Leads are elongated fractures in the sea ice cover. They form under stresses on the sea ice forced by wind and ocean currents [1
]. The open water refreezes in the cold environment, so leads may contain unfrozen water or ice of varying thicknesses. Leads may be a few meters or a few kilometers in width and may be tens of kilometers in length. While leads occupy a relatively small area of the pack ice overall (e.g., less than 5% of the pack ice surface area), the open waters provide a significant source of heat and moisture to the atmosphere, particularly during winter [2
]. In the summer, leads absorb more solar energy than the surrounding ice, warming the water and accelerating melt. In the winter, spring, and autumn, leads impact the local boundary layer structure [4
] and cloud properties because of the large heat and moisture fluxes into the atmosphere. Leads affect atmosphere and ocean chemical exchanges, such as carbon dioxide, mercury and bromine (e.g., [5
]). The spatial and temporal descriptions of sea ice lead characteristics can provide useful information in studying the chemical properties of the Artic region. From an operational perspective, knowledge of lead characteristics can aid in navigation, with direct benefits to security, subsistence hunting, and recreation. Given the rapid thinning and loss of Arctic sea ice over the last few decades [7
], changes in leads can be expected. Lead formation is driven by stress imposed on the sea ice by wind and ocean currents, so they respond to changes in the wind and the resulting ice deformation [10
]. From a climate perspective, identifying changes in lead characteristics (width, orientation, area coverage and spatial distribution) will advance our understanding of both thermodynamic and mechanical processes in the Arctic.
A number of studies used satellite data to detect sea ice leads, dating back at least to the early 1990s. Key et al. [11
] developed a method to detect and characterize leads using thermal infrared and visible satellite imagery, primarily from Landsat, and explored the sensitivity of lead detection using the Advanced Very High Resolution Radiometer (AVHRR) thermal imagery under various atmospheric conditions. A follow-on study examined the effect of sensor pixel resolution on lead detection [12
]. Lindsay and Rothrock [13
] produced binary lead maps for specific Arctic regions from the AVHRR data in 1989. Miles and Barry [14
] used AVHRR data to study leads in the western Arctic for the winters from 1979 to 1985. Drüe and Heinemann [15
] applied the potential open water concept of Lindsay and Rothrock [13
] to Moderate Resolution Imaging Spectroradiometer (MODIS) data to derive sea ice concentration based on infrared window brightness temperatures. Willmes and Heinemann [16
] presented a technique for lead detection using the thermal channels of MODIS.
Satellite microwave observations, both passive and active, have also been employed in leads detection, the advantage being that most clouds are transparent in the microwave spectral range. Röhrs and Kaleschke [18
] and Röhrs et al. [19
] used the low-resolution Advanced Microwave Scanning Radiometer-Earth Observation System (AMSR-E) where the 18.7 and 89 GHz brightness temperatures were mapped to a 6.25 km grid and an emissivity ratio method was used to detect thin ice. A spatial high-pass filter was applied to retain linear thin-ice areas. It was determined that subpixel-resolution leads could be identified. This work was extended by Bröhan and Kaleschke [20
] to develop a nine-year “climatology” of lead orientation and frequency. Synthetic Aperture Radar (SAR) provides the best spatial resolution of microwave sensors but is limited in coverage, both spatial and temporal. Nevertheless, SAR observations have been used to characterize leads and to validate lead information derived from other sensors [21
]. Murashkin et al. [22
] found that using both horizontally-polarized channels of Sentinel-1 SAR identified more leads than single co-polarized observations. Their lead classification algorithm uses a random forest classifier based on polarimetric features and textural features.
Satellite altimetry can also be used for the detection of leads. For example, Zakharova et al. [23
] used the SARAL/AltiKa (satellite with Argos and ALitKa) altimeter to detect leads. They defined a threshold of the maximal power of waveform to discriminate the leads from 200 m to 2-4 km in width. Lead area fraction and widths have been examined using the Cryosat2 radar altimeter [24
], where a supervised classification of CryoSat-2 data was performed by a comparison with visual scenes. As in [23
], the maximum power of the waveform showed the best classification properties.
This paper presents a methodology for detecting and estimating characteristics of sea ice leads using infrared satellite data. It describes the algorithm, demonstrates applications, and provides examples of the analysis. Our approach improves upon the Key et al. [11
] algorithm, and provides lead characteristics such as width, orientation, and area that are not produced by more recent thermal infrared algorithms. While the application of the methodology described here is demonstrated using MODIS on NASA’s Terra and Aqua satellites, it is equally applicable to other satellite imagers.
The design philosophy of our algorithm is to minimize the errors of commission; i.e., to minimize overestimation of leads. As presented in Table 2
, the area classified as potential leads is in better agreement with the Willmes and Heinemann [37
] leads area than our positively identified leads. Quantitative results are presented annually; qualitative comparisons of an example day are shown in Figure 11
and Figure 12
. The quantitative results from the day of the case study are similar to the annual results (not shown). Although both products are derived from the same satellite, there are some differences in the domain. Willmes and Heinemann [16
] does not extend as far south however, our algorithm limits the view angle (creating a coverage gap at the pole). Cloud coverage results in some difference as well. We apply a cloud screening technique that allows for some lead retrievals in areas where the ice surface temperature is not retrieved due to cloud, and therefore Willmes and Heinemann [16
] cannot process a location for leads. Some of these differences are visible in the case study shown in Figure 11
Furthermore, our positively identified lead area is significantly smaller. When comparing statistics where both have cloud-free coverage, our technique finds leads in approximately 3% of the domain compared with approximately 10% of the Willmes and Heinemann domain. Some of the differences can be attributed to resampling of the Willmes and Heinemann [37
] product; the nominal native resolution of that product is 2 km resolution while our product resolution is 1 km. Using the coarser resolution, the contribution of sub-resolution leads may overestimate the total lead area. We are able to achieve the finer resolution by limiting the scan angle to 30°, though the trade-off is a coverage gap over the pole (north of approximately 81°N). The effects of sensor resolution on lead characteristics were examined in [11
]. Using Landsat Multi-Spectral Scanner 80 m data degraded to lower resolutions, Key et al. [12
] found that (a) while the manner in which widths of individual leads changes with increasing pixel size is highly variable, the mean lead width over an image changes in a more predictable way. For example, a mean lead width of 400 m derived from data with a 200 m pixel may “grow” to a mean width of 800–1000 m with a 700 m pixel (Figure 6 in [12
]). (b) The change in total lead area with increasing pixel size is generally exponential. (c) Leads narrower than approximately 250 m disappear as the resolution of Landsat images is degraded to 320 and 640 m. (d) Lead orientations change if they are anisotropic (i.e., have a preferred orientation), but do not change substantially if they are isotropic. The actual effect of pixel size depends in part upon the temperature or reflectance contrast between a lead and the surrounding ice; see [12
] for a quantification of the combined effect.
The difference in lead area is also a result of the different linear identification techniques that our algorithm employs. If there are applications that are less sensitive to commission error and more sensitive to omission error, users can refer back to Table 1
for a summary of lead rejection categories and they may choose to include some rejection categories in addition to our positively identified leads category. A polynya is an example of an ice feature with the thermal signature of leads, but it is irregularly shaped. [41
] Our algorithm is tuned towards identifying leads as linear features within the sea ice pack.
We use Level1B 11 μm brightness temperatures [28
] rather than the MODIS MxD29 ice surface temperature product [42
]. The MODIS noise-equivalent temperature difference is 0.05 K at 11 μm channel, which is sufficiently accurate for lead detection. There is, of course, a strong correlation between ice surface temperature and the 11 μm brightness temperature. Also, the actual surface temperature is less important than the contrast in temperature. Leads become undetectable when the thermal contrast between leads and surrounding ice pack is small (e.g., less than 1.5 K, although the local spatial variability is also a factor, as described earlier). The primary cause of thermal contrast in the Arctic winter would be leads—the contrast between solid ice and either open water, ice and water mixed, or thin ice. In warmer seasons the contrast between ice temperature and water temperature becomes smaller. Detection capabilities decreases as the surface temperature increases. In summary, the temperature contrast becomes small as ice becomes thicker within a lead or when the surface temperature approaches the melting point of ice in the case of an unfrozen lead.
We agree with the findings of Fraser et al. [34
] that cloud masks often misidentify leads as clouds. We are able to employ a similar cloud clearing method to reclassify false cloud detections as clear and detect leads using the 11 μm brightness temperature. It would not be possible to detect leads using the sea ice temperature in these locations because the sea ice temperature product would fail to provide a temperature retrieval in these areas flagged as cloudy. Another area of investigation was to use an ice concentration product as the basis for identifying potential leads, but we found that much like ice surface temperature products, cloud product omission errors (associated with leads) often prevented ice concentration retrievals just as clouds also prevent ice surface temperature retrievals.
Clouds present several obstacles in lead detection. The infrared thermal signature of clouds is often similar to that of leads, which is why cloud detection algorithms often errantly flag leads as clouds. Polar cloud detection in the winter is difficult; cloud detection algorithms are designed to detect cold and highly reflective features, which happen to also be characteristics of sea ice. Further, cloud mask performance is in general better during the day—when reflective bands contribute to better performance of cloud detection algorithms—but the polar region is predominately dark during the winter. Also, clouds in the polar winter are commonly warmer than the surface. Cloud detection algorithms use ancillary data to try to identify in advance where snow or ice may exist at the surface in an attempt to detect clouds over snow or ice. The problem with this technique is that the ancillary data might not accurately reflect the actual conditions; the product may be observation based and reflective of past conditions that may be obsolete or based on a forecast that is not reflective of current conditions. Also, the resolution of the snow/ice mask may be too coarse resolution to account for narrow features like leads.
One of the weaknesses of the algorithm is that it was not designed to work in the summer months. The thermal contrast between leads and the surrounding ice would be very different and the surface can be more complex. Melt ponds, for example, may appear in the summer and these could potentially fool the algorithm. When the ice surface temperature approaches the water temperature, the algorithm would not detect thermal contrast and therefore not detect leads. Persistent cloud coverage in the warmer months is another factor for the limited period of coverage. The January through April time period was chosen because this is generally the best season for lead detection and it is consistent with the season provided by Willmes and Heinemann [37
One of the advantages of the polar region is the frequent coverage from polar satellites. One of the assumptions of the algorithm is that most locations in the domain will have been covered by multiple satellite overpasses. As illustrated in Figure 6
, most locations have over six overpasses per day, though cloud coverage will limit the number of overpasses where lead detection is possible.
In addition to frequent repeating coverage, another assumption the algorithm makes is that leads will be stationary or slow moving in order to satisfy the requirement that a lead must be detected in multiple overpasses. To account for lead movement there is a requirement that 90% of the lead area must be detected in multiple overpasses. A fast-moving lead could fail to meet these conditions; however, the alternative has drawbacks. If a lead is moving quickly such that it appears in different locations throughout the day, a daily composite of lead area could over-report lead area because the fast-moving lead area would be counted towards the daily area in each location it was detected (versus stationary leads that would only contribute towards the daily area once). We believe the smearing effect from slow moving leads that could artificially inflate the lead width and area is more acceptable than considering each detection of a fast-moving lead as a contribution towards the daily lead area. We also believe that repeat detections are important to establishing confidence in a lead detection. The type of phenomena that would produce a thermal contrast signature similar to that of leads would either be a broken ice feature like a lead or else related to a cloud. Cloud edges, cloud shadows, and cloud mask omission errors could all produce a thermal contrast signature similar to a lead. These cloud related features are more likely to be short lived; we would expect viewing geometry and cloud movement to change overpass-to-overpass and make the reoccurrence of false lead signatures less likely than for repeat observations of a real lead. Establishing a confidence parameter would be a non-trivial undertaking. Providing the number of times a potential lead was detected can be used subjectively to help establish user confidence in lead detections. In future work a confidence may be established based on observation frequency as well as other factors such as making an attempt to track a previously detected lead or perhaps quantifying the thermal contrast signature into a confidence parameter.
This paper describes an algorithm to detect sea ice leads (fractures) over the Arctic using measurements from MODIS on the Terra and Aqua satellites. We describe the algorithm and demonstrate results by applying the algorithm over the entire MODIS Aqua/Terra period, 2003–2018, for the winter and early spring months of January–April. The algorithm consists of three main steps. The first step makes a 1-km grid composite of the potential leads based on the thermal contrast in the satellite infrared window data over the Arctic Ocean. The second step defines lead objects using a series of image analysis techniques to identify pixels that may have spatial characteristics of leads. In the third and final step, lead objects are broken into smaller individual branches and the characteristics of each branch of a potentially multi-faceted lead are calculated. This step also determines the area, length, width and orientation of the lead. The results of this algorithm are compared with the products from Willmes and Heinemann [37
] who also use MODIS but employ a different methodology.
Cloud coverage constrains lead detection, optically thick clouds in particular, thus an infrared algorithm for lead detection requires an accurate cloud mask. Unfortunately, the most difficult conditions for automated cloud identification techniques—nighttime darkness, bright surfaces, and cold surfaces—are also the most prevalent conditions in the Arctic winter. Also, the thermal contrast and shape of cloud edges and cloud shadows often have a similar shape and temperature contrast as lead. Passive microwave observations overcome many problems associated with cloudy conditions but have a larger footprint. Combining collocated microwave and imager observations may be the best approach to determining changing lead characteristics in the Arctic.
Instrument resolution is another limiting factor. By limiting the sensor scan angle we are able to keep our observation resolution at a relatively consistent 1 km. We did not attempt to define the minimum detectable lead size; it would be a function of the lead size, water temperature (and sea ice concentration) within the leads, and the temperature of the surrounding sea ice.
Future work will explore the relationship between cloud cover and the lead detection, trends in lead length and width, relationships between lead azimuth angle and anomalies in wind and ocean currents, and relationships between changes in lead properties and sea ice thickness. Application of the algorithm indicates that the annual variation in lead coverage is large, both spatially and temporally. The algorithm will be adapted to the Visible Infrared Imaging Radiometer Suite (VIIRS), which has more consistent spatial resolution of 375 m across the entire 11 μm swath. With a wider swath than MODIS, the two-satellite system of NOAA-20 and the Suomi National Polar-orbiting Partnership satellites will be able to provide even greater coverage than has been available with MODIS.