The UMD-MPA was developed to identify pond surfaces and their bathymetry in the ICESat-2 ATL03 photon height product . First, we used a cloud indicator based on the apparent surface reflectance parameter provided as a flag in ATL07 to identify cloud-free sections of along-track surface height data. If at least 20 % of the track within the study region was cloud-free, we manually examined the ATL03 photon height data for evidence of melt ponds. Figure  demonstrates the methodology to determine the surface and bathymetry of a pond using the UMD-MPA. Figure a shows the ICESat-2 ATL03 photon cloud, where we see photons outlining the two-dimensional iconic bowl shape of a melt pond (between 400 and 700 m along track), with photons returned from both the surface and bottom of the pond. We manually identified the start and end of ponds as the points where two surfaces diverge and rejoin, respectively. We defined pond width as the distance between the start and end points. To determine the surface height hs, we binned all photons across the width of the pond into 0.1 m vertical bins (Fig. b), and hs is the mode of the distribution: 4hs=Pn_ma, where Pn_ma is the bin containing the maximum count in the vertically binned histogram for all photons across the width of the pond; hs was reset to an elevation of 0 m, and all photon heights were recalculated relative to hs (Fig. c). Then we constructed a new two-dimensional histogram of photon height data with vertical elevation binned at 0.1 m using 10 m wide horizontal along-track bins in order to distinguish the surface photons from the bathymetric photons. For each vertical bin, we added the photons from the bins on either side to increase the photon count for each bin. In this way, the vertical bins were overlapping with an effective bin height of 0.3 m at 0.1 m intervals (Fig. d). For each 10 m horizontal along-track segment (as shown in Fig. c), we examined the resulting histogram of vertical elevation (Fig. d). We assumed photons within the two bins on either side of the identified pond surface mode could be associated with the surface and removed all photons in those bins for the subsurface analysis (Fig. d, green bins), and thus the minimum retrievable pond depth was 0.3 m (0.23 m after correction for refraction of light in water). We located modes in the histogram below the surface that contained at least 5 % of the number of surface photons in Pn_ma (Fig. d, blue bin). If there were no modes that met this threshold, pond depth was not estimated at this location and we moved on to the next horizontal segment. If there were multiple modes, the one closest to the surface was defined as the bathymetry of the pond, as it was unlikely there are modes within a pond because the green laser is able to penetrate through the water column. The bathymetric elevation, hb, of the pond was determined as the elevation of the subsurface mode: 5hb=Pni_ma1.

Bathymetric elevation was determined for each 10 m horizontal section across the pond (Fig. d, blue, Eq. ). Next, we estimated pond depth by differencing the pond surface and bathymetry. We then multiplied this depth by the ratio of the refractive index of air to water following to derive the true melt pond depth, hmp, as follows: 6hmp=hs-hb×ηaηw, where hmp is the depth of the melt pond, hs is the elevation of the pond surface, hb is the elevation of the bathymetry, ηa is the refractive index of air (1.00029), and ηw is the refractive index of water (1.33567) . Pond depth (hmp) was determined for each 10 m along-track segment. To increase along-track resolution, a linear interpolator with a 5 m length was applied to obtain pond depth at 5 m intervals across the pond. The true elevation within each bin could be ±0.15 m from the estimated value (half of the 0.3 m bin width). When the melt pond surface and bathymetry elevations are differenced to determine the depth, uncertainty doubles because the pond surface and bathymetry uncertainties are additive (0.3 m), resulting in a total depth uncertainty of ±0.23 m after correction for refraction (0.3×ηaηw). At least one depth measurement and the melt pond start and end points are required for pond detection, and thus the minimum retrievable pond width is 20 m. The advantage of the UMD-MPA is that individual ponds were manually selected so that false positives are minimized. However, the manual process of identifying ponds is arduous and vulnerable to human error.

The DDA constitutes a family of fully automated algorithms designed to track complex surfaces in micro-pulse photon-counting lidar altimeter data, such as those of ICESat-2 . The DDA-bifurcate-seaice algorithm was designed to track height in complex sea ice topography and has the ability to simultaneously track two diverging surfaces. A full description of the algorithm can be found in , but we briefly describe it here. The DDA utilizes the full geolocated photon height point cloud as provided in the ATL03 data product . The algorithm employs the calculation of a density field for data aggregation and principles of auto-adaptive signal-to-noise thresholding and roughness determination (as described in ). The DDA has the ability to detect bifurcating reflectors and can accommodate situations where the stronger reflector can be the lower or the higher reflector, and the two reflectors may have different spatial distributions and material and reflection properties. The DDA includes a layer follower with automated adaptation to layer roughness. On rough surfaces, the DDA tracks at 2.5 m intervals to capture the varying surface and on smooth surfaces at 5 m intervals. These parameters are adjustable. At least three sequential depth measurements are required for pond detection, and hence the minimum retrievable pond width is 7.5 m on a rough surface and 15 m on a smooth surface. For comparison and consistency with the UMD-MPA, we resample the surfaces tracked by the DDA at 5 m intervals. The minimum elevation difference between the two tracked surfaces is adjustable within the DDA, and for the purposes of this work it is set at 0.2 m within the photon cloud, allowing for a minimum retrievable pond depth of 0.15 m (after correction for refraction). The DDA is automated, requiring no manual input, and can be applied in a systematic way. We use the DDA for comparison with the UMD-MPA and to extend the time series of the melt pond depths in summer 2020. The limitations of both the UMD-MPA and DDA are discussed in Sect. .

We examine the evolution of surface classifications throughout the melt season (Fig. ). Vertical gray bars indicate signal loss due to the requirement of 90 % cloud-free images (Sect. ) when there are fewer than 10 images in the 5 d period. At the beginning of the melt season a high percentage of pixels (>90 %) are classified as ice, and this is followed by a sharp drop to <80 % in mid-June. The ice pixel percentage decreases through mid-August and then becomes more variable. Melt pond pixels increase from 3.0 % on 13 June to 10.0 % on 15 June, and the maximum coverage is 15.3 % on 24 June. The percentage of pixels classified as melt ponds remains greater than 10 % until 2 July and makes up less than 5 % of each image from 23 July through the end of the study period (15 September). The open-water percentage is low (<5 %) at the beginning of the season and then increases and becomes more variable later in the season, with the highest open-water percentage from mid-July through mid-August. This indicates an increase in lateral melting of floes and a more dynamic, divergent ice cover. On average, open water makes up 14 % of the surface pixels in July and 17 % in August. The open-water percentage decreases in late August and September as leads begin to refreeze. Throughout the season, the pixels classified as other remain below 10 %. Towards the end of the season, refrozen leads appear in the Sentinel-2 scenes, and the algorithm classifies these areas as other, explaining the increase in other pixel percentages in September (Fig. ).

We examine the SIC derived from Sentinel-2 data in the study region. Mean SIC in the region was 91.6 % with a standard deviation of 15.0 %, and the median was 97.2 %. The difference between the median and mean indicates that there are some Sentinel-2 tiles with very low SIC or entirely open water. The SIC values ranged from 0 %–100 %, with 75 % of the SIC values greater than 92.8 % and 25 % greater than 99.0 %. As the melt season progressed, individual images had more variable SIC and the median SIC value decreased. Figure  shows the seasonal evolution of the melt parameters with SIC shown in Fig. a. SIC was consistently greater than 90 % through mid-June. On 27–28 June, imagery shows the ice separated from the landfast ice in the Lincoln Sea and along the western coasts of the Canadian Arctic Archipelago . At the same time, sea ice drift data indicate westward ice drift . These dynamics opened leads and reduced local ice concentration. Throughout July, the sea ice continued to separate from the coast, leaving large areas of open water. As the ice cover receded, Sentinel-2 images along the edge of the pack ice captured lower SIC (<80 %), and in the lowest latitudes of the study region, SIC values dropped below 20 % (Fig. a). The consolidated ice cover evolved into a mosaic of smaller floes with leads that grew in size as the floes melted laterally. Median SIC dropped below 80 % in late July, consistent with , who observed a sharp decrease from 95 % to 80 % SIC in early August in aerial observations of the SHEBA site.

We calculated MPF from the Sentinel-2 images in the study region with SIC >15 % (Fig. b). The average MPF in the region in summer 2020 was 6.5 % with a standard deviation of 6.5 %. The highest MPF for an individual Sentinel-2 scene was 31.6 %. This image is located just outside of the mouth of Nansen Sound at 82.3∘ N, 95.3∘ W but is far enough from that coastline that it does not contain landfast ice. Median MPF remains low, <5 %, through 17 June. We then see a sharp increase to 12.1 % in MPF on 18 June. The imagery is scarce between 18 and 22 June due to widespread cloud coverage. This weather system likely enhanced the melt , and when it passed, MPF was high, averaging 15.2 % between 24 and 29 June. The peak 5 d running mean MPF was 15.9 % on 24 June (Fig. b). MPF slowly decreased in July, and by August, MPF remained below 5 % for the remainder of the season. The evolution of melt in WorldView images, presented in Fig. f, follows a similar pattern: a sharp increase in MPF earlier in the season and a decrease in MPF by September. However, the images show a sustained high MPF (>20 %) through early August (Fig. f), indicative of the variability in MPF regionally and at smaller scales.

Small features on sea ice pose a challenge for satellite-derived classifications. We discuss the potential for misclassification of pixels and how the algorithm handles pixels containing multiple surface types (mixed pixels). When the algorithm encounters uncommon surfaces that do not fall squarely into the classification categories, the pixels may be misclassified. This includes ridge shadows, submerged sea ice, and very light melt ponds as previously discussed in . Misclassifications occur more frequently as the image resolution is lowered because each pixel covers more surface area. Given ponds can range in size from less than 1 m to hundreds of meters in diameter , there may be several surfaces within a Sentinel-2 10 m pixel. Mixed pixels are those pixels with a combination of surface conditions, whether the edge of an ice floe, containing ice and open water, or small melt ponds and drainage channels surrounded by sea ice. In these cases, it is difficult to robustly determine the pixel designation since the reflectance signature is not indicative of one particular surface type. So to mitigate the pixel misclassification errors, we introduce a category labeled other.

We examine the classification of Sentinel-2 imagery and temporally near-coincident (12 min time difference), but higher-resolution, WorldView commercial satellite imagery in the Canada Basin as the surface undergoes melt on 27 July 2020 (Fig. ) to demonstrate the occurrences of mixed pixels and pixel misclassification. We set the Sentinel-2 image to the bounds of the WorldView image applying the same classification algorithm as described above and compare results. In Fig. c–f, we show a segment of the WorldView and Sentinel-2 images and their classification masks. Figure e illustrates the high-resolution features visible in the WorldView-2 imagery. In the center of Fig. e, small melt ponds are connected by long, narrow drainage channels. As these drainage channels are on the order of 5–10 m in width, the Sentinel-2 imagery does not resolve these features, and pixels in this area are composed of both ice and meltwater (Fig. c). Here, pixels consisting of small melt ponds and drainage channels are classified as ice or other (Fig. d). Also, along the sea ice edge, where pixels contain both ice and water, the pixels are classified as other (Fig. d, green, center bottom). The other category includes complex ice types such as new ice, which appears gray in imagery and is not bright enough to be classified as ice. This occurrence is rare and happens towards the end of the melt season as leads and areas of open water start to freeze. Pixels categorized as other are not considered in the calculation of the derived parameters of MPF and SIC. Our analysis suggests that other pixels represent on average less than 10 % of all image pixels (see Sect.  and Fig. , green).

We estimate MPF of 7.6 % and 25.5 % from the Sentinel-2 image (Fig. d) and WorldView image (Fig. f), respectively, a difference of 18 percentage points. The underestimation of MPF (especially as the ice reaches the maximum MPF) in the lower-resolution image is consistent with previous studies . We also find SIC is 6.2 percentage points higher in the Sentinel-2 image than in the WorldView image. The next section further discusses the impact of these errors on the derived parameters.

Given the biases revealed between the Sentinel-2 and WorldView analysis shown in Fig.  and described in the previous section, we investigate the robustness of the parameters derived during the Sentinel-2 classification. In Sect.  we discussed the ability to resolve small melt features in WorldView imagery that are not resolvable in Sentinel-2 and showed an example in Fig. . Our goal is to assess the level to which MPF may be biased low due to the 10 m pixel resolution. In Sect.  we discussed the distribution of melt pond sizes detected in the WorldView imagery. The cumulative probability distribution (Fig. b) illustrates the prevalence of small ponds, showing that 73 % of ponds are smaller than 100 m2. This implies that approximately 73 % of individual ponds are not captured by the Sentinel-2 imagery, which has a 100 m2 pixel area (10 m pixel size). However, since these are small ponds, they account for only 38 % of the total pond area in the WorldView scenes, and Sentinel-2 imagery is able to capture approximately 62 % of the total pond area.

We compare MPF and SIC derived from Sentinel-2 with MPF and SIC derived from the higher-resolution WorldView imagery (Fig. ). We identify Sentinel-2 imagery captured within 24 h of the same 18 WorldView images (Sect. ), and we subsample the Sentinel-2 tiles to the extent of the WorldView image by matching ice features in the imagery. Although the imagery spans 2 years, we organize the findings by day of year to understand if there is a seasonal trend in the bias.

A comparison of the derived melt parameters from the classification of coincident images is shown in Fig.  and Table . In the beginning of the melt season, both datasets show consolidated ice with few or no signs of melt. The classification of the images results in a good agreement in the derived MPF and SIC. In the five scenes in the first half of June, MPF is less than 1 % in all Sentinel-2 and WorldView images. SIC is high in all the images (>90 %), and the Sentinel-2 SIC agrees to within 3 percentage points of the coincident WorldView SIC (Fig.  and Table ). As the melt season progresses, sea ice floes are more susceptible to breakup due to structural weakening induced by melt pond formation and enhanced dynamics as sea ice is in free drift. For this reason, there are smaller features that appear in the imagery scenes: smaller floes, brash ice, melt ponds, and drainage channels. Small features are not as well resolved by the lower resolution of Sentinel-2 imagery, and thus misclassification and mixed pixels are more common. This leads to weaker agreement of the derived parameters in Sentinel-2 versus those from the higher-resolution WorldView imagery, which still may be able to resolve these small features. The Sentinel-2 MPF is lower than MPF derived from WorldView images, as small ponds can go undetected or are classified as other pixels. From the end of June through mid-September, Sentinel-2 MPF is on average 12 % lower than the equivalent MPF derived from coincident WorldView imagery. There are two cases where MPF calculated for the Sentinel-2 image is greater than that of the WorldView image. In the imagery collected on 17 June 2020 the calculated MPF is 8.4 % and 2.4 % for the Sentinel-2 and WorldView images, respectively (Table ). In this scene (see Fig. b), there is level bare ice that appears blue in color, classified as melt pond in the Sentinel-2 imagery and ice in the WorldView imagery, resulting in a higher MPF for the Sentinel-2 scene than the WorldView scene. In the imagery captured on 3 September 2020 (WorldView subset shown in Fig. f), there are many ice fragments smaller than Sentinel-2's pixel size (10 m) classified as other or melt ponds in the Sentinel-2 imagery, falsely increasing the melt pond fraction. Our analysis shows that MPF can be biased low in the Sentinel-2 results by up to 20.7 % and averaging 7.2 % when small ponds are widespread across the surface. SIC is biased high by up to 16 % and averaging 4.3 %, increasing as the melt season progresses (Table ). The WorldView images better resolve these features and properly classify pixels as ice or open water.

In order to quantify the impact of pixel size on derived MPF, we look at the melt pond size distribution for each WorldView image with a coincident Sentinel-2 image. With knowledge of the WorldView pixel size, we can determine the area of each object in the binary image (as in Sect. ). For each WorldView image, we determine the total area of ponds with a size smaller than the Sentinel-2 pixel area (100 m2). The Sentinel-2 classification cannot resolve these small features as they are smaller than the pixel size. To quantify this, we calculate an adjusted Sentinel-2 MPF that adds the area of unresolved melt ponds into the MPF calculation for each pair of coincident Sentinel-2 (S2) and WorldView (WV) images: 8S2_MPFadj=S2_MPA+WV100_MPAS2surf_A×S2_SIC, where S2_MPFadj is the adjusted Sentinel-2 MPF, S2_MPA is the Sentinel-2 melt pond area, WV100_MPA is the area of ponds less than 100 m2 in size in the coincident WorldView image, S2surf_A is the surface area in the Sentinel-2 image, and S2_SIC is the Sentinel-2 SIC. The denominator on the right-hand side is a calculation of the total area of sea ice and melt ponds in the Sentinel-2 scene. We make the assumption that melt ponds smaller than the Sentinel-2 pixel size are classified as ice, so when making this adjustment, we hold the sea ice concentration constant, and the area of WorldView melt ponds less than 100 m2 replaces sea ice in the original Sentinel-2 classification. Table  provides the adjusted MPF results for each pair of WorldView and coincident Sentinel-2 images. Figure b shows the adjusted Sentinel-2 MPF as the sum of the royal blue and light blue bars.

Although the addition of ponds smaller than the Sentinel-2 pixel size through this adjustment increases the S2 MPF, making it more comparable to the WorldView-derived MPF, it does not account for the entire discrepancy between MPF derived from Sentinel-2 and WorldView (Fig. b, Table ). The average difference between MPF derived from Sentinel-2 and WorldView decreased from 7.2 % to 3.6 % when the subpixel-size WorldView ponds were accounted for. Although this methodology accounts for individual small ponds identified in WorldView imagery, we have not accounted for subpixel-size areas that are connected to larger ponds. Where the Sentinel-2 pixels may be classified as ice along the edges of ponds, a portion of that pixel may be a melt pond and properly classified as such in the WorldView classification. This scenario was not accounted for in the adjusted Sentinel-2 MPF and may account for some of the remaining bias between Sentinel-2 and WorldView imagery. We also note that the Sentinel-2-derived SIC is on average 4.3 percentage points greater than that derived from WorldView, and with a higher SIC and sea ice area per scene, this contributes to a lower MPF.

We also consider the minimum resolvable pond area when the UMD-MPA and DDA are used to map pond depths, based on our melt pond size distribution analysis in Sect. . The algorithm capabilities are linked to the minimum pond width that can be detected, which is 20 m for the UMD-MPA tracking (Sect. ) and 7.5 m for the DDA tracking (Sect. ). To estimate the approximate area of ponds with such widths, we assume a circular melt pond, resulting in the minimum detectable melt pond area being 314 and 44 m2 for the UMD-MPA and DDA, respectively. Figure b shows the cumulative distribution of individual melt pond area, with the minimum retrievable pond areas marked in magenta for the UMD-MPA and green for the DDA. Note that the cumulative distribution is shown for individual ponds, not total ponded area. According to the WorldView imagery, approximately 83 % of the total ponded area is made up of ponds with an area smaller than the UMD-MPA minimum resolvable size, suggesting the UMD-MPA is missing a large majority of ponds. However, the WorldView imagery melt pond distribution suggests that only 14 % of the total ponded area is made up of ponds smaller than 44 m2, the minimum detectable area for the DDA. The divergence in the UMD-MPA and DDA results is due to the inability of the UMD-MPA to track ponds smaller than 20 m in width, which make up 96 % of the melt ponds by number based on the WorldView imagery classification.

The primary advantage of the DDA over the UMD-MPA is that it can be run over any segment of ICESat-2 ATL03 data. The DDA includes an algorithm module that avoids clouds, including low-lying clouds down to 150 m above the digital elevation model (DEM). The DDA does not rely on the cloud flag reported in the atmospheric data product ATL09 , which is more restrictive; thus the DDA increases the amount of data where sea ice surfaces and melt ponds can be detected. The DDA is fully automated without any manual user input beyond prescribed parameters . However, as a result of this automation, two scenarios associated with complex sea ice topography can result in false positive melt pond detection by the DDA. These cases are discussed in more detail in , but here we illustrate two cases (Fig. ) and briefly describe our approach to reduce the impact of these issues. In the first case, complex surface topography associated with heavily deformed and ridged ice can result in the DDA tracking two surfaces between sea ice ridges (Fig. a–b). Here, the surface tracking is not across a level pond surface, but instead the algorithm bifurcates, and the first pass connects ridge sails and the second pass tracks rubble between the ridges. Anomalies such as these can be detected and discarded by flagging ponds that have surfaces with a standard deviation >0.05 m. This scenario has been eliminated with updates to the algorithm described in .

The second type of false positive pond detection occurs due to the dead time of the ATLAS photon detectors when a strong surface return results in saturation of the detectors and a period of 3.2 ns when no additional photons can be detected . Following the detector dead time, photons are once again reported, resulting in a secondary “surface return” 0.5 m below the true surface. Over sea ice surfaces, detector saturation commonly occurs over very bright surfaces such as specular leads and melt ponds . In this scenario, the DDA tracks the secondary return as the bathymetry of a pond as seen in Fig. c–d. To detect these occurrences, we look at the mean density of the surface return. The distribution of the mean density of the surface returns reveals a bimodal histogram. We have determined that the higher mode may correspond to a scenario where the surface is saturated and a secondary surface return results in false positives in the DDA tracking algorithm. Retracking anomalies due to dead time can be identified by depth measurements corresponding to the dead-time effect (0.5–0.6 m), where the surface mean density is greater than the minimum between the two modes in the mean density distribution.

From 1 June through 17 June, MPF was less than 5 % and SIC greater than 99 % (Fig. a–b). Figure  shows an example of an unponded ice surface on 9 June (Fig. a), and early melt occurred in the image observed on 17 June (Fig. b). During this time the median DDA-derived depth remained below 0.5 m, and there were no ponds tracked by the UMD-MPA. At the MOSAiC site in the same melt season as our study but in the Fram Strait east of our study region, continuous melt started in mid-June 2020 expanding existing ponds and increasing the pond areal coverage .

Our results show a sharp increase in MPF in mid-June (Fig. b), consistent with where aerial observations over the field site show a rapid increase in pond fraction over the study area from 5 % to 20 % in just a few days. also find a sharp increase in the modeled MPF early in the melt season in their standard multiyear case. We found MPF greater than 10 % from 23 June through 2 July, with a maximum MPF of 16 % on 24 June. At the MOSAiC site on primarily second-year ice, ponds greater than 100 m in diameter were observed on 1 July 2020 . Maximum pond coverage occurred later in the season at SHEBA in 1998 (24 % on 7 August) and at the MOSAiC site in 2020 (21 % on 26 July) . In the second half of June, the first ponds were tracked by the UMD-MPA (22 June), and both UMD-MPA and DDA median pond depths increased through the end of June. Similarly, at MOSAiC, melt pond depths increased through early July. found gradually increasing pond depth in their model. analyzed 220 pond depth measurements on multiyear ice within the Canadian Arctic Archipelago from 27 May to 26 June 1994. These ponds had a mean depth of 0.27 m with a standard deviation of 0.13 m. The UMD-MPA measurements revealed a higher mean melt pond depth for this time period (0.75 m ± 0.66 m), but this is likely due to the part of the melt pond size distribution sampled by ICESat-2 and the UMD-MPA minimum observable depth of 0.23 m. From 4–26 June 2020, the DDA-tracked ponds had a mean pond depth of 0.54 m with a standard deviation of 0.4 m.

Following the maximum MPF on 24 June, there was a decrease in MPF, consistent with and , who both described a decrease in pond coverage as meltwater was efficiently routed through drainage channels, ice permeability increased, and meltwater percolated through the sea ice. At MOSAiC, a drainage event that occurred in mid-July reduced the pond area by 5 % . and suggest a second mode of MPF as melt ponds spread laterally and connect through drainage channels, but we did not see this in our results. This could be a result of the low resolution of the Sentinel-2 imagery where the smaller drainage channels that occurred later in the melt season were not resolved well in the imagery, and melt pond pixels were classified as ice pixels instead. The UMD-MPA median pond depth increased throughout July, from 0.32 m on 1 July to 0.93 on 30 July, whereas the DDA depth increased through 16 July, reached a median depth of 0.78, and then was less than 0.5 m from 19 July through 14 August. These contrasting results demonstrate the bias of the UMD-MPA towards the identification of larger melt ponds. However, evolution is highly dependent on local weather and sea ice conditions , and it is likely that the UMD-MPA and DDA were tracking ponds under different atmospheric and sea ice conditions. The simulated pond depth in surpassed 1 m in early July and remained above 1 m for the remainder of the melt season, agreeing well with the UMD-MPA observations. The gradual increase in melt pond depth throughout the season was also observed at SHEBA . Observations at MOSAiC show that pond depth increased over time and melted through the first-year ice by late July .

The formation of ice lids on melt ponds is a sudden process, as temperatures below freezing will quickly freeze the top layer of the pond, drastically reducing pond fraction . Small, shallow ponds form lids before larger, deeper ponds . The MOSAiC observatory was relocated to the central Arctic (approximately at 89∘ N) in mid-August, and although MPF was greater than 30 % on 4 September, all ponds had refrozen by 6 September 2020, effectively reducing MPF to 0 %. The number of melt ponds tracked by the UMD-MPA and DDA (gray histogram in Fig. 5c) significantly decreased towards the end of August, indicating either that ponds had drained or that a lid had formed preventing laser penetration into the pond. This is consistent with our findings of low MPF during this period. The WorldView image observed on 3 September (Fig. e) shows light gray ponds, indicative of pond lid formation.