Below, highlights from a very readable N-ICE2015 paper relevant to the current melt season, on how to interpret radar imagery (such as Sentinel-1AB) based on ground-truthing by simultaneous ship, helicopter and ground measurements co-located to a satellite track. Here the goal is unattended interpretation of scenes as leads, froze-over leads, pressure ridges, rafted ice, snowed-over features, nilas, open water etc (ie assignment of each image pixel to one of those classes).
The work was bedeviled by frost flowers (below, first frame of 1st gif), internal layers of wetted and refrozen snow, and drifted snow masking features; considerations that raise further questions about ice thickness algorithms that are already mutually inconsistent (2nd frame and 2nd gif).
On these forums, we make two main uses of satellite data. The first is literal, accepting as-is photo-like products, such as a Landsat scene of Nares Strait. We might tile these together, make a time series, or dink with contrast or color palette but don’t often apply a segmentation (classification) algorithm ourselves that allocates each pixel to a bin. Instead we rely on others to provide those products.
For example, NSIDC sea ice age uses data from 6 satellites + buoys to bin Arctic ice into five age classes (which are highly correlated with ice thickness
http://www.mdpi.com/2072-4292/8/6/457/htm). Bin occupancy is then scored and graphed, showing the older ice classes pinching out over time in recent years (3rd image below, underlay). It is very rare on the forums to see a palette scored but common to see bin integrals (eg piomas volume summed over thickness cells).
Every segmentation product has
four distinct parts: initial satellite image, product map colored by bin, color key to bin definitions, and bin occupancy graphic. Within university-grade cartographic GIS,
if the classification results in N bins, the product map uses exactly N colors which are exactly those of the color key. Outside the map, embedded text or grid lines might be dithered (or offered as a separate layer) but colors within the map (or bin usage graphic) never stray from the palette. If they don't conform, the end user is
just guessing at what map colors represent.
Image classification can be done either by defining categories in advance (eg land, open water, slush or ice) using pixel properties of fiducial areas to partition the rest of the image, or by algorithmic ab initio determination of best N bin segmentation with class interpretation left to the investigator. Both methods develop issues over time series and even year-on-year.
Combined observations of Arctic sea ice with near-coincident co-located X-band, C-band, and L-band SAR satellite remote sensing and helicopter-borne measurements
AM Johansson JA King, AP Doulgeris, S Gerland, S Singha, G Spreen and T Busche
doi: 10.1002/2016JC012273 (see doi:10.5194/tc-11-755-2017 for related article on melt-pond sensing)
Here a low-flying helicopter (with camera and altimeter) tows a device that induces and detects eddy currents in (conductive) sea water but not in snow or ice (non-conductive unless briny) as various radar satellites pass overhead in mid-April 2015 north of Svalbard. This yields 40m pixels of snow + ice thickness and snow surface roughness that can be compared to same-afternoon satellites (L-, C- and X-band) passing overhead. Here the radar sees the ice, the altimeter the snow surface, and the magnetic field sensor the salt water surface.
The radar is polarized and channel ratios prove quite informative. In conjunction with surface texture, taken as kurtosis (heavy-tailedness) of the roughness distribution, the ‘extended polarimetric feature space’ segmentation algorithm can produce 13 sea ice bins (those specified by the WMO in 1970), though that was simplified to 4 here that aren’t quite any in that system.
The drifting RV Lance saw one of these -- refrozen leads -- develop asymmetric thickness on the downwind side. A second lead had sheets of young grey-white ice with some rafting and a thin layer of snow and frost flowers with thin ice having elevated roughness corresponding to deformation and edge effect transitioning between thin and thick ice.
They also encountered a snow-free lead with barely solid first-year ice adjacent to nilas/young grey ice, followed by wind-blown nilas and young grey-white ice with frost flowers. Here the GoPro camera on the helicopter and snow pit studies were essential as radar struggles with such subtleties (4th image).
Some scenes have large-scale ridge structures. Variations in ice thickness of up to 5.5 m documented by the helicopter unfortunately had no counterpart in radar backscatter despite non-vertical incidence.
Ridges increase surface roughness and indeed the camera saw evidence of old, consolidated ridges now completely snow-covered. These have a star-shaped signature that forms from snow drifts but those are more easily detected by eye.
The snow had undergone thaw and refreezing by mid-April, presumably from a rain-on-snow event giving an ice layer at the surface, later buried by more recent snowfall. This created a uniform ice layer within the snowpack but above the main ice which then
fools the radar return; this ice crust was confirmed within the snowpack in snow-pit studies. If the non-coastal piomas thickness blob is artifactual, this is a conceivable explanation.
In short, the authors see a limited potential for future automatic classification of SAR images to distinguish different ice classes because the ice is complicated enough already and processes that fool the radar could be more common as multi-year ice diminishes (or vanishes) as the Arctic amplification of warming precedes.
All the current segmentation schemes like extent, area, age, and volume will need to be seriously tweaked to keep up with once-rare ice formations and interactive atmospheric processes becoming dominant. It may not be feasible however to maintain time series if new bin classes become necessary.