Two automated weather stations located in the La Joie Basin are operated by the provincial hydroelectric utility BC Hydro (Fig. 1). The Downton Lake Upper automated weather station (1829 ma.s.l.) and Green Mountain automated weather station (1725 ma.s.l.) provide hourly records of temperature, precipitation, and SWE. Downton Lake Upper is situated on a northwest-facing slope and is classified as disturbed forest (Fig. 1). Green Mountain is situated on a northeast-facing slope and is classified as mature forest (Fig. 1).

A 5 m bare-earth lidar DEM was used to create slope, aspect, and elevation classes in the La Joie Basin. Lidar data were collected on 22 September 2017 from a fixed-wing aircraft using methods described elsewhere (Pelto et al., 2019). The DEM was resampled to 30 m and classified into 4 aspect classes, 6 slope classes, and 20 elevation classes.

The Sentinel-1 constellation acquires SAR scenes every 2 to 3 d over the La Joie Basin. The constellation consists of two satellites orbiting the Earth in ascending and descending directions. Each satellite has two tracks that capture the La Joie Basin. In the descending direction both tracks (13 and 86) completely capture the La Joie Basin; however, in the ascending direction only one track (137) has full coverage, while the other (64) only captures the eastern side of the basin (Table 1). Local incidence angles vary slightly by satellite track (Table 1). On average, incidence angles are higher for tracks in the descending direction (Table 1). Local incidence angles in descending Sentinel-1 images at Downton Lake Upper measure between 33 and 42 ∘, whereas at Green Mountain they measure between 26 and 34∘.

Images over the watershed are acquired in the interferometric wide-swath mode, the default collection mode over land. These SAR scenes have a resolution of 10 m and have two polarization bands available, VV (vertical–vertical) and VH (vertical–horizontal).

The Level-1 ground-range-detected (GRD) Sentinel-1 SAR dataset from Google Earth Engine (GEE) is pre-processed using the Sentinel-1 Toolbox (European Space Agency, 2021). The images undergo five steps during correction:

application of the precise orbital file,

The Committee on Earth Observation Society Analysis Ready Data for Land (CEOS-ARD) identify additional steps needed for SAR pre-processing (Lewis et al., 2018) that are not included in the Sentinel-1 GEE catalogue. These steps include radiometric terrain correction and pixel-based identification of shadow or layover (Lewis et al., 2018).

The operational land imager and thermal infrared sensor aboard Landsat-8 from the National Aeronautics and Space Administration acquire imagery over the La Joie Basin every 16 d. Images are provided at a 30 m resolution, and bands provide coverage across the visible, near-infrared, and shortwave-infrared spectrums. Top-of-atmosphere (TOA) reflectance Landsat-8 scenes are publicly available in the GEE data catalogue.

The multispectral instrument aboard Sentinel-2 from the European Space Agency collects multispectral imagery over the La Joie Basin every 6 d. Images are provided at 20 m resolution, and bands provide coverage across the visible, red-edge, near-infrared, shortwave-infrared, and thermal spectrums. TOA reflectance Sentinel-2 scenes are freely available in the GEE data catalogue.

Dates of snowpack saturation from 2018 to 2021 were estimated from Sentinel-1 SAR images. GEE was used to access, correct, and download Sentinel-1 SAR scenes. To detect the onset of snowmelt in the basin, we restricted our analysis for SAR scenes acquired between February and August. We created eight image collections from Sentinel-1 SAR scenes, with two collections created for each study year based on the directional pass of the satellite (ascending or descending) (Table 2). To prevent border noise from impacting results, we only selected images containing the whole basin (i.e., those captured by tracks 13, 86, and 137) to create the stacks. We corrected the Sentinel-1 image collections for slope-induced radiometric distortion using a volumetric model (Vollrath et al., 2020) and applied a filter to reduce radar-induced speckle (Lopes et al., 1990). After corrections, we resampled the image collections to 30 m for consistent analysis with optical imagery and downloaded them. We performed these steps for images captured in both VV and VH polarizations.

We used RStudio, an integrated development environment for R (RStudio Team, 2021), to estimate snowpack saturation from the created Sentinel-1 image collections. Before we extracted saturation dates, we temporally smoothed time series of backscatter values in each pixel using a locally weighted least-squares regression with a span of 0.2 (Supplement). This correction reduces noise in SAR time series and thus minimizes false positives in the final saturation date maps. After smoothing, we identified the date of the minimum backscatter value for each pixel.

We used backscatter minima to approximate the date of snowpack saturation (Marin et al., 2020) for all pixels in the basin, regardless of their snow-covered status. After we extracted saturation dates, we calculated mean values for each elevation, slope, aspect, and land cover class. To correct for any remaining artefacts of SAR speckle and distortion, we removed any pixels with estimated saturation dates that were outside of 2.5 standard deviations of the calculated means. We then infilled saturation dates with the median value from that pixel's elevation, slope, aspect, and land cover class (Table 3).

To further validate the behaviour of SAR signals over complex terrain, we focus on SAR time series extracted from pixels between elevations of 1600 and 1800 m and average them by land cover type and slope. We also extracted SAR time series at the Green Mountain and Downton Lake Upper automated weather stations, where continuous measurements of SWE are available. To create SAR signals at the location of each automated weather station, we used the average pixel values from all pixels located within a 40 m circular radius of each station. All SAR time series were smoothed with a LOESS regression with a span of 0.2.

We determined snow disappearance from Sentinel-2 and Landsat-8 images in GEE. We filtered Sentinel-2 and Landsat-8 scenes to create collections of images containing the La Joie Basin captured between February and September from 2018 to 2021 (Table 4). Images were further filtered to remove scenes with more than 60 % cloud cover, resulting in the use of 53 %, 51 %, 56 %, and 55 % of all available images each year from 2018 to 2021, respectively. We masked images for clouds and combined them into a single image collection for each study year (Table 4).

We utilized a hybrid approach to approximate snow disappearance from the multi-source images. Over non-glacierized areas, we calculated Normalized Difference Snow Index (NDSI; Hall et al., 1995) and Normalized Difference Forest Snow Index (NDFSI; Wang et al., 2015) values to classify each image for snow cover using a threshold of 0.4 (Table 5). For glacierized areas of the La Joie Basin, we used a k-means clustering algorithm to classify the initial image collection for snow cover. We created two classes on the glacier, snow and ice, based on reflectance values in the visible, near-infrared, and shortwave-infrared spectrums (Shea et al., 2013).

For all areas in the basin, we selected the first snow-free date per pixel and extracted the corresponding date. We then exported the maps of snow disappearance, or snowmelt end, from GEE and calculated mean values for each elevation, slope, aspect, and land cover class. To correct for errors due to clouds and cloud masking, we removed any pixels with estimated snow disappearance dates that were outside of 2.5 standard deviations of the calculated means. We then infilled snow disappearance dates with the median value from that pixel's elevation, slope, aspect, and land cover class (Table 3).

We approximated snowmelt duration from the created maps of snowpack saturation and snowmelt end. We took the difference in days between disappearance and saturation estimates as the duration of melt production for each study year. Elevation-based thresholds were used to identify improbable melt duration estimates, such as negative estimates or estimates exceeding 120 d. Improbable melt duration estimates were infilled with the median value from that pixel's elevation, slope, aspect, and land cover class.

To validate SAR time series, estimates of snowmelt onset, and estimates of snowmelt duration we used continuous SWE records from the Downton Lake Upper and Green Mountain automated weather stations. We averaged hourly SWE measurements for each day and calculated maximum SWE for each snow season between 2018 and 2021. We used a piecewise linear regression to infer the period of melt at each station. Regression models were determined using the “segmented” package (Muggeo, 2017), which implements a bootstrap restarting algorithm (Wood, 2001) to estimate breakpoints in the time series. We used the first breakpoint after maximum SWE to approximate melt onset, and we used the last breakpoint to approximate the end of melt (Fig. 2).

Estimates of snowmelt onset and duration from the linear piecewise regression were consistent during the study period. At Downton Lake Upper, snowmelt duration ranged from 40 to 51 d among study years, with the shortest duration observed in 2019. Snowmelt onset dates were consistent at Downton Lake Upper, ranging from 4 to 10 May, with the earliest onset recorded in 2018. At Green Mountain, snowmelt duration ranged from 27 to 35 d among study years, with the shortest duration estimated in 2021. Estimates of melt onset at Green Mountain ranged from 29 April to 13 May, with the earliest onset observed in 2018.

At automated weather stations, the minimum value in each SAR time series coincided with rapid (10–20 mmd-1) declines in snow water equivalence (Figs. 3 and 4). At Downton Lake Upper, SAR time series minima occurred within 0–10 d of SWE melt onset estimates for both polarizations (Fig. 3). Comparing across four melt seasons at Downton Lake Upper, the maximum difference (+4 d) between onset estimates from VV polarized time series and telemetry records was produced in 2019. The minimum difference was produced in 2018 when there was no difference between telemetry records of melt onset and SAR estimates of melt onset. The maximum difference (-10 d) between estimates from VH polarized time series and telemetry records was produced in 2020 at Downton Lake Upper. The minimum difference was produced in 2018, when (similar to the VV time series) there was no difference between telemetry records of melt onset and SAR estimates of melt onset. At Green Mountain, SAR minima occurred within 1–13 d of SWE melt onset estimates (Fig. 4). Among study years, the maximum difference (+11 d) between onset estimates from VV polarized time series and telemetry records was produced in 2019. The minimum difference was produced in 2019 (-1 d). The maximum difference (-13 d) between onset estimates from VH polarized time series and telemetry records was produced in 2020 at Green Mountain. The minimum difference was produced in 2021, when there was no difference between telemetry records of melt onset and SAR estimates of melt onset. Minima from VV polarized images produced more accurate approximations of melt onset compared to those produced from VH polarized images at both stations. At Downton Lake Upper, all VV-derived minima occurred within 7 d of SWE melt onset estimates. At this location, only 50 % of minima from VH polarized images occurred within 7 d of melt onset. At Green Mountain, 75 % of VV-derived minima occurred within 7 d of SWE melt onset estimates. Similar to the results at Downton Lake Upper, only 50 % of minima at Green Mountain from VH polarized images occurred within 7 d of melt onset.

SAR backscatter time series from across the basin have a characteristic “U” shape, with their lowest values occurring in late spring (Fig. 5). In open areas, the seasonal decrease is most pronounced on gentle slopes (< 10∘) and decreases in amplitude as slope increases.In forested regions of the La Joie Basin, the seasonal decrease in backscatter showed the greatest amplitude on steep slopes (40–49∘). Among land cover classes, SAR minima are the most pronounced over bare earth and become harder to detect under dense vegetation, especially on moderate (10–29∘) slopes.

Due to the lower frequency of images in the ascending direction, results are shown for imagery captured by the descending pass of the satellite only. As time series from VH polarized images produced less reliable estimates of melt onset at Green Mountain and Downton Lake Upper, only results from VV polarized images are shown. For results from VH polarized images, see the Supplement. From 2018 to 2021 in the La Joie Basin, onset estimates were consistent by elevation with slight interannual variability (Fig. 6).

Generally, SAR estimates of melt onset indicate that snowmelt initiates in early March at low elevations in the La Joie Basin, with later snowmelt onset dates observed at higher elevations. The highest-elevation pixels are estimated to initiate melt in mid-July. When examined by elevation, 2018 showed the earliest average melt onset dates when compared to other study years, with low elevations estimated to initiate melt as early as 15 March. At low elevations, pixels were estimated to initiate melt the latest in 2019 when compared to other study years, with estimates averaging as late as 7 April. The largest range in mean melt onset estimates was produced in 2018, with 97 d between the earliest and latest elevation-binned averages. The smallest range was produced in 2019, with 83 d between the earliest and latest elevation-binned averages. At high elevations, the latest average melt onset dates were estimated in 2020, with pixels initiating melt on 5 July. The earliest average melt onset dates at high elevations were estimated in 2018, with the highest-elevation pixels initiating melt on 20 June. Melt onset estimates across elevations were influenced by aspect, slope, and land cover (Fig. 7). From 2018 to 2021, south- and west-facing slopes initiated melt earlier than north- and east-facing slopes. SAR melt onset estimates indicate that melt initiates earlier in the La Joie Basin on steep slopes (40–49∘) and initiates progressively later as slope decreases. Melt onset dates are similar between land cover types, with the exception of seasonal snow on glaciers, which initiated later than other classes in all study years.

Snow-free dates estimated from optical and multispectral imagery (Fig. 8) show a similar elevation dependence to melt onset. However, we observe a greater variability in the snow-free dates in forested regions and at lower elevations.

The detection of snow-free dates from Landsat-8 and Sentinel-2 images provided consistent estimates by aspect, elevation, land cover type, and slope (see the Supplement). The impact of aspect is most pronounced at upper elevations, with northern and eastern slopes melting later than southern and western slopes. Similar to SAR estimates of melt onset, snow disappearance estimates were the most variable on steep slopes (> 30∘). When compared to the other land classes, snow cover on glaciers disappears much later in the ablation season. Trends in snow-free dates are consistent among study years.

With SAR-derived estimates of melt onset and optical and multispectral estimates of snow disappearance, we map snowmelt duration in the La Joie Basin (Fig. 9).

SAR snowmelt duration estimates formed consistent patterns based on elevation (Fig. 10). Snowmelt duration estimates were the longest between elevations of 2200 and 2400 ma.s.l. in the La Joie Basin from 2018 to 2021. Between 2200 and 2400 ma.s.l. the snowpack produced runoff for 46 d on average, whereas the basin-wide average was 30 d. At elevations above 2400 ma.s.l. duration estimates shorten to an average of 38 d. Below 1200 ma.s.l., estimated snowmelt duration is the shortest, lasting 14 d on average. Among study years, elevation-averaged snowmelt duration estimates were the shortest in 2021 at 27 d on average and the longest in 2018 at 32 d on average.

Snowmelt duration is more sensitive to land cover, slope, and aspect when compared to melt onset dates, especially at elevations of greater than 1400 ma.s.l. Melt durations were longer on north- and east-facing slopes compared to southern and western slopes. Increasing slope duration decreases melt duration in the La Joie Basin, with the longest durations observed in relatively flat areas (0–9∘) above 2000 ma.s.l. Land cover strongly influences melt duration. On average, from 2018 to 2021 snowmelt duration estimates were the longest over glacierized terrain, followed by mature forest. Snowmelt duration estimates were the shortest over disturbed or immature forest, followed by bare earth.

Multi-source snowmelt duration estimates did not always reflect melt duration from SWE records at automated weather stations (Tables 6 and 7). Optical-radar melt duration ranged from 19 to 38 d at Downton Lake Upper and from 14 to 36 d at Green Mountain. Higher inaccuracies are seen in the optical estimates of snow-free dates when compared to SAR estimates of snowmelt onset.

While VV polarized time series produced more accurate estimates of melt compared to VH polarized time series at automated weather stations, increased noise in VV time series is observed in the La Joie Basin (Fig. 5). The amplified noise in the VV time series can be attributed to the steep slopes and forest cover present in the La Joie Basin. Vegetation covers a significant proportion (∼ 45 %) of the total area of the La Joie Basin. Manickam and Barros (2020) report that a common thresholding approach for mapping wet snow using co-polarized SAR images (Nagler and Rott, 2000) failed in Colorado, USA, due to the presence of conifer forest. Backscatter coefficients from cross-polarization were more sensitive to snow cover below vegetation compared to those from co-polarization and enabled wet snow mapping below the treeline (Manickam and Barros, 2020). Furthermore, the high relief in our study area may amplify the observed noise in co-polarized images, as steep slopes reduce the ability of VV polarization to distinguish between wet snow and snow-free surfaces (Nagler et al., 2016). The combination of steep slopes and forested regions present in the La Joie Basin increase noise in VV time series; however, VV polarized SAR images are shown to be more sensitive to wet snow compared to VH polarized images (Nagler et al., 2016). The increased sensitivity to wet snow in VV polarized images is reflected in the La Joie Basin through the greater accuracy of VV time series for approximating snowmelt onset at Downton Lake Upper and Green Mountain.

The U shape of SAR time series is more pronounced in open areas compared to mature forest (Fig. 5). As a result, estimates of melt onset and duration were less reliable in forested areas. The decreased sensitivity in forested areas can be attributed to the scattering of SAR signals by the forest canopy. Forests effectively scatter radar energy depending on their structure, composition, and stem density (Bernier, 1987). Conifers, in particular, strongly scatter C-band SAR signals (Bernier, 1987). Spruce and fir trees (conifers) dominate forests in the La Joie Basin, leading to increased noise in melt onset maps from Sentinel-1 data. The locally weighted least-squares regression temporal smoother helped to reduce the variability of melt onset estimates in forested regions; however, ground truthing in these regions is required to verify the validity of this shift.

In open areas, the U shape of SAR time series is more pronounced on flat slopes when compared to steep slopes (Fig. 5). Manickam and Barros (2020) presented similar findings in the Swiss Alps, where between April and May backscattering coefficients were less sensitive to wet snow as slope increased. In forested regions, however, slope has a less uniform effect on SAR signals. The most distinct U-shaped signals are created from SAR time series collected on steep (> 30∘) slopes and the least distinct are collected on moderate (10–29∘) slopes. Scatter from forests is impacted by terrain, which induces changes in trunk–ground and crown–ground scattering mechanisms (Park et al., 2012). Vegetative scattering mechanisms can create more negative values based on ground surface tilt (Park et al., 2012). In the La Joie Basin, the increased visibility of SAR minima on steep slopes in forested regions may be related to terrain-induced changes in vegetative scattering mechanisms.

Optical-radar estimates of melt onset and duration were impacted by interannual climate fluctuations. At Downton Lake Upper, SAR estimates of melt onset from VV times series from 2018 to 2021 were within 0–4 d of telemetry estimates. For VH polarized time series, melt onset estimates were within 0–4 d of telemetry estimates in 2018 and 2021 only, with minima occurring 10 and 9 d early in 2019 and 2020, respectively. At Green Mountain, VH estimates of melt onset were also early in 2020, with minima occurring 13 d before telemetry estimates. VV and VH estimates were late in 2019 at Green Mountain, with both minima occurring 11 d after telemetry estimates. In 2020, early estimates of melt onset at both stations may be the product of large temperature fluctuations in early April. Melt–refreeze cycles alter snowpack LWC and grain size (Yamaguchi et al., 2010), and they will thus impact SAR signals (Liu et al., 2006). The late and early melt onset estimates in 2019 are also a likely result of temperature fluctuations. In 2019, daytime and nighttime temperatures at the La Joie Basin exceeded 0 ∘C from 17 to 22 March, increasing snowpack LWC and thus the possibility for early melt onset estimates. However, after the initial warming period in 2019 temperatures decreased, with average nighttime temperatures of below 0 ∘C for the entire month of April. Snowpack refreezing can decrease seasonal differences in SAR backscatter values (Floricioiu and Rott, 2001), and, as a result, minima in SAR time series may become less pronounced during refreezing periods. Among study years, interannual variations in climate, and therefore snowpack metamorphosis, influenced the accuracy of SAR snowmelt retrieval in the La Joie Basin and can partially explain the variability in snowmelt onset at automated weather stations.

Melt onset and duration estimates displayed varying agreement with continuous SWE records at Downton Lake Upper and Green Mountain. While all VV SAR minima from the extracted time series were within 5 d of melt onset at Downton Lake Upper, they occurred up to 11 d apart at Green Mountain. Green Mountain is located on a northeast-facing slope and receives morning sunlight. For this analysis only descending images were used, which are captured around 07:00 PST (Pacific standard time) in the La Joie Basin. Descending images in the La Joie Basin may be more susceptible to false positives from morning wetting on east-facing slopes during the ripening stage. Morning wetting of the snowpack may attribute to the greater variability seen in Green Mountain time series when compared to Downton Lake Upper. At both locations, time series estimates of melt onset were improved in VV polarization when compared to VH polarization. Final melt onset maps, however, were more accurate from VH polarized melt onset estimates when compared to VV, with 75 % of VH estimates at Green Mountain occurring within 1 d of telemetry estimates (see the Supplement). Co-polarized SAR signals have shown greater sensitivity to wet snow when compared to cross-polarized images (Nagler et al., 2016); however, cross-polarized images are preferable when mapping wet snow below the treeline (Nagler et al., 2016; Manickam and Barros, 2020). Green Mountain is more heavily vegetated compared to Downton Lake Upper, which may result in the increased accuracy of VH estimates at this location. As the accuracy of VH observations at Green Mountain in final melt onset maps was not mirrored in time series estimates, increased physical observations of the snowpack are required to further quantify the impact of polarization when using SAR minima as proxies for snowmelt onset.

Inaccuracies in snowmelt duration is attributed to errors in SAR onset estimates and optical snow disappearance estimates. Among study years, SAR acquisitions are available every 5 to 12 d, averaging 6 d in each year. Optical observations are variable among study years (Table 4); however, average revisit intervals are consistent at 4 d in 2018 and 2019 and 3 d in 2020 and 2021. We calculated errors in snowmelt duration (σD) using Eq. (1). 1σD=σONS2+σEND2, where error in the onset (σONS) and disappearance dates (σEND) are taken as the average revisit interval for each data type and study year. Rounding to the nearest day, this yields an average error in the duration estimates of ± 7 d for all study years. Ideally, a revisit time for SAR observations of 3 d would reduce this error to ± 4 d. Daily SAR observations could reduce error in duration estimates to ± 3 d; however, high-resolution optical imagery is needed at more frequent revisit intervals for reduced error or operational applications.

Snowpack monitoring with optical remote sensing data is challenging in the La Joie Basin, as imagery is frequently obscured by cloud cover. Cloud cover is persistent at high elevations during the ablation season in the La Joie Basin. Missing data from cloud cover reduce the temporal resolution at which imagery is available, decreasing the accuracy of snow disappearance estimates. The inclusion of optical data with a more frequent revisit interval or the use of complex snow detection algorithms could improve this analysis. Support vector machine algorithms have yielded promising results for multi-source snow detection with optical and SAR data in mountainous environments (He et al., 2015; Lui et al., 2020). Further, downscaling and gap-filling algorithms can improve snow cover detection using multi-source optical data (Premier et al., 2021; Revuelto et al., 2021). Revuelto et al. (2021) produced 20 m snow cover products by downscaling observations from the high-frequency, low-resolution Moderate Resolution Imaging Spectroradiometer with Sentinel-2 observations. In future studies, downscaling or machine-learning methods can be adopted to reduce error in snow disappearance estimates.

Despite errors in snow-free estimates, the observed trends in snowmelt duration agree with trends in snow depths observed in the study area (Sergio Tagle and Brian Menounos, personal communication, 2021). The snowpack is generally deepest at middle to high elevations (2200 to 2400 ma.s.l.) in the La Joie Basin, as at the highest elevations there is heavy redistribution of snow by wind and avalanching.

SAR retrievals of snowmelt onset and duration rely on the high resolution and revisit frequency of Sentinel-1. On 23 December 2021, Sentinel-1B malfunctioned and has not been communicating data since (European Space Agency, 2022). The loss of Sentinel-1B reduces the number of observations available for snowpack monitoring, which means that there may not be sufficient data to capture melt onset until the launch of Sentinel-1C in 2023. Alternative C-Band SAR datasets can be explored for snowmelt monitoring, such as the RADARSAT Constellation Mission.