2.1. Background and detection algorithm

Time series of brightness temperature measured by microwave radiometers present very clear peaks throughout the summer in areas where melting occurs, especially at 19 GHz in horizontal polarization (Reference Zwally and FieglesZwally and Fiegles, 1994). We briefly explain the physical background for this phenomenon and its application to surface melting detection.

The brightness temperature T _b is the product of the emissivity ε and the surface/subsurface physical temperature T _s:

This equation is a first approximation and neglects atmospheric effects and snow penetration. However, for dry snow, emission originates from a layer extending from the surface down to a few meters deep (Reference Sherjal and FilySherjal and Fily, 1994).

Emissivity of dry snow is low, typically 0.65–0.80 at 19GHz in horizontal polarization (Reference Surdyk and FilySurdyk and Fily, 1995), because upwelling waves emitted by the lower layers are scattered by the upper layers and become trapped by the medium. When snow melts at or near the surface, the presence of liquid water increases the absorption. The emissivity increases to about 0.9, resulting in an increase in brightness temperature by about 40 K (physical temperature is around the melting point). This amplitude is large and cannot be caused by seasonal variations of the physical temperature (Reference Zwally and FieglesZwally and Fiegles, 1994; Reference CagnatiCagnati and others, 2004). Therefore, detecting such a clear signature in the signal can be performed by a simple algorithm using a threshold: a pixel is classified as melted if T _b is above a fixed or adaptive threshold T _b0.

In practice, however, difficulties arise. The resolution of space-borne microwave radiometers is coarse, a few tens of kilometers at 19 GHz. When a pixel is partially melted, the increase in T _b is lower than 40 K. Detecting such weak events requires a low threshold that in turn increases the risk of confusion between an increase really due to melting and an increase of the physical temperature (see Equation (1)). The solution adopted in our algorithm consists in using an adaptive threshold. The threshold is different for each pixel and depends on the winter brightness temperature mean and standard deviation (see Reference Torinesi, Fily and GenthonTorinesi and others, 2003) so that only an increase in T _b well above the typical local seasonal variations is classified ‘melted’. The algorithm settings ensure a good compromise between detecting weak events and avoiding false alarms (unmelted pixels detected as melted). Also, in some particular conditions, refreezing creates thick ice layers. These ice layers do not transmit upwelling waves emitted by the lower layer. The emissivity is then reduced with respect to the previous season (Reference Abdalati and SteffenAbdalati and Steffen, 1998). During the following summer, the emissivity is still lower than usual, and peaks in T _b caused by melting are then lower than usual. They may be missed by the algorithm using a constant threshold as in Reference Zwally and FieglesZwally and Fiegles (1994). Our algorithm deals with this issue by adapting the threshold every year.

A detection technique based on wavelet transforms has recently been proposed (Reference Liu, Wang and JezekLiu and others, 2005), but no extended inter-comparison has yet been performed. At this point, we note that the sensitivity/robustness balance is different between the two algorithms. Large differences in the number of detected melting events are observed, but the variations and trends are similar (see, e.g., CMS in the peninsula in Reference Picard and FilyPicard and Fily (2006) and Reference Liu, Wang and JezekLiu and others (2006)).

It is worth noting that detection of surface melting in Greenland uses a different approach: it uses a ratio of brightness temperatures at different frequencies (37 and 19 GHz instead of 19 GHz only) (Reference Abdalati and SteffenAbdalati and Steffen, 1997) rather than a single brightness temperature. A recent study shows that using two frequencies allows melt detection but also distinguishes between different stages in the melting/refreezing cycle (Reference Ashcraft and LongAshcraft and Long, 2005). A drawback, however, concerns the brightness temperature sensitivity to precipitation at high frequency (37 GHz) which causes bias in the melting series (Reference Fettweis, Gallée, Lefebre and van YperseleFettweis and others, 2005, Reference Fettweis, Gallée, Lefebre and van Ypersele2006).

2.2. Melting indices

Our algorithm classifies each pixel on the continent, every day since 1979, as melted or not melted. It produces a huge amount of information that needs to be reduced before analysis. Various ‘melting indices’ have been proposed in different studies including:

1. 1. for individual pixels: melt-onset and melt-end dates and melt-season duration (number of days with melting per year or month) (Reference Zwally and FieglesZwally and Fiegles, 1994; Reference Liu, Wang and JezekLiu and others, 2005). The results are shown as maps

2. 2. for entire regions: CMS measures the number of days× km² melted in a region for a year or a month; mean melting day (MMD) (Reference Zwally and FieglesZwally and Fiegles, 1994) measures the melt duration average for a region; and the maximum melting surface (MMS) measures the surface of all the pixels melting at least 1 day (Reference Torinesi, Fily and GenthonTorinesi and others, 2003). These indices are usually shown as annual series.

Regions of Antarctica were defined by Reference Zwally and FieglesZwally and Fiegles (1994) and have been reused almost ‘as is’ in Reference Torinesi, Fily and GenthonTorinesi and others (2003), Reference Liu, Wang and JezekLiu and others (2006) and Reference Picard and FilyPicard and Fily (2006).

The annual CMS integrates both spatial and temporal information. It has been analyzed in recent studies (Reference Liu, Wang and JezekLiu and others, 2006; Reference Picard and FilyPicard and Fily, 2006). In the present paper, we focus on the spatial variability and analyze melt-season duration on a pixel basis.

2.3. Satellite change issues

In Reference Picard and FilyPicard and Fily (2006) we showed that the change of observation hours between the different satellites used for composing the 27 year long series has a significant impact on the CMS index time series. We proposed a correction method that specifically applies to the CMS index.

Change of observation hours also affects the melt duration, but no correction method exists yet. The method developed for the CMS cannot be easily applied for melt duration, as it requires statistics of melt duration that are not accurately estimable on a pixel basis. In the present paper, no correction is applied but our analysis takes this effect into account by noting that the hours of observation have been less and less favorable to melt detection from the SMMR (1979–88) up to the current SSM/I F13. As a consequence, a pixel experiencing no climatic change would have a slightly negative trend in melt duration, due to the satellite replacements (Reference Picard and FilyPicard and Fily, 2006). As a corollary, observed melt duration trends are negatively biased.

2.4. Trends, correlation analysis and statistical issues

The melting signal provides non-continuous information about climate as opposed to temperature for example. There are two reasons:

1. 1. surface melting occurs when the temperature reaches the limit of 0˚C. Below this temperature, no quantitative information is carried in the melting signal

2. 2. at 0˚C, the liquid-water content is a continuous variable, as the temperature is. Although liquid-water content is a variable of interest for climatology, it cannot currently be retrieved by microwave remote sensing.

Trend analysis (i.e. the least-squared optimization, interpreting the correlation value or Student test result) needs to be carefully conducted.

Where melting occurs frequently, the number of melting events is high and the melt duration behaves as a continuous variable. The distribution of melt duration is then approximately Gaussian (see Fig. 1a) and the usual trend analysis can be used. In contrast, where melting is rare, the distribution looks like an exponential (Fig. 1b), with many years experiencing no melting. In this case a trend analysis is less robust.

Fig. 1.

Distribution of melt duration for two pixels: (a) a pixel with frequent melting in the peninsula; and (b) a pixel with rare melting in Dronning Maud Land.

To deal with this issue, we retain the pixels where melting occurs at least 1 day over 20 or more years (for the 27 years of the series) and reject the other pixels. The selected area covers 1.0×10⁶ km² (˚% of the continent). By comparison, regions melting at least 1 day every year cover only 0.3× 10⁶ km² (2%), although regions melting 5 or more years cover about 3.3×10⁶ km² (25%).

3.1. Long-term trends on the peninsula

Figure 2 shows a map of melt-duration trends for the full time series 1980–2006. Trends are expressed in per cent of annual variation and calculated by the following procedure: (1) a least-squares linear regression is performed on the melt-season duration time series md(t):

Fig. 2.

Trends of melt duration on the peninsula over the summer 1979/80 to summer 2005/06 period. Unit is per cent of change per year.

where t is the time in years; (2) the coefficient of linearity a is normalized by the mean melt duration:

The normalization is required to deal with the large dynamic range in melt duration (0–120 days a^–1).

Melting trends on the peninsula are contrasted, with mountains and ice shelves behaving oppositely. The following features are observed:

By only studying time series and trends of CMS, previous studies (Reference Torinesi, Fily and GenthonTorinesi and others, 2003; Reference Liu, Wang and JezekLiu and others, 2006; Reference Picard and FilyPicard and Fily, 2006) gave an incomplete picture of melting evolution on the peninsula: CMS is nearly steady over the two decades. This observation apparently contradicts some facts: the warming observed at meteorological stations, and the ice-shelf disintegration and glacier surges (Reference Cook, Fox, Vaughan and FerrignoCook and others, 2005). The new depiction provided by Figure 2 is clearer. The CMS is steady because of the compensation between opposite melt-duration trends in different areas.

The ice shelves are experiencing more and more melting, correlating with the warming observed by the meteorological stations, all located near the coasts or on islands (Reference ComisoComiso, 2000; Reference TurnerTurner and others, 2005). Summer mean temperature (November–March) increased over the period 1980–2006 for most stations, for instance from north to south, by +0.14˚Cdecade^–1 at O’Higgins, +0.18˚Cdecade^–1 at Faraday (65.4˚ S, 64.4˚ W) and +0.27˚C decade^–1 at Rothera. This also correlates with the disintegration of the shelves observed in the area (Reference Scambos, Hulbe, Fahnestock and BohlanderScambos and others, 2000) and the positive trends observed by thermal infrared remote sensing on the open ocean (Reference ComisoComiso, 2000).

The shortening of the melt season on the mountainous western coast is a more surprising observation, as the closest meteorological stations record a warming (e.g. Faraday at 65.4˚ S). Altitude and topography may be the key factors explaining this difference in trends. However, the upper-air temperatures measured (Reference Antarctic Data for Environmental Research (READER) project (Reference TurnerTurner and others, 2005)) at the nearest station (Bellinghausen, located at the northernmost end of the peninsula) do not support this hypothesis, as they show strong warming in the summer at both 850 hPa (+0.5˚Cdecade^–1) and 700 hPa (+0.9˚Cde-cade^–1). The only data correlating with the shortening are the pronounced cooling trend observed by thermal infrared (–0.25˚Ca^–1) in Graham Land (Reference ComisoComiso, 2000, fig. 17d).

In addition, surface melting changes are not necessarily explained by changes in temperature. Surface melting depends on all the terms of the surface energy balance. For instance, an increase of the mean albedo (perhaps due to more frequent snowfalls) reduces the shortwave radiation absorbed in the snow and thus reduces the energy available for melting. Also, melting can occur when air temperature is below 0˚C if the incoming short waves are intense enough. In such conditions, an increase in wind speed increases the sensible flux toward the atmosphere which decreases the energy available for melting in the snow. This latter process may have compensated the warming observed at Faraday as the summer-mean wind speed at Faraday increased by as much as 25% over the period 1979–2004 (data from the READER project (Reference TurnerTurner and others, 2005)). More model studies are required to further interpret the complex relationship between trends in temperature and melt duration (e.g. Reference Liston and WintherListon and Winther, 2005).

3.2. Trends between summer 1995/96 and summer 2005/06 in Antarctica

The trend analysis presented here is limited to the last 11 summers (1996–2006). This period is chosen to avoid the artifacts caused by the satellite replacements. The SSM/I F13 has operated since 1995 and has the longest lifetime in the SSM/I family. Its observation hours have been fairly constant so that the trend derived from the limited series is more accurate than for any other periods.

Over such a short period, the trends are not representative of long-term changes, as the Antarctic climate is known to be influenced by several natural oscillations that operate on multi-decadal timescales, mainly the Southern Annular Mode (SAM) and the El Niño–Southern Oscillation (ENSO) (Reference Genthon and CosmeGenthon and Cosme, 2003). Therefore, the trend analysis presented here only represents a snapshot of the current variations.

Figure 3a shows the melt-duration trends over the continent. The main feature is a clear east/west dipolar pattern: the melt season is shortening in the west (Antarctic Peninsula, Filchner Ice Shelf and eastern Dronning Maud Land) and lengthening in the east (East Antarctica and the Ross Ice Shelf). Both zones are homogeneous, with only a few spots having opposed trends.

Fig. 3.

(a) Trends in melt duration in Antarctica over the period 1995/96–2005/06. Unit is per cent of change per year. (b) Trends in measured air temperature (where >90% confidence only). Unit is ˚Ca^–1.

Using a regional climate model, Reference Van den Broeke and van LipzigVan den Broeke and Lipzig (2004, fig. 5a) showed that summer near-surface temperature trends had a dipolar pattern when the SAM index was increasing (1980–93), with the peninsula warming and the rest of the continent cooling. For the period 1995–2005, the trend is reversed: the SAM was decreasing, with the probable consequence of a reversed dipolar pattern in the temperature trends, i.e. cooling on the peninsula and warming over the rest of the continent. This pattern closely mirrors our observed melt-duration trend and is confirmed in the air-temperature trend measured on the ground (READER project (Reference TurnerTurner and others, 2005)) (Fig. 3b).

Over most of the peninsula, the melt season has been shortening over the last 11 years, although it was lengthening on the shelves over the last 27 years. This is in apparent contradiction to the disintegration of the Larsen B in 2002 (Larsen A disintegrated in 1995, at the beginning of the analyzed period) and the southward migration of glacier acceleration (Reference Cook, Fox, Vaughan and FerrignoCook and others, 2005). The contradiction may be because the contribution of surface melting to ice-shelf break-up or glacier surges is not obvious and probably acts on multiple timescales. Recent warming observed at stations over the past 50 years on the peninsula (Reference TurnerTurner and others, 2005) and visible in the 27 year long trends of melt duration (section 3.1) may be the primary cause for the disintegration and surges. But at shorter timescales, surface melting is expected to be involved in the break-up as hypothesized by Reference Scambos, Hulbe, Fahnestock and BohlanderScambos and others (2000) and Reference Van den BroekeVan den Broeke (2005). This mechanism involves intense surface melting, creation of melt ponds and the filling of crevasses which propagate within the shelf. This results in a weakening of the shelf that may lead to the break-up. Although our data agree on a long-term increase in melting, the melt duration in 2002 is not exceptionally long and the trend over the 11 years is not increasing.

A possible explanation is given by Reference Van den BroekeVan den Broeke (2005) who studied the energy budget in 2002 on the Larsen B ice shelves and noticed that the melt amount was much higher than in previous years (1995–2001), even if the melt duration was not significantly different. The timing of the melting days is important, as the solar energy varies during the summer. Another explanation lies in the chaotic dynamic of pond formation. This chaotic nature comes from the fact that free water (i.e. ponds) has a much lower albedo than a wet snow surface. The formation of ponds is thus controlled by a positive feedback: more melting → lower albedo → more absorbed energy → more melting. A slight change in the initial or meteorological conditions may result in one of two cases: a low-albedo surface with many growing ponds or a high-albedo surface with only wet snow. In both cases, microwave radiometers provide the same signal and may have missed the important pond formation in 2002 on the Larsen B. The retrieval of a quantitative value (such as the melt flux or the liquid-water content) from passive microwave is unlikely.