Detecting Melt Pond Onset on Landfast Arctic Sea Ice Using a Dual C-Band Satellite Approach

Abstract

The presence of melt ponds on the surface of Arctic Sea ice affects its albedo, thermal properties, and overall melting rate; thus, the detection of melt pond onset is of significant importance for understanding the Arctic’s changing climate. This study investigates the utility of a novel method for detecting the onset of melt ponds on sea ice using a satellite-based, dual-sensor C-band approach, whereby Sentinel-1 provides horizontally polarized (HH) data and Advanced SCATterometer (ASCAT) provides vertically polarized (VV) data. The co-polarized ratio (VV/HH) is used to detect the presence of melt ponds on landfast sea ice in the Canadian Arctic Archipelago in 2017 and 2018. ERA-5 air temperature and wind speed re-analysis datasets are used to establish the VV/HH threshold for pond onset detection, which have been further validated by Landsat-8 reflectance. The co-polarized ratio threshold of three standard deviations from the late winter season (April) mean co-pol ratio values are used for assessing pond onset detection associated with the air temperature and wind speed data, along with visual observations from Sentinel-1 and cloud-free Sentinel-2 imagery. In 2017, the pond onset detection rates were 70.59% for FYI and 92.3% for MYI. Results suggest that this method, because of its dual-platform application, has potential for providing large-area coverage estimation of the timing of sea ice melt pond onset using different earth observation satellites.

1. Introduction

The dramatic decline in the Arctic Sea ice extent (i.e., about 13.1% per decade since 1979) [1] indicates that the Arctic is likely to be practically sea-ice-free in September at least once before 2050 [2]. The Arctic Sea ice cover plays a key role in the global climate system by regulating the transfer of moisture, heat, and energy between the atmosphere and ocean. Arctic Sea ice is transitioning from a regime of old, thick, multiyear ice to a younger, thinner, first-year ice regime due to increasing ice melting [3,4,5,6]. The presence of melt ponds on the surface of sea ice can affect its albedo, thermal properties, and overall melting rate; thus, the detection of melt pond formation and evolution is important for understanding the Arctic’s changing climate.

Melt pond onset (PO) refers to the time when melt ponds start to form on sea ice. They form annually on sea ice when the surrounding snow and ice begin to melt during the transition from spring to summer. This shifts the surface’s thermodynamic and radiative state and initiates a positive ice–albedo feedback, which enhances the surface melting process [5,7]. The formation of melt ponds earlier in the summer, which can cover 50–60% of the Arctic Sea ice extent [8,9,10,11], can lead to increased solar absorption and a decline in sea ice area and thickness. Considering the significance of melt ponds in Earth’s energy budget, it is critical to continuously monitor their formation and evolution on Arctic Sea ice, as they are not adequately parameterized in climate models, which impacts seasonal sea ice forecasting.

Both the detection and prediction of sea ice melt pond onset timing are important yet challenging tasks for sea ice monitoring, as sea ice models are not yet fully optimized to account for the spatiotemporal variability of the formation and distribution of melt ponds [12]. Although the dynamics of melt pond onset, formation, and evolution have been well researched at the local to regional scales [13,14,15,16], our understanding at larger spatial and temporal scales is limited. Low- to medium-resolution optical satellite data, i.e., Moderate-Resolution Imaging Spectroradiometer (MODIS), Medium-Resolution Imaging Spectrometer (MERIS), and the Sentinel-2 and Landsat series, have been widely used to distinguish melt ponds from sea ice and open water by analyzing the distinctive spectral albedo signature of ponds [17,18,19,20,21,22]. However, as the optical satellite data application is limited to daytime and cloud-free conditions, it restricts observation to detect melt ponds on Arctic Sea ice.

The efficacy of microwave remote sensing for detecting melt ponds on Arctic Sea ice is well established because of its all-weather and continuous imaging capability from regional to pan-Arctic scales [15,23,24,25]. Synthetic Aperture Radar (SAR) is an active microwave sensor that offers high-resolution (<100 m) sea ice data, but its limitations include spatial coverage and temporal resolution; alternatively, spaceborne scatterometers are active sensors that provide regional sea ice coverage on a daily time scale with lower resolution [16,24,26,27,28,29,30,31].

The evolution of SAR and/or scatterometers has witnessed a notable advancement, transitioning from a conventional single polarization channel (VV or HH mode) to quad-polarimetric modes, which encompass HH, HV, VH, and VV polarizations [15,32]. Both single and quad-polarimetric data have potential in interpreting ambiguities in the SAR backscatter signature because of wave roughness on melt pond surfaces during the advanced melt stage [24]. Combining the channels into polarimetric SAR features, such as channel ratios, emphasizes the relative properties of different polarization rather than the absolute intensity of the radar signal. Such ratios (co-pol-VV/HH or cross pol-HV/VH) are sensitive to the geometric or shape properties of the snow–ice system [33]. Several studies [34,35,36,37,38] have utilized the co-pol ratio from C-band SAR and scatterometers and classified Arctic Sea ice types (first-year ice (FYI) and multi-year ice (MYI)) and open water areas based on backscatter magnitude and polarization differences. These studies suggested that the co-pol ratio could be a useful measure to detect surface water on sea ice rather than using a single polarization (VV or HH). One such study [39] found that the co-pol ratio from ERS-1 SAR images of summer MYI exhibited daily backscatter variability due to wind-driven changes in melt pond surface roughness. Another study [40] observed that the co-pol ratio from ENVISAT ASAR data had potential for detecting melt ponds on FYI and pond fractions at low incidence angles. In addition, the authors of [24,32] used the co-pol ratio to identify melt ponds on FYI due to the dielectric property contrast between ponds and sea ice. Scharien et al. [24] utilized in situ polarimetric C-band scatterometer measurements of an individual pond and bare ice patches on smooth FYI and found that co-pol ratio values were higher for melt pond sea ice than winter sea ice.

Our primary goal is to evaluate the efficacy of a dual-sensor C-band co-pol ratio method for providing robust estimates of PO on both FYI and MYI ice types in the Canadian Arctic Archipelago. In doing so, we address a fundamental earth observation research question. Can two different satellite C-band sensor platforms (one a SAR and the other a scatterometer) provide the spatial and temporal resolution necessary to resolve the rapid transition from late winter to melt pond sea ice?

2. Materials and Method

2.1. Study Area

The study area for this analysis is landfast sea ice in the Canadian Arctic Archipelago (CAA; Figure 1) during 2017 and 2018. To illustrate the spatiotemporal variability in PO timings, the CAA was divided into three latitudinally separate regions: the northern CAA, the Parry Channel (Central CAA), and the southern CAA. The northern CAA, often referred to as the Queen Elizabeth Islands (QEI), comprises numerous islands interconnected by channels containing both FYI and MYI. Sea ice in this region remains landfast and stable for nearly all of its lifecycle, only transitioning to a mobile/drifting state during the late melt season, rendering it ideal for time series analysis due to continuous and stationary imaging availability. Thick MYI consistently dominates this region, serving as a significant repository for MYI transported from the Arctic Ocean [41,42,43]. The Parry Channel is the northwest sea passage extending from Lancaster Sound, passing through Barrow Strait, and westward into Viscount Melville Sound. This area is largely dominated by MYI in both summer and winter [42]. During the melt season, the sea ice starts breaking up rapidly and the region becomes mostly ice-free, making passage available for shipping. The southern CAA consists primarily of FYI, interspersed with areas of MYI transported to the region during the previous summer.

2.2. Data and Processing Method

2.2.1. Space-Borne Scatterometer Data (ASCAT VV Data)

We use VV-polarized data acquired from ASCAT from 1 April to 31 July in 2017 and 2018. ASCAT is a C-band, active, real aperture remote sensing system (5.255 GHz center frequency) on board the polar-orbiting European Organization for the Exploitation of Meteorological Satellites (EUMETSAT), MetOp-A and MetOp-B, launched in 2006 and 2012, respectively. We used the VV-polarized ASCAT GDS Level 1 sigma-naught (${\sigma }_{VV}^0$) 12.5 km swath grid data available from the EUMETSAT Data Centre (Figure 2a).

The Level 1 data provided σ° triplets (from the three beams of ASCAT). These were incidence angle normalized to 35° [32] and then the mean of the triplets was taken, which constituted one ASCAT point.

To ensure data quality, we retained only points that met the criteria of good radiometric quality and low land fraction. We used a land fraction of 1% instead of 0% because there are many locations in the CAA that have very small islands and/or fractions of land. Thus, a compromise was made to maintain enough data points while minimizing land contamination.

For the study area and period, up to 29 swaths were available per day. The different orientations of each day’s swaths resulted in a heterogeneous distribution of points at different locations and distances from each other (Figure 2b). The points for an entire day were interpolated to a regular 5 km grid using Inverse Distance Weighting with a maximum search radius of 7.5 km (Figure 2c). This produced a daily weighted mean backscatter product, which was constrained by a 25 km land buffer to further reduce land contamination (Figure 2d). The mean daily backscatter for the time series (1 April to 31 July) were derived for 2017 and 2018. ASCAT data were projected to Canadian Polar Stereographic WGS84 (EPSG: 5937).

In the daily time series of ASCAT data, occasional gaps arose due to factors such as instrumental malfunctions, satellite orbit characteristics, or adverse weather conditions, affecting signal acquisition [44]. To mitigate this, we employed a weighted linear-moving-average temporal imputation method [45]. This approach effectively addresses missing values, ensuring continuity and reliability of the dataset.

2.2.2. SAR (Sentinel-1 HH Data)

We used Sentinel-1 (S-1) SAR images acquired from 1 April to 31 July from 2017 and 2018. We used HH-polarized imagery acquired in EW (extra-wide swath) mode, available at Copernicus Open Access Hub. We accessed S-1 data using the Google Earth Engine (GEE) cloud processing platform. This collection includes S-1 (A and B) Ground Range Detected (GRD) images in both ascending and descending orbits. In the study area and period, usually one to four images were available per day, with occasional days with no data. GEE uses the Sentinel Application Platform (SNAP) to derive the HH-polarized backscatter coefficient, ${\sigma }_{HH}^0$. The processing steps included thermal noise removal, radiometric calibration, and orthorectification to Canadian Polar Stereographic WGS84, EPSG: 5937. S-1 data were incidence angle normalized to 35° using a dependency of −0.19 dB/1° (mean between FYI and MYI) [32,46]. The available daily S-1 scenes were averaged to produce a daily mean backscatter product. While S-1 offers extensive spatial coverage in the Arctic, constructing high-temporal-resolution SAR datasets (e.g., 1–2 days) remains challenging, leading to gaps in the data [47]. To fill data gaps in the time series, we used a weighted linear-moving-average temporal imputation [45].

2.2.3. ERA-5 Data

ERA-5 is a fifth generation European Centre for Medium-Range Weather Forecasts (ECMWF) comprehensive atmospheric re-analysis model of the global climate, coupled with a land surface and a wave model (covering the period from 1950 to present) [48]. It is produced by the Copernicus Climate Change Service at ECMWF and provides hourly estimates of many atmospheric, land, and oceanic variables. These data cover the Earth on a 30 km grid, and we used the daily-averaged 2 m surface air temperature and 10 m wind components. The data are available at: Daily statistics calculated from ERA5 data.

2.2.4. Canadian Ice Service Digital Archive (CISDA)

We employed weekly regional western and eastern Arctic ice charts from the Canadian Ice Service Digital Archive (CISDA) to obtain sea ice break-up dates in the CAA for years 2017 and 2018, to support visual verification of the advanced melt season and break-up of landfast ice via GEE time series.

2.2.5. Validation Data

For PO date validation, we used Landsat 8 red-band reflectance data using the Google Earth Engine (GEE Extraction tool) (see Section 3.4.). In addition, we used Sentinel-2 level 1C optical data as manual visualization validation for PO identification.

3. PO Detection Algorithm Development

3.1. Selection of Sample Points

The 5 km regular grid points used for ASCAT interpolation represent the available sample points. The 25 km land buffer and constraint to landfast ice limited the number of available points in the study area. We selected 34 FYI and 26 MYI sites in 2017, and 26 FYI and 25 MYI sites in 2018. Discrimination of FYI and MYI was performed using an ASCAT threshold of −17 dB, with FYI represented by values below the threshold [36]. The selection criteria were further aided by CISDA ice charts and visual inspection of S-1 scenes, to ensure homogeneous regions at the sites. To characterize the variability in surface conditions, we selected a variety of relatively smooth and rough FYI and MYI points by visual inspection of backscatter.

Each sample site encompassed approximately 1 to 7 individual ASCAT measurements. Sampling of S-1 data used a 3.5 km radius, resulting in 18 to 20 SAR pixels per point. The spatial scales and number of measurements that constituted ASCAT and S-1 data were different. However, the large number of measurements and spatial averaging that make up the mean values of each sample site helped to reduce statistical artifacts stemming from these differences.

3.2. Dual-Sensor Co-Pol Ratio

Our dual-sensor method is dependent on ${\sigma }_{VV}^0$ from ASCAT and ${\sigma }_{HH}^0$ from S1-A/B. The co-polarized ratio (co-pol ratio) ${\gamma }_{co}$ (dB) is expressed as (1):

$${\gamma }_{co}=\frac{{\sigma }_{VV}^0}{{\sigma }_{HH}^0}$$

(1)

The primary factor that influences the co-pol ratio is the microwave interaction with different interfaces of varied dielectric properties [37]. For sea ice, horizontal polarized waves can reflect more from smooth surfaces, whereas vertical polarized waves can transmit proportionately more into smooth surfaces, resulting in negative ${\gamma }_{co}$ values. This phenomena can be explained by Fresnel reflection and is often observed on smooth FYI and smooth, thin ice with substantial dielectric mismatches [37]. As the sea ice surface becomes rougher, the co-pol ratio tends toward zero. Water usually exhibits positive ${\gamma }_{co}$ values due to Bragg scattering and its high dielectric properties [37,49]. For melt ponds, backscatter mainly depends on surface roughness, which is largely dictated by surface wind conditions [16,32,50,51]. Under calm conditions where the wind speed is less than ~3 m/s (as measured from a 10 m height), scattering from melt ponds is mostly specular and exhibits very low σ° values [15,24], which may also result in ${\gamma }_{co}$ values near 0 dB. Melt pond sea ice will exhibit ${\gamma }_{co}$ values that are a mix of water and sea ice. In such cases, we rely on ${\gamma }_{co}$ from the melt pond water to dominate the signal.

3.3. Pond Onset Thresholds’ Development

We required three conditions to be met for PO to be identified. We used a site-specific co-pol threshold (${TR\gamma }_{co})$, a wind speed threshold (${TR}_{WS}$), and an air temperature threshold (${TR}_{temp}$).

3.3.1. Co-Pol Ratio Threshold

During the pre-melt season on snow-covered FYI and MYI, microwave backscatter, σ°, was generally stable from day-to-day. The transition from late winter to early spring introduces liquid water into the snow volume, which increased σ° for FYI and decreased it for MYI [28,52,53,54]. The primary reason is due to the thickness of the snow cover on MYI, leading to higher amounts of microwave absorption, thereby lowering σ°. This change in σ° can be detected using time series of both ${\sigma }_{VV}^0$ and ${\sigma }_{HH}^0$, which also results in a change in time series of ${\gamma }_{co}$ values. We estimated PO using a threshold approach based on the change from winter mean, ${\gamma }_{co(w.mean)}$, which has been previously employed to detect melt onset [16,27,52,55]. Measurements during the month of April were used for the winter baseline, since air and snow temperature-induced changes to dielectric properties and σ° were nearly negligible because temperatures remained substantially below freezing [16].

First, we calculated the April (day of year, DOY 91–120) ${\gamma }_{co}$ mean (${\gamma }_{co(w.mean)}$) and standard deviation (${\gamma }_{co(w.std)}$) for each site, in dB, to obtain candidate estimates of the co-pol threshold:

$${TR\gamma }_{co}={\gamma }_{co(w.mean)}+{\gamma }_{co(w.std)}$$

(2)

We added $2{\gamma }_{co(w.std)}$, $2.5{\gamma }_{co(w.std)}$, and $3{\gamma }_{co(w.std)}$ to site-specific ${\gamma }_{co(w.mean)}$ for estimating the variability in the PO date, which is based on fitting the selected addends to the training datasets (see Section 4).

To calculate the mean ${\gamma }_{co}$ change/transition from winter to PO, we subtracted ${\gamma }_{co(w.mean)}$ from mean ${\gamma }_{co}$ at PO (${\gamma }_{co\left(PO\right)}$). We used the following equation to measure the transition threshold:

$${TH\gamma }_{co}={\gamma }_{co\left(PO\right)}-{\gamma }_{co(w.mean)}$$

(3)

3.3.2. Wind Speed Threshold

Wind speed is an important aspect of this detection method because it changes the surface roughness of the melt ponds, which affects microwave backscatter. When the wind speed is low, there is little or no effect on Bragg scattering, but as the wind speed increases, backscatter becomes more pronounced [9,16,56,57]. Wind-roughened ponds create a difference in ${\sigma }_{VV}^0$ and ${\sigma }_{HH}^0$, resulting in a ${\gamma }_{co}$ increase, from which PO can be detected [15]. In this study, we used a wind speed >3 m/s, irrespective of direction, as a threshold $\left({TR}_{WS}\right)$ to detect PO.

3.3.3. Air Temperature Threshold

We calculated our ${TR\gamma }_{co}$ statistics during DOY 91–120 (April 2017 and 2018), when the air temperature is representative of winter conditions. We used the 2 m air temperature from ERA-5 to create our air temperature threshold. When the 2 m air temperature reached freezing, we used it as threshold (${TR}_{temp}$) for PO detection.

3.4. Validation

We used spectral albedo/reflectance (636–673 nm wavelength, Band 4: red band) of Landsat 8 to validate PO [10,58]. The areally averaged contrast between snow and melt ponds, presented in [56], was strong in red-band wavelengths. The authors of [56] noted that the shift from snow/ice to the initial formation of melt ponds occurs, for 600–650 nm wavelength, with a decrease in reflectance starting from 0.8. We used a threshold reflectance ($W_{refl}$) window, whereby PO is expected to occur between the last late winter reflectance of >0.7 and the first post PO reflectance of <0.7. If our detected PO date fell within this window, we considered the PO date to be accurate (see Section 4.3 and Supplementary Tables S1 and S2). The use of a 0.7 reflectance value is somewhat arbitrary; however, in practice, the red-band reflectance dropped precipitously from winter (>0.8) to post-PO (<0.5) conditions [56], and since optical imagery is only occasionally available, the window straddles the PO date.

4. Results and Discussion

This section presents a comprehensive analysis. First, Section 4.1. examines the effect of meteorological parameters. Then, Section 4.2. provides detailed statistics. Finally, the results of the PO detection algorithm are discussed through site analysis in Section 4.3. and spatiotemporal variability in Section 4.4. Additionally, the limitations of the research are addressed to provide a complete understanding of the findings.

4.1. Factors Affecting the Co-Pol Ratio

As the air temperature rises during the seasonal progression, the snow-covered sea ice surface starts to melt to form melt ponds. Thereafter, both air temperature and wind speed affect the ${\gamma }_{co}$ of melt pond sea ice. To assess the prevalent air temperature conditions during the CAA melt season, we observed the 2 m air temperature for all sites between DOY 90 and 120 (from late winter to advanced melt; Figure 3). The mean daily 2 m air temperatures from ERA-5 data in April (DOY 91–120) for 2017 and 2018 among all study sites were −17.2 ± 2.6 °C and −16.7 ± 2.8 °C, respectively.

We assessed the wind speed conditions at all sites (both FYI and MYI) for the period DOY 130 to 191 of both years. The mean daily wind speed during the melt season for all sites in 2017 was 4.11 ± 1.15 ms⁻¹, with a minimum of 1.8 ms⁻¹ and a maximum of 7.21 ms⁻¹. Similarly, for 2018, the mean daily wind speed for all sites was 4.9 ± 1.5 ms⁻¹, with a minimum of 2.3 ms⁻¹ and a maximum of 8.5 ms⁻¹. The mean daily wind speed for all sites during the melt season in both years combined was 4.5 ± 1.3 ms⁻¹. Wind speed distribution over both years is illustrated in Figure 4.

4.2. C-Band Statistics for PO Date Detection

We analyzed our site-specific ${\gamma }_{co}$ time series by adding $2{\gamma }_{co(w.std)}$, $2.5{\gamma }_{co(w.std)}$, and $3{\gamma }_{co(w.std)}$ to each site-specific ${\gamma }_{co(w.mean)}$ in 2017 as training data. We validated our time series site analysis using the threshold reflectance window, $W_{refl}$, and performed a sensitivity analysis with the results from our validated sites. The results showed that using ${\gamma }_{co(w.mean)}+3{\gamma }_{co(w.std)}$ provided less false detection of PO dates than the other two addends. Using the addends $2{\gamma }_{co(w.std)}$ and $2.5{\gamma }_{co(w.std)}$ showed early PO detection in sites when we assessed this with the validation window. Conversely, addends 3.5${\gamma }_{co(w.std)}$ and 4${\gamma }_{co(w.std)}$ to each site-specific ${\gamma }_{co(w.mean)}$ were not considered in order to mitigate potential outlier effects, thereby enhancing statistical robustness and the accuracy of representing typical site conditions. The detection accuracy was 70.59% (2017) for FYI and 92.3% (2017) for MYI (

Table 1. PO date detection accuracies for selected addends to site-specific April means.

Year	FYI Overall PO Detection Percentage			MYI Overall PO Detection Percentage
	${\gamma }_{co(w.mean)}+2{\gamma }_{co(w.std)}$	${\gamma }_{co(w.mean)}+2.5{\gamma }_{co(w.std)}$	${\gamma }_{co(w.mean)}+3{\gamma }_{co(w.std)}$	${\gamma }_{co(w.mean)}+2{\gamma }_{co(w.std)}$	${\gamma }_{co(w.mean)}+2.5{\gamma }_{co(w.std)}$	${\gamma }_{co(w.mean)}+3{\gamma }_{co(w.std)}$
2017	61.75	67.6	70.59	73.07	76.9	92.3
2018	57.7	57.7	61.53	64	64	96

). We applied the same technique to our 2018 data to compare the results using the selected addends to ${\gamma }_{co(w.mean)}$. Table 1 presents detection of the percentage of validated PO date results from 2017 and 2018 datasets. Considering the higher accuracy of the $3{\gamma }_{co(w.std)}$ addend, all subsequent analysis used ${\gamma }_{co(w.mean)}+3{\gamma }_{co(w.std)}$.

Supplementary Tables S1 and S2 present the FYI and MYI statistics for 2017 and 2018, respectively (Supplementary Materials). The validation window indicates the late winter reflectance value and DOY before PO and the post-PO reflectance value and DOY. If the ${\gamma }_{co}$ detected PO date (DOY) was ≥ the late winter DOY and ≤ the post-PO DOY, we considered it an accurate PO date. The empty cells represent no value due to the limitations discussed in Section 4.5.

Our method detected PO dates more accurately from MYI than FYI. We assumed that this occurred because the sea ice type characteristics related to backscatter and environmental factors (wind speed and air temperature) were more consistent over the MYI pond surface compared to the FYI pond surface. It has been shown that FYI melt pond fractions are larger and more regionally variable than MYI pond fractions owing to smoother ice surfaces and thinner snow covers on FYI compared to MYI [59]. C-band April ${\gamma }_{co(w.mean)}$ for FYI 2017 and 2018 sites were 0.4 ± 0.98 and 1.42 ± 2.1, respectively, and MYI 2017 and 2018 sites were −0.16 ± 0.81 and −0.07 ± 0.82, respectively. ${\gamma }_{co(w.mean)}$ for all sites in both years (2017 and 2018) was 0.39 ± 0.53 dB, where FYI showed more site-to-site variability in ${\gamma }_{co(w.mean)}$ than MYI (see Supplementary Tables S1 and S2). ${\gamma }_{co(w.std)}$ for all FYI sites was 0.58 ± 0.02, and for all MYI sites was 0.37 ± 0.01. A larger mean standard deviation for FYI could suggest larger site-to-site variability or more heterogeneity in the FYI ice types resulting from a mix of both thick and thin snow covers that produce more variable melt pond fraction surfaces. The greater ${\gamma }_{co(w.std)}$ for FYI means that a larger ${\gamma }_{co}$ increase is required to detect PO for FYI than for MYI.

For the FYI 2017 and 2018 sites, ${TH\gamma }_{co}$ was 2.43 dB and 1.42 dB, respectively (mean ${TH\gamma }_{co}$ for FYI = 1.92 dB), whereas for the MYI sites in 2017 and 2018, ${TH\gamma }_{co}$ was 1.94 and 2.03 dB, respectively (mean ${TH\gamma }_{co}$ for MYI = 1.98 dB). For all sites, the mean ${TH\gamma }_{co}$ was 1.96 dB, which indicated that from ${\gamma }_{co(w.mean)}$ to ${\gamma }_{co\left(PO\right)}$, the mean ${\gamma }_{co}$ value increase was 1.96 dB to detect PO in ${\gamma }_{co}$ time series. The mean wind speed for all sites at PO was 5.5 ms⁻¹, and the mean air temperature was 1.48 °C.

We found that there was a significant (95% confidence) positive correlation (r = 0.485) between co-pol ratio values and 2 m air temperature (from DOY 140 to 170) for all the sites, which indicated the variability of site-to-site PO DOY (Section 4.3 and Section 4.4).

4.3. Co-Pol Ratio Time Series

In this section, we assess the ${\gamma }_{co}$ time series method for 2 FYI and 2 MYI representative sites, out of the 111 sample sites under different climatic parameters (wind speed, air temperature, and reflectance values). We acknowledge that the changes in ${\sigma }_{VV}^0$ from ASCAT and ${\sigma }_{HH}^0$ from Sentinel-1 were not associated with each other since these were from different sensors. As a result, we resampled the values from both sensors, since they affected the ${\gamma }_{co}$ values in the time series. Samples were selected based on their ice type and, most importantly, availability of cloud-free optical imagery near to or after PO occurred.

FYI_26_2017, located in the Southern CAA, and FYI_9_2018 located on Northern CAA in the Parry Channel, were selected as FYI test cases (Figure 1).

4.3.1. FYI

In April (DOY 91–120), site FYI_26_2017 had a ${\gamma }_{co(w.mean)}$ of −0.68 ± 0.59 dB and the ${\sigma }_{VV}^0$ and ${\sigma }_{HH}^0$ values were low and stable, ranging from −21 dB to −19 dB, with air temperatures consistently below −5 °C (Figure 5). The backscatter values indicated smooth FYI [37]. The stable and low winter ${\sigma }_{VV}^0$ and ${\sigma }_{HH}^0$ observed for FYI were largely due to specular surface scattering and minimal volume scattering from smooth interfaces between snow and sea ice. Considering the air temperature, there were likely only small changes in snow volume dielectric properties until melt onset (MO) began [53,55].

Time series indicated that early melt started around DOY 139 when increased warming (−0.5 °C) first caused an increase in the brine volume at the snow–sea ice interface, which in turn increased the volume scattering, resulting in the backscatter upturn [16,27,43] on DOY 140 as MO. As daily mean air temperatures approached 0 °C, the snowpack transitioned from the pendular regime, characterized by liquid water held in the interstices, to the funicular regime, where sufficient liquid water content breaks grain bonds, with the transition occurring at approximately 7% liquid water by volume [60].

As ${\sigma }_{VV}^0$ started to increase and became greater than ${\sigma }_{HH,}^0, {\gamma }_{co}$ began to increase when the ${TR}_{temp}$ also reached 0 °C and the possibility of PO increased. The reason for the backscatter increase was the increasing surface roughness from the moister snow surface due to the air temperature rise and the volume scattering from the denser, wet snow surface. The daily wind speed occasionally reached ≥ 3 ms⁻¹, which would cause enough surface roughness on the melt ponds to produce a positive ${\gamma }_{co}$ value. As soon as ${\gamma }_{co}$ exceeded the site-specific ${TR\gamma }_{co}$, and ${TR}_{temp}$ and ${TR}_{WS}$ were reached, the PO detection algorithm detected PO on DOY 142. The ${TH\gamma }_{co}$ for FYI was 1.3 dB. The $W_{refl}$ validation window (Supplementary Table S1) confirmed that the DOY 142 fell between the window, where we indicated late winter PO reflectance (0.8) on DOY 141 and post-PO reflectance (0.4) on DOY 149, ending the PO window, to validate our PO identification (Figure 5 and Figure 6). Based on the time series presented in Figure 5, both ${\sigma }_{VV}^0$ and ${\sigma }_{HH}^0$ exhibited a decline starting around DOY 170, indicative of the initiation of pond drainage, aligning with the observations regarding the microwave backscatter response to surface water drainage in sea ice during the advanced melt stage [61,62]. Site FYI_26_2017 showed a rapid change in temperature and, thus, for this site, PO started shortly (within three days) after MO. Following MO, there was notable variability in ${\sigma }_{VV}^0$, ${\sigma }_{HH}^0$, and ${\gamma }_{co}$, aligning with the suspected rise in snow surface wetness, increase in brine volume in basal snow layers, a shift from pendular to funicular regimes, pond initiation on DOY 142, and eventual pond drainage, which started around DOY 170. Sites FYI_24_2017 to FYI_ 31_2017 showed similar trends (Supplementary Table S1; Figure 1a).

For Site FYI_9_2018, the April ${\gamma }_{co}$ mean was −0.58 ± 0.70 dB and the ${\sigma }_{VV}^0$ and ${\sigma }_{HH}^0$ values ranged from −17 to −19 dB (Figure 7). The high backscatter values in April indicated that the site may have a rougher surface [36,37]. As the temperature started to gradually increase from −6 °C near DOY 140, it initiated early melt. As the temperature rose to −2 °C, an upturn from winter backscatter (3.5 dB) on DOY 146 indicated MO. We interpreted it as a combination effect from the previously described surface and volume scattering that occurred during MO, which resulted in stronger backscatter. To summarize, warming air and snow temperatures caused the basal snow layer brine volume to increase [55]. As a result, snow grains enlarged and could become brine-coated, acting as strong scatters at the C-band [41,63].

Both ${\sigma }_{VV}^0$ and ${\sigma }_{HH}^0$ increased, which resulted in an increase in ${\gamma }_{co}$. Here, we observed ${\sigma }_{VV}^0$ > ${\sigma }_{HH}^0$, which initiated the change in values of ${\gamma }_{co}$. When ${\gamma }_{co}$ exceeded the site-specific ${TR\gamma }_{co}$, ${TR}_{temp}$> 0 °C, and the daily wind speed reached ≥ 3 ms⁻¹, the algorithm detected the PO date on DOY 157. On this day, the mean ${\gamma }_{co\left(PO\right)}$ was 2.24 dB, exceeding the estimated ${TH\gamma }_{co}$ of 1.6 dB. During PO, the mean air temperature reached 1.1 °C and the wind speed was 8.7 ms⁻¹. The $W_{refl}$ window indicated the last late winter reflectance (0.85) on DOY 151 and the first post-PO reflectance (0.43) on DOY 180, validating our PO identification (Figure 7 and Figure 8). Sites FYI 10, 13, 17, and 19 (Supplementary Table S1; Figure 1b) of this study exhibited similar trends in 2018.

The difference between FYI_26_2017 and FYI_9_2018 appeared to be primarily influenced by the rate of the 2 m air temperature rise. FYI site 26 showed a rapid increase in temperature, causing an increase in ${\sigma }_{VV}^0$ and ${\sigma }_{HH}^0$ and causing PO detection soon after MO. In contrast, for FYI_9_2018, we found a gradual change in temperature and a gradual change in ${\gamma }_{co}$ of nearly six days after MO before detecting PO. The mean ${TH\gamma }_{co}$ change was also larger (by 0.3 dB) in FYI_9_2018.

4.3.2. MYI

The PO detection algorithm worked with higher accuracy to detect PO on MYI than on FYI (Table 1). During winter, the time series of MYI ${\gamma }_{co}$ was relatively stable, with little variability within the MYI sites. We analyzed two sample sites from 2017 and 2018, respectively, along with visual representation of optical imagery from Sentinel-2.

The April (DOY 90–120) ${\gamma }_{co}$ for site MYI_11_2017 was –0.19 ± 0.37 dB (Figure 9) and the ${\sigma }_{VV}^0$ and ${\sigma }_{HH}^0$ values were < −9.8 dB. MYI backscatter was much higher than that for FYI in winter because the snow on MYI was nearly transparent to C-band microwaves, and air bubbles trapped within MYI produced significant volume scattering to enhance the rough surface scattering of MYI [53,64]. We did not see any significant changes in backscatter during the early melt period. However, a small rise in backscatter was observed after a dip on DOY 143. This dip occurred due to a sudden temperature rise on DOY 140 to near 0 °C, which caused microwave absorption due to sub-surface snow melting. This also caused ${\gamma }_{co}$ to decrease. The dip caused an increase (3.8 dB) in ${\sigma }_{VV}^0$ compared to ${\sigma }_{HH}^0$ prior to MO in time series (Figure 9), which indicated that temperature-induced backscatter changes were more visible in ${\sigma }_{VV}^0$ than ${\sigma }_{HH}^0$.

MO occurred on DOY 162 with increasing temperature, where ${TR}_{temp}$ > 0 °C, and ${\gamma }_{co}$ exceeded ${TR\gamma }_{co}$ for the first time on DOY 168. The temperature after MO was sufficient; however, due to not having enough wind speed (1.6 ms⁻¹), the PO detection algorithm did not detect the first upturn, rather it detected the PO on DOY 170, where both ${TR\gamma }_{co}$ and ${TR}_{WS}$ > 3 ms⁻¹ conditions were met. ${TH\gamma }_{co}$ was 1.90 dB, which is larger than the FYI sites discussed. The PO date 170 fell within the validation window, $W_{refl}$ from 167 (0.8 reflectance value) to 181 (0.3 reflectance value), confirming a reasonable PO detection date for this site (Figure 9 and Figure 10).

The April (90–120) ${\gamma }_{co}$ for site MYI_13_2018 was –0.13 ± 0.38 dB (Figure 11) and the ${\sigma }_{VV}^0$ and ${\sigma }_{HH}^0$ values were approximately −11 dB. Backscatter values were consistently high through the pre-melt period. Diffuse surface scattering was much more dominant in MYI due to its rougher surface compared to FYI, which was dominated by specular surface scattering [16].

MO occurred on DOY 154 due to the increasing air temperature. The air temperature slightly increased around DOY 157 to 1 °C, which increased the moisture in the snow cover, causing a greater decrease in ${\sigma }_{HH}^0$ compared to ${\sigma }_{VV}^0$. As a result, ${\gamma }_{co}$ became more positive. With rapid warming, ${\gamma }_{co}$ exceeded the ${TR\gamma }_{co}$ on DOY 158, where the ${\gamma }_{co\left(po\right)}$ was 1.01 dB, marking this date as PO. The wind speed was 4.9 ms⁻¹ and the air temperature was 0.9 °C. The small validation window, from 157 (0.78) to 164 (0.5), supported the presence of melt ponds at the site on this date (Figure 11 and Figure 12).

Compared to the two FYI sites discussed, the temperature evolution of MYI ${\sigma }_{VV}^0$ and ${\sigma }_{HH}^0$ remained very stable after reaching 0 °C, helping to moderate melt and backscatter after PO. The difference between the two MYI sites was similar to that for FYI. Variability of PO date occurrence across the ice types and regions is discussed in the next section.

4.4. Spatial and Temporal Variability of PO Date (2017 and 2018)

Figure 13 illustrates the temporal and regional variability of PO dates. The mean PO date for FYI throughout the CAA in 2017 was 173 ± 22. For the FYI sites in 2017, the mean PO was estimated to occur on DOY 174 ± 3 for the Parry Channel. Only one validated FYI site was available in the northern CAA, and this had a PO date of 176. The mean PO date was 147 ± 4 near Cambridge Bay and Victoria Strait, which is 27 days earlier than in the Parry Channel. It is expected that PO dates occur first in the south and then gradually proceed to the north (Figure 13). The southern sites in the CAA had earlier PO dates because of the temperature and latitudinal influence, and the results showed regional spatial variability in the occurrence of PO dates in 2017 throughout the CAA. In contrast to 2017, the mean PO date for FYI in 2018 occurred on DOY 164 ± 4 throughout the CAA, with less spatial variability. Therefore, near-homogeneous and widespread early PO occurred over the CAA in 2018.

We saw a similar trend of MYI PO date occurrence throughout the CAA. The mean PO was estimated to occur on DOY 172 ± 2 throughout the Queen Elizabeth Islands in 2017. At the Parry Channel, the mean PO date over MYI was on DOY 170 ± 2. The mean PO date was observed three weeks earlier near Cambridge Bay and Victoria Strait on DOY 145 ± 4, compared to the northern CAA. The mean PO date over MYI sites in 2018 was DOY 166 ± 4. The air temperature evolution from winter to the melt period was rapid in 2017 compared to 2018. This initiated earlier melting of MYI, and the transition from MO to PO was also shorter in 2017 compared to 2018. Both FYI and MYI ice types showed similar PO date progression in each year but different interannual trends. This indicated that site-to-site variability of PO depends largely on climatic parameters (air temperature, snow thickness, and wind speed conditions).

According to [52], in years of faster progression of surface air temperature from winter to the melt period, the melt is advanced, resulting in earlier MO, and vice versa. The liquid moisture slowly accumulates in snowpacks, resulting in a longer time to appear on the ice surface, which causes delayed MO signal detection in backscatter. We saw similar trends in 2017 and 2018 for both ice types (Figure 13), which explains our PO date occurrence variability in the ice types and years.

4.5. Limitations

For some of the test sites in northern CAA, the PO detection algorithm was unable to detect the PO date on very smooth FYI due to not fulfilling the criteria. Figure 14 shows FYI_4_2018, where the PO algorithm could not detect the PO date. Upon additional investigation, we observed a significant noise component in the ASCAT data caused by selecting very smooth FYI. It is notable that ASCAT backscatter for this site did not go below −20 dB due to substantial land and/or MYI contamination that made it into each ASCAT point measurement. Depending on the antenna direction and the orientation of the three beams of ASCAT, the backscatter from as far away as 30 km can contribute to this noise. Although, using three beams of ASCAT reduced the radiometric noise, which increased the chances of mis-selecting the ice type using our simple thresholding approach. For FYI, the PO detection algorithm could not identify the PO date for three sites in 2017 and seven sites in 2018. For these reasons, we found that using FYI ${\sigma }_{HH}^0$ > −20 dB would greatly reduce noise but would limit PO detection from very smooth FYI.

The applicability of this research is limited to Arctic landfast sea ice types. The PO detection techniques developed for Arctic landfast sea ice may not be directly applicable to moving ice due to significant differences in their behavior. Landfast ice is stationary, and it changes predictably, allowing for consistent observation points and straightforward detection of melt ponds through time series analysis. In contrast, moving ice is dynamic, subject to continuous drift and deformation due to wind, ocean currents, and thermal effects, leading to rapid and unpredictable changes. To adapt the PO detection algorithm for moving ice, it must include motion-tracking capabilities, higher temporal and spatial resolution, and adaptive thresholds. In addition, accurate validation requires real-time data and techniques to handle the spatial and temporal variability of moving ice. These enhancements are essential to effectively monitor PO in the complex and dynamic environment of moving ice in Arctic regions.

5. Conclusions

In this paper, we presented a novel sea ice melt pond onset detection method utilizing a C-band multi-sensor approach from Sentinel-1 and ASCAT. Our investigation centered on the utility of the co-pol ratio threshold method for detecting pond onset (PO) dates over the Canadian Arctic Archipelago (CAA) during the years 2017 and 2018. This method incorporated factors such as wind speed and air temperature, further validated by Landsat 8 reflectance data in conjunction with available Sentinel-2 RGB imagery. Our findings demonstrated the efficacy of the co-pol ratio in predicting PO dates across the study region, showcasing accuracy amidst spatial and temporal variability.

Our study assessed site-specific time series of ${\gamma }_{co}$ in the context of PO detection on sea ice. Different addends ($2{\gamma }_{co(w.std)}$, $2.5{\gamma }_{co(w.std)}$, and $3{\gamma }_{co(w.std)}$) were tested against site-specific ${\gamma }_{co(w.mean)}$ for the years 2017 and 2018. Results indicated that using $3{\gamma }_{co(w.std)}$) led to fewer false PO detections compared to the other multipliers, with detection percentages of 70.59% and 92.3% for FYI and MYI, respectively, in 2017. The same technique was applied to 2018 data, and the detection rate was 61.53% for FYI and 96% for MYI. Site-to-site variability in ${\gamma }_{co(w.mean)}$ was higher for FYI than MYI, making FYI PO detection less robust. Four case studies were presented. The observed differences between 2017 and 2018 highlighted the influence of meteorological parameters on PO occurrence, with faster air temperature progression leading to earlier melt onset (MO) and subsequent PO dates. MYI exhibited more stable temperature evolution compared to FYI, influencing melt dynamics and increasing the PO detection accuracy. We also presented a case study where the PO detection algorithm encountered challenges in identifying PO dates on very smooth FYI in northern CAA test sites due to significant noise in ASCAT data, particularly when FYI ${\sigma }_{HH}^0$ < −20 dB. This noise, caused by substantial land and/or MYI contamination within ASCAT point measurements, emphasized the need for refined thresholding approaches to enhance detection accuracy while accounting for data noise limitations.

Finally, the dual-platform approach adopted in this study successfully mitigated uncertainties regarding spatial and temporal resolution. However, we acknowledge the potential for further enhancement, particularly in the context of a larger region, such as the CAA. Therefore, upcoming studies should focus on incorporating snow thickness and pond fraction from post-MO to post-PO in model input and testing this algorithm over different ice types in regions for a similar period in a year in CAA. The overall performance of this PO detection method provided accurate results, confirmed by validation from optical satellite data in the Canadian Arctic Archipelago. Moreover, the methodology shows potential for mapping and detection of melt pond spatial and temporal variability on a pan-Arctic scale by using the co-pol ratio. With the launch of the RADARSAT Constellation Mission creating more availability of SAR data, a multi-sensor co-polarization approach may offer the most robust approach to detect PO over sea ice across the Arctic. The method has been tested on C-band SAR; however, it could also be applied using Ku- or L-band SAR. Such advancements hold promise for progressing our understanding of sea ice thermodynamic and dynamic processes and their implications in the context of Arctic climate change.