1 Introduction

Arctic sea ice area has undergone a significant decline in recent years. The change and variability of time series of Arctic sea ice concentration has been studied in numerous scientific papers. Many studies show that the decline of Arctic sea ice has accelerated since the inception of the satellite era (1979 to present). Parkinson and Cavalieri (2008), Comiso et al. (2008), and Perovich and Richter-Menge (2009) all report a declining trend for September sea ice extent. Ice phenology defines the characteristics of the ice growth and decay cycle of a year. Ice phenology change deserves special attention because it is a direct indicator of climate change. Notably, a warming Arctic has led to ice phenological timing shifts toward later start of freeze-up and the earlier start of melting and resulted in the lengthening of the melt season (Markus et al. 2009). This contributes to the severe sea ice loss (Serreze et al. 2007). It has been found that Alaskan polar bear denning has been shifted landward and eastward (Fischbach et al. 2007), while the sea bird hatch dates also advanced in the high Arctic (Moe et al. 2009) due to sea ice phenological change, including the reductions in stable old ice, the increase in unconsolidated ice, and the lengthening of the melt season. Using a coupled ice-ocean ecosystem model, Ji et al. (2013) showed that change in sea ice phenology at a specific location lead the changes in the timing of ice algae and pelagic phytoplankton peaks.

When measuring the interannual trend in sea ice change and variability, the ice extent for specific dates or the monthly and seasonal means are sampled from high-resolution time series (e.g., daily sea ice maps). The selection of a few isolated dates for change detection is appropriate only when it can be assumed that there is no significant seasonal variation in the spatial pattern of the icescape. Moreover, sea ice extent trends are often calculated on a regional scale and hence the spatial patterns at pixel level are generally lost. However, the geographic location of the areas with phenological timing shifts are changing from year to year, with some years being more severe than other years. For example, the Western Arctic, including the Beaufort Sea, Chukchi Sea, Eastern Siberian Sea, Laptev Sea, and Kara Sea suffered the most significant summer sea ice loss compared to that of the other regions of the Arctic in summer 2007 (Serreze et al. 2007). Thus, the identification of the geographic locations where the yearly sea ice phenological changes most often occur is important for improved understanding of the processes of sea ice change.

Change detection from multitemporal remote sensing images has long been studied for a variety of applications. Lu et al. (2004) illustrated a comprehensive review of the change detection techniques using remotely sensed data. Im and Jensen (2005) noted that remote sensing change detection generally has three goals, namely detect the location of the change, identify the types of change, and quantify of the amount of change. For sea ice distribution patterns, principal component analysis is often used to explain most of the variance in the data for the analysis period (Piwowar and LeDrew 1996), but the derived principle components are aggregated patterns, and although statistically meaningful, there is no guarantee that the patterns are physically meaningful unless there is extensive visual validation. In response to this issue, temporal mixture analysis was proposed for sea ice change detection (Piwowar et al. 1998). Three to four end members are defined as the purest pixels, and each pixel is assigned a percentage of purity. Ice change can be captured by each end member’s percentage of change in a pixel. However, the identification and interpretation of the end members are not always straightforward. Small (2012) proposed a new method to characterize multitemporal imagery by the combined use of principal component analysis and spectral mixture analysis. These methods, however, do not have sufficient incorporation of high temporal resolution data for continuous monitoring of change, such as temporal trajectory analysis which has been commonly used to monitor vegetation change based on land surface phenology at regional, continental, and global scales (Coppin et al. 2004; deBeurs and Henebry 2006; Hargrove et al. 2009). Temporal mixture analysis results in a fusion of temporal information at different times, but phenology-based clustering regions of the temporal trajectory analysis provide direct and easy to interpret maps of the subject studied. When the clustering regions of different years are compared, the location where changes have occurred is directly available.

Arctic sea ice edge is not a symmetric ring around the pole in either summer or winter. Melling (2012) proposed that the ice edge is the location where the rate of ice loss through melting equals that of delivery via drift. Ocean bathymetry influences the distribution and mixing of warm and cold waters and plays a profound role in controlling the location of sea ice edges by influencing the Arctic sea ice formation and seasonal evolution (Nghiem et al. 2012). The wintertime ice edge depends primarily on coastline geometry, ice motion, and the melt rate at the ice-ocean interface, driven by absorbed solar radiation and the convergence of heat transported by ocean currents (Bitz et al. 2005). The controls of the summertime sea ice edge position have been poorly understood.

The spatial and temporal changes are not coherent when different regions and different months of year are examined. Temporal variations occur for periods ranging from days, weeks, and months to periods of years, decades, and longer. Spatial variations occur at regional to synoptic and hemispheric scales. Sea ice phenology provides information about the sea ice annual cycle, such as the maximum/minimum sea ice extent (SIE) of a region, the timing of the maximum/minimum, the timing of melt/freeze up onset, and the length of melt season. Thus, the daily passive microwave sea ice concentration (SIC) data accumulated for the past three decades provide a hypertemporal dataset to identify where sea ice phenological change has occurred, what is the nature of the phenological change, as well as the degree of change. In this study, sea ice phenological information is used to investigate the sea ice spatial and temporal changes in the Arctic in two steps. First, we generate a phenoregion map for each year in the study period based on decision tree classification of the temporal SIC profile. A phenoregion map integrates the sea ice melt information of a year for an interval that we have defined as the active period from April to September. Secondly, the spatial and temporal characteristics of the phenoregion maps are analyzed. In the following section, the phenoregion maps of different years are compared, and the location and degree of change of the phenological classification map are identified. Thus, the spatial and temporal details about the timing and the locations where the changes occur are answered.

2 Sea ice data and methods

2.1 Study area and data

Daily sea ice concentrations for 1989 to 2010, provided by the National Snow and Ice Data Centre (NSIDC), University of Colorado, Boulder, CO, are used in this study. Since we are only interested in the sea ice melt process during spring and summer, daily data from April to September are selected. The study area is the ocean area inside the Arctic Circle as shown in Fig. 1, north of 66.56° N. The National Aeronautics and Space Administration (NASA) Team algorithm SIC dataset is the most widely used one for calculating the sea ice extent and sea ice area, because of its consistency through the years and between different generations of satellites (Parkinson and Cavalieri 2008). The presence of melt ponds produce ambiguities in passive microwave signature and cause the underestimation of ice concentration significantly during melt. The total ice-covered area derived using the NASA Team algorithm (Cavalieri et al. 1984) was underestimated by 20.4 to 33.5 % during ice melt in the summer compared with regional charts from the Canadian Ice Service (Agnew and Howell 2003). Since the delineation of locations where ice phenology changes occur through the years is the main research goal, the analysis of our results is not hampered because the ice concentrations are consistent through the years. Thus, the NASA Team dataset is selected in this work due to its consistency.

Fig. 1
figure 1

a Study area and b two examples of sea ice concentration profiles for two Arctic locations in three different years 1998, 2004, and 2007

Full size image
Fig. 2
figure 2

Flowchart of the generalization of yearly phonological map for change detection

Full size image

2.2 Annual daily SIC profiles

The annual SIC profile of a pixel is defined here as the time series of its daily SIC from April 1 to September 30 of a year, which shows the pixel’s sea ice decay process during that year. The SIC profile contains all information about the pixel’s SIC evolution. Phenological metrics, such as melt/freeze onset and the length of the melting/melting season, can be used to characterize the sea ice profile (Table 1). Changes in any of these factors, such as an earlier melt onset or a longer melting season, lead to changes in the SIC profile. The SIC profile captures detailed information about the time trajectory of the sea ice annual evolving patterns compared to methods that only use one value at a particular time period to represent a year.

Table 1 Representative attributes of sea ice melt phenology
Full size table

To illustrate, the SIC profile of a pixel in the Western Arctic basin and another pixel in Canadian Arctic Archipelago (CAA) (both shown in red on the map in Fig. 1) for three different years, 1998, 2004, and 2007, are shown in different colors in the right panel in Fig. 1. These years are of interest for they represent extreme ice extent years for the CAA: 2004 was a high ice year and 1998 is a very low ice year. 2007 is represented because it is a low year for the whole Arctic. For the selected pixel in the Arctic basin, the SIC profiles of the three years shown are quite different. Approximately before year day 181, the SIC are similar, but after that, the SIC profile differs from each other greatly, mainly due to the different rates of melt which resulted in different end-of-summer SIC values. In 2004, SIC remained above 0.5 for all summer, but SIC dropped to 0.4 in August 2007 and the pixel became ice-free in August 1998. As for the selected pixel in CAA, year 2007 became ice-free in YD 222 but remained partly covered by ice during the entire melt season in 1998 and 2004. Moreover, these three years had different melt onset dates: 1998 is the earliest and 2004 is the latest.

To create the phenology database, the daily SIC data for a year is used to generate one sea ice melt map for that year, and the sea ice melt map of different years will be compared for sea ice melt change detection (Fig. 2).

2.3 Design of the classification scheme

The annual SIC profile of a pixel provides significant information about sea ice annual evolution. Two phenological metrics are selected here to define a pixel’s yearly sea ice retreat pattern, the minimum SIC for the year, and the day of year when that minimum SIC occurs. All annual SIC profiles for all pixels from the 22-year study period are examined based on the two selected phenological metrics. The histogram for the general distribution of the minimum SIC is shown in Fig. 3a. Here, ice-free pixels are defined as having a SIC less than 0.15. The minimum SIC exhibits a bipolar distribution with 86 % of the pixels having some ice cover all year round and 63 % of the pixels having a minimum SIC of less than 0.15, which is considered ice-free. However, 14 % of the study area is ice-free all year round. Thus, 49 % of the study area has seasonal ice cover, some ice cover at the beginning of the melt, and gets ice-free in the summer. For seasonal ice-covered pixels, the day of year when ice-free occur is explored. The distribution of the day of year when the ice-free conditions occur is shown in Fig 3b.

Fig. 3
figure 3

a The distribution of minimum SIC for all pixels from all years and b the distribution of the ordinal day when ice-free occur for all pixels that is not ice-free at the beginning of melt (April 1)

Full size image

We designed a classification scheme that uses the two phenological metrics and subsequently classifies the yearly SIC profiles into several classes (Fig. 2). The number of classes is set to seven to ensure that each class has some physical meaning regarding ice phenology. One class is reserved for pixels that are ice-free all year round. Three classes are used to define areas with seasonal ice cover, and three classes are used for areas having ice cover all year round. Another consideration for the selection of seven classes and the corresponding thresholds is to ensure that the number of pixels in each class is evenly distributed based on all pixels in 22 years. Thus, a decision tree classifier with seven classes is defined and shown in Table 2. Ice in the areas covered by classes 1 to 3 survives the summer melt and becomes multiyear ice (MYI) in the following year. However, areas in these three different classes have different summer minimum sea ice concentrations. Classes 4 to 6 cover areas that become ice-free in the middle of summer. These three classes are defined based on the timing of the ice-free condition. Class 7 contains areas with no ice cover all year round. The percentage of pixels in each category is evenly distributed except for class 1, which is reflective of the large area of the central Arctic.

Table 2 Decision tree classifier
Full size table

2.4 Generalization of the yearly phenoregion map

For each year, a classification map is generated using the yearly SIC profiles of all pixels from that given year. The classification procedure is unsupervised as long as the classification scheme has been designed. Firstly, for the two proposed phenological metrics, the minimum SIC for the year and the day of year when that minimum SIC occurs are extracted from each yearly SIC profiles. Secondly, all yearly SIC profiles are classified into seven classes using these two phenological metrics based on the classification scheme. Thus, each pixel gets a class number, ranging from 1 to 7. Thirdly, a classification map is generated for each year using the class number of all pixels from that year. Classification maps are named as phenoregion maps. The area covered by each class is defined as a phenoregion. Each pixel belongs to one of the seven phenoregions. A phenoregion map generates an instant history of sea ice melt phenology of a particular year. Since we have data for 22 years, 22 phenoregion maps are formed using the same criteria. It is the phenoregion maps of different that are useful for comparison between different years.

2.5 Sea ice phenological change detection

The yearly phenoregion maps provide a basis for change detection because all maps are generated using the same criteria. A pixel belongs to one phenoregion in one year may belong to another phenoregion in another year due to ice phenology change, such as the earlier melt onset. A 22-year climatological mean phenoregion map is generated by defining the phenoregion number of each pixel as the median phenoregion number of the 22-year period. The phenoregion map of each year is then compared with this climatological mean map. Pixels that have different sea ice phenology profiles compared to their normal conditions are identified.

3 Results

3.1 Spatial-temporal variability of the phenoregion map

Based on the decision tree classifier, a phenoregion map is generated for each year (Fig. 4). Each year, each pixel has a unique phenoregion number. The climatological phenoregion map depicts the average sea ice retreat process for the study period from 1989 to 2010 (Fig. 5a). A large part of the central Arctic Ocean is classified as phenoregion 1 and represents pixels that always have SIC higher than 60 %. The pixels classified as phenoregion 2 to 3 have high SIC at the beginning of the ice retreat, and SIC decrease to a relatively low level at the end of summer. Phenoregions 4 to 6 represent pixels that have some ice cover at the beginning of the melt season and turn completely ice-free in the summer. The timing when these pixels get to ice-free varies greatly, ranging from early June to late September, as shown by the mean SIC profile for each phenoregion in Fig. 5c. From the standard deviation map of the phenoregion number for the 22 years (Fig. 5b), we can see that western Arctic has the biggest variation in phenoregion number, indicating that sea ice retreat in this region follow quite different patterns in different years for the time period selected. Based on our definition of ice phenoregions, ice in phenoregions 1 to 3 survives the summer melt and become MYI in the following year, while ice in phenoregions 4 to 6 experience melt in April to September. Generally speaking, the boundaries of the phenoregions follow a distribution, with areas in lower latitude becoming ice-free earlier than higher latitude regions, even though we do not take into account the spatial location of the pixels in the decision tree classifier. Moreover, the spatial boundaries of each phenoregion on the mean map generally follow the oceanographic features of the Arctic. The detailed descriptions of each phenoregion are as follows:

  1. 1.

    Phenoregion 1 (shown in brown): Highly ice covered throughout the year, which includes the central part of the Arctic Ocean, the Northern CAA, and a small tip along the northeast corner of Greenland. The spatial distribution of phenoregion 1 is highly influenced by the clockwise Beaufort Gyre and Transpolar Drift, which move sea ice to the Canadian side of the Arctic Ocean and through the Fram Strait to the Atlantic Ocean (Brandon et al. 2010).

  2. 2.

    Phenoregions 2 and 3 (shown in pink and orange): This region has high ice cover at the beginning of April and SIC drops to relatively low values at the end of summer melt. These areas are transitional regions between fully ice-covered regions and ice-free regions. The transition region between high ice areas and ice-free areas is quite narrow. Once the ice is broken into pieces, it melts away fairly quick and cannot remain as partially ice covered due to the ice albedo feedback which accelerates ice melt.

  3. 3.

    Phenoregion 4 (shown in yellow): This region has high ice cover at the beginning of April and becomes ice-free typically after August 1. In the Atlantic side of Arctic, the boundaries of this phenoregion generally resemble the farthest northern extent of the warm Atlantic Water in the central Arctic Ocean.

  4. 4.

    Phenoregion 5 (shown in green): This region has high ice cover at the beginning of April and ice melts away in July.

  5. 5.

    Phenoregion 6 (shown in blue): This region has low sea ice cover at the beginning of April and ice melts away in May and June. This includes peripheral seas and several areas where major polynyas in the northern hemisphere are located, such as the North Water Polynya, Cape Bathurst polynya, and Northwest Water polynya (Smith and Barber 2007). The polynyas are the areas that maintain very low SIC to ice-free conditions all year round.

  6. 6.

    Phenoregion 7 (shown in cyan): This region is ice-free before April 1 and generally has no ice cover all year round.

Fig. 4
figure 4

Phenoregion map for each year from 1989 to 2010, the last one is the climatological classification map based on the mean daily SIC map. Phenoregion number 1 to 7 is color-coded as brown, pink, orange, yellow, green, blue, and cyan, respectively

Full size image
Fig. 5
figure 5

a Mean phenoregion map for each class based on the mean SIC profiles, b the standard deviation map of the phenoregion number. The darker the blue color, the larger the standard deviation of the phenoregion number. c Mean SIC profiles. The color coding is the same as in Fig. 3

Full size image

3.2 Temporal variability of the phenoregion maps

The phenoregion maps provide information about the SIC melt pattern, as shown in Fig. 4. Phenoregion 1 was a large region in the 1980s. However, phenoregion 1 evolved into two sectors in 1996, with some areas close to the North Pole that were then classified as phenoregions 2 and 3. In 1998, the areas classified as phenoregions 1 to 3 were highly mixed with each other, indicating that the ice was broken into pieces, probably due to ice motion. In 2010, phenoregion 1 almost disappeared from the eastern part of the Arctic Ocean. Due to the high index of Arctic oscillation (AO) from 1989 to 1994, anomaly wind caused increasing drift of old ice out through Fram Strait (Rigor and Wallace 2004), and the area of phenoregion 1 decreased precipitously. However, AO has been neutral since 1995, and it turned strongly negative during winter 2009/2010 (Stroeve et al. 2011).

Year 1998 marks the lowest sea ice extent for CAA since 1979 (Howell et al. 2008). Large areas in the southern Beaufort Sea were classified as phenoregion 6 shown in blue (Fig. 3), indicating that this area became ice-free as early as April. However, this blue area almost disappeared from the Beaufort Sea region in 1999 and 2000, implying that the ice loss was not as large as that in 1998. In the CAA, sea ice cover has recovered from 1998 to 2004 after the record low in 1998, indicated by the growing area of phenoregion 2 (shown in pink). From north to south and from west to east, more and more ice have survived the summer melt with eventual higher SIC at the end of summer than in other years.

The percentage of the total area covered by each phenoregion varies greatly from year to year (Fig. 6). The sum of the area covered by phenoregions 1 to 3 for the study area decreased significantly since 2007 (Fig. 6a), from nearly 50 % of the area inside the Arctic Circle to about 30 % for 2007 to 2010. The numbers of pixels classified as each phenoregion for the CAA region and the Arctic Ocean are also shown. The region is delineated following Parkinson and Cavalieri (2008). The black line on each panel shows the variation of the sum of percentages for phenoregions 1 to 3. The correlation coefficient between the sum of pixels labeled as 1 to 3 and the minimum sea ice extent for September all exceed 0.9 for the three regions considered, which is reasonable because phenoregions 1 to 3 are the areas where ice survives the summer melt and generally represents the September monthly mean. Sea ice in the central Arctic has been observed to be increasingly dominated by thin first-year ice (Stroeve et al. 2011). This can be supported from our result. The percentages for phenoregions 4, 5, and 6 are increasing greatly for the 2007 to 2010 period for the Arctic Ocean (Fig. 6b).

Fig. 6
figure 6

Time series of the percent area of each phenoregion for 1989 to 2010 for a all ocean pixels inside the Arctic Circle, b the CAA, c Arctic Ocean, d the Kara and Barents Sea, and e Greenland Sea. The last stacks in each time series are the percentage calculated based on the climatological mean map. The black lines on each time series are the yearly sum of percentage of class 1 to 3

Full size image

3.3 Phenoregion anomaly maps and phenoregion changes in recent years

The phenoregion maps of different years can be used to understand the patterns of interannual change. Due to the changes in the SIC melt profiles, an individual pixel belonging to one phenoregion in one year may belong to another phenoregion in another year. By comparing the class labels of the same pixel from one year to another, we know whether a pixel changed its ice melt pattern. Thus, the geographic location and the degree of change can be mapped. Based on our definition of phenoregions, a pixel’s increasing classification label from one year to another year implies earlier melt, while a decreasing label indicates later melt. Here, all years are compared with the climatological mean, and the anomaly maps are generated and shown in Fig. 7 for selected years. Considering that the differences between two neighboring classes are quite small, pixels with a change of label of 0, −1, and 1 are shown in white, which means no change. Pixels with a change of class label of 2 or more are shown in red, indicating a sea ice trend toward earlier melt; while pixels with a change of class label smaller than −1 are shown in blue, indicating sea ice trend toward later melt. The darkness of the red/blue defines the magnitude of the change. Thus, areas shown in color are pixels with significant changes of ice melt pattern. From the anomaly maps, we find that the most striking declines occur in marginal ice zones and coastal regions. Moreover, ice in western Arctic melts much earlier than before, replacing perennial ice with annual sea ice types.

Fig. 7
figure 7

Anomaly maps for different years

Full size image

3.4 Discussion

Since the criteria used to define the phenoregions are subjective, it is impossible to evaluate the classification results using objective measures. Our classification results are evaluated by close visual inspection of the classification maps. First, phenoregions are homogeneous, and the boundaries of the phenoregions are smooth and mainly follow the longitudinal zonation. Second, several locations stand out as earlier or later melt compared to other regions with the same latitude. This is influenced by local oceanographic conditions. For example, phenoregions in Baffin Bay area are developed around the north water polynya and are influenced by the ocean currents. For the northern CAA, land blocking helps preserve the old ice. However, the warm Baffin Current makes ice in eastern Lancaster Sound melt much earlier than western Lancaster Sound. Third, adjacent regions on each side of Novaya Zemlya belong to totally different phenoregions because warm Atlantic Water can only arrive at the Barents Sea but not the Kara Sea. All of this evidence illustrates that the proposed clustering criteria are reasonable, and the classification results are logical.

Though each phenoregion is well defined, the thresholds used to separate the phenoregions in the decision tree classifier are subjectively selected. Thus, the boundaries of the phenoregions are not precise. Despite the limitations described above, this pilot study presents several previously undocumented findings. By proposing a decision tree classification that divides sea ice pixels into different phenoregions based on sea ice melt profile, we have established a framework to integrate high temporal resolution sea ice data for change detection at spatiotemporal information of the original data. Our results are also comparable to the ice age distribution of Maslanik et al. (2011). However, the ice age algorithm is developed for the Arctic Ocean (Fowler et al. 2004).

4 Conclusion

Arctic sea ice cover has been shifting from MYI to seasonal ice in recent years. Direct measurements of ice age or ice thickness, which is typically required for an assessment of multiyear versus seasonal ice, are only available for selected areas in recent years. In this paper, we present an innovative way of classifying the Arctic into seven classes using the ice phenology information inferred from the long time series of daily SIC data. Ice phenology in the Arctic region during melt season has different characteristics, including different timing of start of melt dates and different durations of the melt. Based on these ice phenology profiles, a phenoregion map is developed by decision tree classification for each year in the study period. The yearly phenoregion maps exhibit significant variations from year to year, which set the foundation for change detection. By comparing the yearly phenoregion maps with the climatological mean map, yearly anomaly maps are formed. The anomaly maps clearly show the areas where ice melts vary most dramatically. Compared to most of the studies that utilize one value to represent the entire sea ice region, the current analysis at the pixel scale provides detailed spatial information of Arctic sea ice change. This analysis also provides a new way of visualizing the daily sea ice data. By integrating key sea ice retreat information of a year into a single map, these yearly phenoregion maps present the spatial and temporal distribution of sea ice retreat patterns. The understanding of sea ice retreat is much more comprehensive when compared to analysis only using a monthly or seasonal means of SIC.