanting to get into a tent and warm up and not wanting to miss a second of watching the aurora dancing over the milky way, with passing satellites and the occasional shooting star.
While on a personal level this was an incredible treat, it also presented some pretty major challenges for working with drones on the ice. The aurora knocked out the radio communications linking our drones to their controllers, meaning they could only be controlled over local wifi, reducing their range from a few hundred metres to about 30!
A new paper, led by Johnny Ryan, shows that a consumer grade digital camera mounted to a drone can be used to estimate the albedo of ice surfaces with an accuracy of +/- 5%. This is important because albedo measurements are fundamental to predicting melt, but satellite albedo data is limited in its spatial and temporal resolution and ground measurements can only be for small areas. Methods employing UAV technology can therefore bridge the gap between these two scales of measurement. The work demonstrates that this is achievable using a relatively simple workflow and low cost equipment.
The full workflow is detailed in the paper, involving processing, correcting and calibrating raw digital images using a white reference target, and upward and downward shortwave radiation measurements from broadband silicon pyranometers. The method was applied on the SW Greenland Ice Sheet, providing albedo maps over 280 km2 at a ground resolution of 20 cm.
This study shows that albedo mapping from UAVs can provide useful data and as drone technology advances it will likely provide a low cost, convenient method for distinguishing surface contaminants and informing energy balance models.
After testing the UAV performance in Svalbard in March, I realised the original ‘tripod’ landing assembly was not going to cut it for work in the Arctic. To prevent damage from landing in cryoconite holes and to spread the drone’s weight when landing on snow, I have added some ski’s modified from off-the-shelf landing gear for RC helicopters. This also has the added advantage that if one attachment point fails, the UAV is still landable, which is not the case for the tripod design.
As well as the ski’s, I have now added the Red-Edge camera’s down-welling light sensor to the top of the casing. This will automatically correct the images for changes in the ambient light field in each wavelength.
Here are a few things I learned after ten days of field testing the UAV multispectral data acquisition in Svalbard…
Video showing take off in stabilize mode, switch to loiter mode at about 5 m, quick control test then into automatic mission.
1. The UAV is surprisingly robust.
The aircraft was transported to the sites on a sled on the back of a snowmobile, hitting sastrugi and general lumps and bumps in the landscape, was hold baggage on three flights, launched and landed in snow, flown in winds and operated at temperatures down to -25 degrees. It is still is perfect working order and flew every mission without issue. I’m now pretty confident in the flight case and arrangement of kit inside, and trust that there won’t be significant flight issues in Greenland in summer.
2. Very low temperatures rapidly deplete the LiPo batteries
One noticeable, and unsurprising, effect of flying in these conditions was that the battry ran down very quickly. The toughest day was about -19 but with strong winds, and we only managed a 2.5 minute automated flight before the battery voltage dropped low enough for me to get twitchy and land the UAV manually.
3. A dGPS and ground control points will be essential for ground truthing the UAV imagery
It was impossible to accurately pinpoint ground sampling locations using handheld GPS and ground feature ID. This was especially true in Svalbard because of the homogenous snow cover, but will also be an issue in Greenland in the summer. This is not really surprising, but the need for an accurate dGPS location lock has been reinforced by the test flights.
4. The workflow for geotagging multispectral images using Photoscan is not as straightforward as I originally thought.
The red-edge camera does not output GeoTIFFs – they have to be postprocessed with ground control point data before they can be stacked and aligned. This is not a big problem, but good to know in advance of summer data collection.
5. New landing gear needed
The three-leg solution currently used on the UAV is better than I expected, but there is still the real problem that the legs sink into soft ground (e.g. snow) which risks pushing the camera lenses and the internal electronics into the snow. This could scratch the lenses or soak the electronics when the snow melts. Also, with the current model, a problem with any one of the three legs compromises the whole UAV because it becomes impossible to land flat. I’m going to develop something new before Greenland.
Video showing manual landing on a patch of compacted snow after an auto mission. It would be more stable on soft ground with a skid-type landing gear rather than the three legs. Includes heckling by Tedstone…!
6. Some a priori knowledge of image area can be useful
High resolution imagery is not available for all locations – for some of our field sites there was not sufficient google earth coverage to orient ourselves or draw a polygon by eye in Mission Planner. It is possible to estimate using the compass and the map scale, but with some existing knowledge of the GPS coordinates of the grid perimeter would help to plan an accurate mission.
7. It’s amazing how small the UAV looks, even when flying quite close by
Even flying the UAV at 30 or 40 m elevation, it quickly becomes difficult to keep track of its orientation when it flies a hundred metres or so on a mission. This does have implications for the length of mission I’d be comfortable flying, since I want to be able to rescue it manually if there are any GPS or autopilot issues – maybe I’m soft but my comfortable range is less than the CAA ‘dronecode’ distance limits- even though these do not currently apply in Svalbard/Greenland. Of course, the conditions (esp. visibility) affect what feels comfortable.
8. The controller is awkward in gloves
At -25 C gloves were pretty essential, but it is also difficult to have fine control over the switches and sticks on the controller. I was flying in a very thin pair of gloves or gloveless, which meant my fingers quickly went numb, especially when there was any wind. This will be less of a problem in Greenland in summer, but I will still get some warmer, thinner gloves with rubber finger pads to help with the UAV control in the cold.
9. Pre-flight checklists are invaluable
It’s so easy to overlook or forget something in harsh conditions or when rushed or excited. The written checklists developed before we went out to Svalbard were extremely useful for making sure everything went smoothly. These are a condition of CAA compliant flights in the UK and our experience in Svalbard demonstrates why! I will probably add a few additional checks or reorder a few things before our Greenland deployment – site specific things like take off and landing zone preparation (in Svalbard it was often a compacted snow platform, in Greenland I intend to use plyboard to avoid melt ponds and cryoconite holes).
The past few weeks have been spent working down in the robotics department at the University of Sheffield building a UAV (unmanned aerial vehicle, a.k.a drone). Ultimately, it will be used to make measurements of spectral reflectance of the ice surface in Greenland. It’s been great fun working in robotics – entering the lab is like walking onto the set of Robot Wars! UAV expert Owen McAree has been a huge help in developing the hardware and software for the drone – affectionately known as ‘albedrone’ in recognition of the albedo work it will enable – and we have now made successful test flights.
It began as an off-the-peg Steadidrone Mavrik quadcopter. However, we have made several modifications. We added a new brushless gimbal powered from the autopilot, machined a new mount that allowed us to better balance the camera and minimize the power being drawn by the gimbal, and added a GPS that can be used to trigger the image capture. We have also invested significant time into tuning the flight parameters and making it as stable and easy to fly as possible.
Flying can still be quite challenging, so I also invested in flight simulator software that interfaces with the real UAV controller, meaning I have been able to get the hang of flying safely without endangering the UAV. Significant time and effort has also gone in to writing a flight manual and logbooks for the batteries, build modifications and flight records.
We have been flight-testing the UAV at the University of Sheffield’s High Bradfield site and have now successfully made a pre-programmed flight and captured overlapping images in five spectral bands. Some examples are shown below. These are interesting as they were captured over an area with a thick cover of green vegetation, perfect for NDVI analysis. Next jobs are to a) keep modifying the UAV to extend the flight time and perhaps add some additional sensors, and b) test the software that will stitch the images and analyse the spectral information…