134
D.P. WINEBRENNER, D.G. LONG, B. HOLT
as 2% renders previously (dry) microwave-transparent snow quite lossy; thus illumination of the strongly scattering bubbly upper layer of multiyear ice is reduced,
and scattered returns are attenuated by their passage through the snow. In the earliest stage of melt onset, this is all the microwave physics that matters. The fact that the
strong scatterer beneath the snow happens to be a bubbly layer is not essential - any
strong backscatterer would produce similar phenomenology because the shielding
by newly wet snow is the source of the change. The difficulty in sensing melt onset
on first-year ice at 5.3-GHz is the weakness of winter first-year ice backscattering.
The drop in backscattering due to shielding of an already weak scatterer is difficult
to observe, especially from space. (Note, however, that these comments apply only to
the earliest stage of melt onset; scattering from both first- and multiyear ice increases to high levels within roughly 2 weeks of melt onset, probably because the snow surface becomes water-logged and rough. Utilization of the information contained in
this latter signature variation is a topic of current research in several groups.) Spaceborne observation of melt onset on first-year ice might well be facilitated by observation at a wavelength where winter first-year ice cross sections are higher. We return
to this point in the next section.
Two considerations directly influence the design of machine automatable algorithms
for melt onset detection. First, the drop in backscattering cross sections varies slightly
between multiyear ice floes, but is large - typically 7-10 decibels - for virtually every
individual floe (Winebrenner et al. 1994). Thus it is not necessary to monitor the
backscattering cross-section history of individual multiyear floes to observe melt
onset; it is only necessary to derive from imagery information on the temporal
backscattering evolution of typical multiyear ice pixels. Second, near-surface air temperatures over Arctic sea ice are highly correlated over distances on the order of at least
100 km (Rigor and Munoz, "Statistics of Surface Air Temperature Observed in the Arctic", manuscript in preparation, 1996). Thus it is necessary only to sample Lagrangian
regions of size on the order of 100 km with SAR imagery to produce a map of melt onset
dates with useful spatial resolution.
For regions covered predominantly by multiyear ice at the time of melt onset, these
considerations suggest a simple algorithm. Divide the region of interest into fixed
(Eulerian) grid cells of the minimum dimension consistent with high temporal resolution of sampling with SAR imagery around the time of melt onset (preferable sizes
will be on the order of 100 km or less). Assign each SAR image to a grid cell, and compute for each image a histogram of pixels with backscattering cross-sections in narrow
ranges. Because the most probable surface cover is multiyear ice (in the selected
regions), simply track the backscattering cross section value of the peak (mode) of the
histogram through time; the time series of such peak values is a reasonable proxy for
multiyear ice backscattering cross sections under the stated conditions. Set a threshold
such that when the proxy for multiyear ice cross-sections crosses the threshold from
above, melt onset is inferred to have taken place within the entire cell. This algorithm
was automated and run using all ERS-1 SAR imagery over the Beaufort Sea collected
by the Alaska SAR Facility (ASF) during spring 1992, with the result shown in Fig. 3.
This particular map used cells of 200 km on a side on the grid defined for Special Sensor Microwave/Imager (SSMII) brightness temperatures. The algorithm required operator intervention in one cell, denoted with a numeral 2 in the map, where an early"false
start" to melt onset occurred - backscattering dropped briefly in late May but then
D.P. WINEBRENNER, D.G. LONG, B. HOLT
as 2% renders previously (dry) microwave-transparent snow quite lossy; thus illumination of the strongly scattering bubbly upper layer of multiyear ice is reduced,
and scattered returns are attenuated by their passage through the snow. In the earliest stage of melt onset, this is all the microwave physics that matters. The fact that the
strong scatterer beneath the snow happens to be a bubbly layer is not essential - any
strong backscatterer would produce similar phenomenology because the shielding
by newly wet snow is the source of the change. The difficulty in sensing melt onset
on first-year ice at 5.3-GHz is the weakness of winter first-year ice backscattering.
The drop in backscattering due to shielding of an already weak scatterer is difficult
to observe, especially from space. (Note, however, that these comments apply only to
the earliest stage of melt onset; scattering from both first- and multiyear ice increases to high levels within roughly 2 weeks of melt onset, probably because the snow surface becomes water-logged and rough. Utilization of the information contained in
this latter signature variation is a topic of current research in several groups.) Spaceborne observation of melt onset on first-year ice might well be facilitated by observation at a wavelength where winter first-year ice cross sections are higher. We return
to this point in the next section.
Two considerations directly influence the design of machine automatable algorithms
for melt onset detection. First, the drop in backscattering cross sections varies slightly
between multiyear ice floes, but is large - typically 7-10 decibels - for virtually every
individual floe (Winebrenner et al. 1994). Thus it is not necessary to monitor the
backscattering cross-section history of individual multiyear floes to observe melt
onset; it is only necessary to derive from imagery information on the temporal
backscattering evolution of typical multiyear ice pixels. Second, near-surface air temperatures over Arctic sea ice are highly correlated over distances on the order of at least
100 km (Rigor and Munoz, "Statistics of Surface Air Temperature Observed in the Arctic", manuscript in preparation, 1996). Thus it is necessary only to sample Lagrangian
regions of size on the order of 100 km with SAR imagery to produce a map of melt onset
dates with useful spatial resolution.
For regions covered predominantly by multiyear ice at the time of melt onset, these
considerations suggest a simple algorithm. Divide the region of interest into fixed
(Eulerian) grid cells of the minimum dimension consistent with high temporal resolution of sampling with SAR imagery around the time of melt onset (preferable sizes
will be on the order of 100 km or less). Assign each SAR image to a grid cell, and compute for each image a histogram of pixels with backscattering cross-sections in narrow
ranges. Because the most probable surface cover is multiyear ice (in the selected
regions), simply track the backscattering cross section value of the peak (mode) of the
histogram through time; the time series of such peak values is a reasonable proxy for
multiyear ice backscattering cross sections under the stated conditions. Set a threshold
such that when the proxy for multiyear ice cross-sections crosses the threshold from
above, melt onset is inferred to have taken place within the entire cell. This algorithm
was automated and run using all ERS-1 SAR imagery over the Beaufort Sea collected
by the Alaska SAR Facility (ASF) during spring 1992, with the result shown in Fig. 3.
This particular map used cells of 200 km on a side on the grid defined for Special Sensor Microwave/Imager (SSMII) brightness temperatures. The algorithm required operator intervention in one cell, denoted with a numeral 2 in the map, where an early"false
start" to melt onset occurred - backscattering dropped briefly in late May but then