Achtergrond

Sea ice extent from satellite microwave sensors

In 2007, the summer extent of Arctic sea ice observed by the Special Sensor Microwave Imager (SSM/I) reached its lowest value on record since 1979.

Satellite sensors provide a convenient way to monitor the vast expanses of sea ice in the Polar Regions and this is done primarily using microwave techniques, which contrary to optical instruments, can operate at night and in all-weather conditions. Sea ice detection is routinely performed by two families of microwave instruments: radiometers, which observe the natural emission from the surface in the range of microwave frequencies, and scatterometers, which observe the energy reflected from a transmitted pulse.

Earlier work has shown reasonable agreement between scatterometer and radiometer sea ice extents, albeit with systematic discrepancies characterized by defective scatterometer extents during the sea ice growth season(1) and deficient radiometer extents during the summer melt(2). In this highlight, we revisit the determination of sea ice extents using satellite microwave sensors, propose an improved approach to sea ice detection for the SeaWinds scatterometer based on a Bayesian methodology, and compare its performance against other existing active and passive microwave algorithms throughout a full annual freeze/melt cycle. The improved Bayesian algorithm has been developed for inclusion in the EUMETSAT-KNMI SeaWinds Data Processor and is currently undergoing pre-operational testing. 

Figure 1. Incidence, azimuth and polarization diversity for the SeaWinds scatterometer(HH and VV refer to horizontal and vertical Transmit/ Receive polarization)
Figure 1. Incidence, azimuth and polarization diversity for the SeaWinds scatterometer(HH and VV refer to horizontal and vertical Transmit/ Receive polarization)

Sea ice detection with the SeaWinds scatterometer 

Scatterometers are active microwave sensors whose primary mission consists in the determination of near-surface winds over the oceans. Similarly to a radar detector, scatterometers radiate energy pulses and collect the backscattered returns from a variety of incidence and azimuth angles, providing a diversity of views that allows for the detection of the wind signature over the ocean (Figure 1). The dielectric permittivity of seawater at microwave frequencies is many times larger than that of sea ice. The wind induced ocean returns arise from the reflective seawater surface, while returns from the semi-transparent sea ice arise from volume interactions deeper in the ice slab. This results in distinct polarization, intensity and directional scattering properties that allow their effective separation. In particular, while ocean scattering is characterized by steep backscatter derivatives relative to the incidence angle, a remarkable azimuthal anisotropy and substantial polarization, volume scattering from sea ice results in smaller backscatter derivatives, azimuthal isotropy and nearly total depolarization. The first sea ice detection algorithms relied on hard threshold tests that capitalized on these properties. They evolved in time into Sea ice extent from satellite microwave sensors Maria Belmonte Rivas The determination of sea ice extents from satellite platforms can be exploited as a marker for climate change, a navigation tool and a reference for coupled ocean-ice-atmosphere models maximum likelihood methods with separate point-wise class clusters for mean backscatter, polarization and anisotropy combinations(3). 

The novel sea ice detection approach proposed here takes advantage of improved knowledge about the distribution of backscatter points in the scatterometer measurement space to replace the former point-wise class clusters by extended geophysical model functions (GMFs) for ocean wind and sea ice(4). For the SeaWinds case, the GMF for ocean winds is the one applied operationally to retrieve wind vectors over ocean surfaces. The empirical GMF for sea ice is drawn from the observed distribution of pure winter ice backscatter, which groups along an extended line in the SeaWinds dB measurement space { σHH,fore, σVV,fore, σVV, aft, σHH,aft}. The advantage of this approach relative to previous algorithms is that the spread of measurements about extended class model functions is smaller than that about point-wise class clusters, allowing class covariances to decrease to instrumental noise levels thus enhancing their discrimination power. Our Bayesian algorithm computes the normalized square distances (or maximum likelihood estimators, MLE) to the ocean wind and sea ice model functions as:

where σ° are backscatter measurements, σ°class are the class model functions and the normalization factors var[σ°class] guarantee that the variance of square distances about each class model is unity. The Bayesian posterior sea ice probability is then calculated as: 

where p°(ice) and p°(wind) are a priori sea ice and ocean wind probabilities, and the conditional probabilities p(s0|ice) and p(s0|wind) reflect the actual location of a measurement relative to the expected backscatter distributions about the ice and wind model functions, expressed in terms of the MLEs defined above. The Bayesian sea ice detection algorithm implemented for SeaWinds at KNMI generates daily sea ice masks using a std[σ°ice] = 1.5 dB and a 50% threshold to posterior sea ice probabilities. The masks are filled with backscatter strength values, which are indicative of ice type or thickness by proxy and archived for dissemination (see Figures 2 and 3). 

Figure 2. NH sea ice mask with backscatter values from SeaWinds (8th May 2009).
Figure 2. NH sea ice mask with backscatter values from SeaWinds (8th May 2009).
Figure 3. NH sea ice mask with backscatter values from SeaWinds (8th May 2009).
Figure 3. NH sea ice mask with backscatter values from SeaWinds (8th May 2009).

Scatterometer versus radiometer sea ice extents 

We would like to compare the sea ice masks observed from passive and active microwave instruments. The radiometer masks (from AMSR, the Advanced Microwave Scanning Radiometer) are calculated using the Enhanced NASA Team algorithm (NT2)9(5), while the scatterometer masks (from SeaWinds on QuikSCAT) are estimated using both the cluster method(3) and the Bayesian GMF algorithm developed at KNMI. It has been proven that passive microwave masks lie within 10 km of the ice edge observed by optical (and synthetic aperture radar) sensors in the wintertime(6), although accuracy becomes worse during the summer months, as these masks become affected by weather effects, unresolved thin/low concentration ice types and surface melt effects(5).

Figure 4. Arctic and Antarctic sea ice extents for Sep’06 to Sep’07 from passive microwaves (PM, dashed line) and active microwaves (AM, the dotted line represents the old cluster method, the continuous line the new Bayesian GMF method developed at KNMI).
Figure 4. Arctic and Antarctic sea ice extents for Sep’06 to Sep’07 from passive microwaves (PM, dashed line) and active microwaves (AM, the dotted line represents the old cluster method, the continuous line the new Bayesian GMF method developed at KNMI).

Having determined the overall convergence of daily sea ice extent estimates from satellite active and passive microwave algorithms during the fall and winter seasons, we are left to evaluate the nature of the observed discrepancies. For this purpose, we make use of higher resolution MODIS and Envisat ASAR imagery. While the extensive and frequent cloud cover in the Polar Regions is a factor against the use of optical techniques for the monitoring of sea ice conditions, the contrast between sea ice and open water is not always well-defined for the cloud penetrating ASAR. The combined use of optical and high resolution microwave datasets allows for more imagery to be used for an in-depth validation study. In general, the combined MODIS and ASAR records helps us confirm that during the winter months, under conditions such that the open ocean is terminated by a compact boundary of consolidated ice, all three algorithms come to agree to within 2 grid pixels or 25 km in their determination of the sea ice edge. Figures 5 to 8 illustrate typical discrepancies found between active and passive microwave sea ice detection algorithms, generally involving mixed ice/ocean scenarios due to thin, water saturated and low concentration sea ice conditions. During the growth season, thick and consolidated sea ice slowly progresses behind a rapidly advancing band of thin ice. The accurate determination of the ice edge in areas of active formation is difficult because the representativity of daily maps degrades rapidly, but also because new ice is so thin and saline that it resembles a smooth ocean surface. 

Figure 5. Sea ice detection discrepancies during the growth season (ASAR truth)
Figure 5. Sea ice detection discrepancies during the growth season (ASAR truth)

Figure 7 provides another instance of passive microwave summer biases, this time focusing on the development of large sea ice bands in the Southern Ocean, which are entirely missed by the passive microwave algorithm. Once more, scatterometer algorithms appear to be more robust in terms of detection of summer ice, partly due to the sensitivity of sea ice microwave emissions to surface melt effects. One last example is shown in Figure 8, taken from two sequential orbital passes of the MODIS sensor over the Baffin Bay and featuring a large but sparse ice floe field that is mistakenly labelled as a cloud field by the optical algorithm.

Figure 7. Sea ice detection discrepancies during the melt season (ASAR truth, Antarctic).
Figure 7. Sea ice detection discrepancies during the melt season (ASAR truth, Antarctic).
Figure 8. Sea ice detection discrepancies during the melt season (sequential MODIS overpasses 1.5 hours apart. White = ice, blue = ocean, magenta = land and cyan = cloud).
Figure 8. Sea ice detection discrepancies during the melt season (sequential MODIS overpasses 1.5 hours apart. White = ice, blue = ocean, magenta = land and cyan = cloud).

The sparse floe field is detected by neither the passive microwave nor the old cluster based scatterometer algorithms. However, the improved KNMI Bayesian sea ice detection algorithm sets out to provide a most inclusive and conservative definition of sea ice edge to date, one that is more in line with that provided by ship observations and well-suited for applications that require reliable indication of sea ice presence all year round, such as satellite-based retrievals of ocean wind and sea surface temperatures.

Conclusions 

KNMI has developed an improved Bayesian sea ice detection algorithm for the SeaWinds satellite scatterometer and has validated its performance against state-of-the-art passive microwave and scatterometer methods throughout a full sea ice growth/melt cycle. It is shown that, although all the methods under study agree very well during the winter months, the new Bayesian approach improves on the misclassification scores that affect earlier scatterometer and passive microwave algorithms and remains sensitive to the summer sea ice species that populate the Arctic edge during the melt season and the Antarctic margin all year round. The improved determination of the sea ice edge provides an all-inclusive daily sea ice mask recommended for use in ground processors that require an effective filtering of sea ice contaminated pixels. The Bayesian scheme has been successfully applied to SeaWinds and is currently under development for the MetOp ASCAT scatterometer. A reprocessing of the entire SeaWinds data record up to 1999 is also foreseen. 

Niet gevonden wat u zocht? Zoek meer achtergrond artikelen