HMA contains several mountain ranges – the Himalaya, Pamir, Karakoram, Hindu Kush, Tian Shan, and Kunlun Shan – from which flow several large rivers serving more than a dozen countries (Fig. ). Many of these catchments, such as the Tibetan Plateau, Tarim, Syr Darya, Amu Darya, and Indus, rely on snowmelt for more than 50 % of their yearly water budget . Many communities – particularly those at high elevations or those that depend on surface water for agriculture – are highly reliant on the timing of snowmelt. An early snowmelt season can create a late-season “water gap” when a dry spell is caused by snow meltwater disappearing before the start of the next rainy season. These water gaps can also negatively impact flora and fauna which depend heavily on the timing of the appearance of ephemeral water bodies . The timing and volume of snowmelt thus have important implications for the environment, direct household water use, agriculture, and hydropower.

Several interacting moisture sources, including the winter westerly disturbances (WWD), ISM, and East Asian summer monsoon (EASM), are responsible for the wide range of snowfall regimes across HMA (Fig. , inset). The interaction of these climatic regimes with the complex topography of HMA – particularly the vast elevation gradients – creates a diverse set of snowfall regimes .

We leverage a combined time series of SSMI (1987–2009), Special Sensor Microwave Imager/Sounder (SSMIS) (2008–2016), Advanced Microwave Scanning Radiometer – Earth Observing System (AMSR-E, 2002-2011), AMSR2 (2012–2016), and Global Precipitation Measurement (GPM, 2014-2016) data, processed to 0.25 decimal degree (dd) resolution by interpolating raw PM swath data at a series of point locations as described in (see Supplement Table S1 for a full dataset listing). In essence, we group all measurements within a 0.125 dd radius of each point on a 0.25 dd grid and generate a spatially weighted mean value for each swath at that point. The dataset is comprised of 6399 point locations, with on average 26 000 PM measurements each (long-term average of 2.4 measurements day-1 for 29 years, with more measurements during the 2002–2016 period).

PM measurements are converted to SWE using the Chang equation (Eq. ) , with modifications for non-SSMI platforms as proposed by and a constant snow density of 0.24 gcm-3 as proposed by .

SD[cm]=1.59[cmK-1]×(Tb18V-Tb36V)[K]

Studies have noted that SWE estimates from the Chang equation have high uncertainties e.g.,, particularly in dense forests. However, as much of our study area is non-forested – and we use SWE only as a rough estimate of snow volume – we choose to rely on the simple Chang equation rather than a more complex algorithm for SWE estimation.

As control data, we analyze Moderate Resolution Imaging Spectroradiometer (MODIS) percentage snow-covered area product MOD10C1 V6, 2001–2016; and High Asia Refined Analysis (HAR) surface temperature 2000–2014;. While these datasets only cover a subset of our study period, they are among the few independent control datasets available across the entire study area.

The shift from dry snow, which can be physically characterized as snow crystals in an air background, to wet snow, which replaces the air matrix with water, shifts the primary interaction between PM radiation and the snowpack from volumetric (dry snow) to surface (wet snow) scattering. These scattering changes are reflected in the Tb data and allow wet and dry snow to be differentiated, as the transition from dry to wet snow drastically increases the measured Tb – particularly in the scattering (Tb37V) channel. The XPGR, as originally described by , is defined as

XPGR=(Tb19H-Tb37V)/(Tb19H+Tb37V).

This algorithm takes advantage of both the channel difference between the Tb19 and Tb37 GHz channels as well as the depolarization effects of snowmelt, which increases the differences between the horizontally and vertically polarized channels . In the original application of the XPGR on the Greenland Ice Sheet, a static value of -0.025 was shown to indicate the presence of liquid water in the snowpack and hence used to separate the year into melting and non-melting phases . We find that for the majority of HMA, the -0.025 threshold is not effective in identifying the onset of snowmelt, as the XPGR–snowmelt relationship is highly variable through time and space.

We modify the XPGR method here to track maximum passive microwave signal separation – or the yearly maximum XPGR, referred to from here on as MXPGR. As the context of seasonal snowmelt in HMA is quite different from that of Greenland, and sufficiently long-term and spatially diverse in situ snowmelt data are lacking, we use the MXPGR as a proxy for snowmelt onset to track changes in the snowmelt season year over year. Thus, while we do not use the classical literature definition of snowmelt – presence of liquid water in the snowpack – we track a consistent metric related to physical snowpack changes that can be broadly interpreted as the onset of the snowmelt season.

However, the MXPGR is not effective for tracking the cessation of snowmelt. To track the end of snowmelt, we leverage two additional datasets: (1) the raw Tb37V time series, which rapidly increases as snowpack thins, and (2) a SWE time series calculated from the Tb19 and Tb37 GHz channels .

We first use a simple peak-finding algorithm, which identifies peaks as points which are larger than their two neighboring samples, to generate a list of potential peaks in the XPGR data. Next, we take the average XPGR value within ±2 days of each peak to determine not only the simple yearly maximum XPGR but also the highest and temporally widest peak in our XPGR data – termed here the MXPGR. We flag years which have multiple strong and temporally distinct XPGR peaks as unconstrained for snowmelt onset estimation, as the algorithm has trouble consistently identifying the MXPGR in these cases.

To determine the end of the snowmelt season, we choose either the date of the yearly maximum Tb37V value, which corresponds to the thinnest snowpack or to a “bare earth” signal, or the first date where 4 out of 5 days have been within 2 cm of the yearly SWE minimum. We choose the yearly SWE minimum instead of zero as our SWE threshold for snow clearance because some regions in HMA have a defined melt season but rarely reach zero SWE. This also helps control for uncertainty in shallow SWE measurements, as detecting shallow snow (<5 cm) with PM data is still difficult . A full description of our melt detection algorithm is available in the Supplement (Figs. S1–S4).

Unfortunately, large-scale and several-decade-long snowmelt onset and end date records are not available for HMA. Instead, we use HAR and MODIS data alongside a manually generated set of control dates for the snowmelt season, determined from the SWE, XPGR, and Tb37V signals by the researchers. We visually identified major peaks (MXPGR), as well as the cessation of snowmelt, by inspection of the time series. We chose a random sample of 25 point locations across our study area and identify snowmelt dates for each year of the time series (n=1400). We use the calculated length of the snowmelt period (days between the MXPGR and the end of the snowmelt season) as an additional control variable (n=700).

Hierarchical clustering is a method used to correlate time series data by intrinsic similarity , which has been used extensively in the environmental research community. We generate clusters from those time series which share the most temporal overlap, or where the periodicity of Tb values have the largest coherence regardless of their spatial correlation.

We choose the XPGR time series as our clustering variable because the XPGR is the most sensitive to melt dynamics, integrates multiple Tb frequencies, and is not sensitive to SWE calibration issues. To improve the robustness of our clustering, we combine the disparate single-instrument PM time series into a single coherent time series which leverages the full temporal extent of each dataset (1987–2016), using the following three steps. (1) We standardize the PM signals of the suite of instruments used in this study to a single set of dates, artificially created at daily resolution from the minimum and maximum dates across all satellite datasets, by resampling all individual satellite time series to a daily time step and dropping dates without data. (2) We homogenize the disparate PM time series based on the overlapping portions of individual satellite time series, using linear regression. The results of these regressions can be seen in Tables S2–S5, with an example regression at a single point shown in Fig. . (3) In order to reduce noise in our cluster analysis, we resample our merged XPGR time series to a 5-day temporal resolution (pentad).

Next, we normalize each merged pentad time series (1987–2016) to a Gaussian distribution, using a percentile mapping approach . We then estimate the Pearson correlation coefficient to classify regions of self-similarity in our XPGR time series . This method computes a Pearson correlation coefficient between each time series and, based on the resulting correlation matrix, computes a set of linkages using the angle between time series in vector space . We use the maximum distance (complete linkage) to split the linkage matrix, which is favorable because it ensures a minimum intra-cluster correlation. An average linkage scheme was tested and produced heterogeneous cluster sizes with outliers. We choose our cluster threshold from the hierarchical clustering dendrogram (Fig. S6), which maximizes cluster size while minimizing cluster internal diversity (Fig. S7). We emphasize that the correlation is based on the temporal co-evolution of the time series and is less sensitive to the relative magnitudes of peaks and troughs between data points. For an oscillating time series, the magnitude of the Pearson correlation coefficient is driven by the synchronization of peak timing, especially in normalized time series. The combination of several sensors in this study may impact the magnitudes of the resultant time series but will not have an outsized effect on the timing, and thus clustering, of our time series.

We apply our algorithm on a pixel-by-pixel and year-by-year basis to identify the onset of the snowmelt season – here proxied by the MXPGR – as well as the end of the snowmelt season. We also use the number of days between the MXPGR and the end of the snowmelt season to calculate the snowmelt period for each year. The long-term average snowmelt period is shown in Fig. .

The length of the snowmelt season varies significantly across HMA (Fig. ). In many low-elevation areas, such as the Ganges Plain, and low-SWE areas, such as the central Tarim Basin, the snowmelt period is very short. Higher-elevation zones, and in particular the Tibetan Plateau, see snowmelt periods of several months. While both elevation and the amount of SWE impact snowmelt, these are not the sole determinants of the length of the snowmelt season (Fig. ). Each of the major catchments (see Fig. ) has a unique MXPGR, snowmelt end date, and snowmelt-period distribution, based on the various climate and topographic forcings present in each catchment.

Cluster selection criteria can be seen in Figs. S6 and S7. We choose our dendrogram cutoff (distance threshold in vector space) based on a combination of the number of generated clusters, the internal variation within those clusters, and the average resultant cluster size. In our case, we choose a distance cutoff of 1 radian from the complete linkage matrix (minimum intra-cluster correlation 0.525), which results in 285 clusters (Fig. ).

While the hierarchical clusters are not based on any explicit spatial relationships, many of the clusters fall into spatially coherent groups. For example, the Pamir Knot and Tarim Basin both form large, coherent clusters based on the similarity of their snowfall and snowmelt patterns. The large number of small clusters throughout the Himalaya indicate that the region is not climatically uniform, and small-scale variations in topography and climate have strong impacts on the snowmelt regime.

As can be seen in Fig. , there exists significant interannual variation in the timing the snowmelt season. This is particularly true of areas impacted by the WWD, which often have multiple snowfall events starting in winter and lasting until spring . As one year may receive a small late-season storm, and thus see a maximum in the spring, while the next year may receive a large summer storm, and thus peak in the summer, analyzing trends at a single point in space is difficult.

To mitigate the influence of interannual variation in determining long-term trends in the timing of the snowmelt season, we group our data into self-similar clusters using hierarchical clustering. We do not filter our generated clusters based on size or self-similarity, as we do not use our clusters to generate a single averaged or representative time series for each cluster, as is often done in climate analyses. Due to interannual variations in SWE and the timing of the snowmelt season, fitting a linear regression through only 29 years of data does not provide statistically significant results for the majority of HMA. Instead, we use our clusters to group sets of algorithmically derived snowmelt dates and fit linear models on a cluster-by-cluster basis. By leveraging the snowmelt dates of a set of time series in parallel, we are able to identify statistically significant changes in the timing of the snowmelt season, as well as changes in the length of the snowmelt period (Fig. ). To reduce noise from low-SWE and very short snowmelt-period areas, we remove areas from the subsequent analyses with long-term average melt periods of less than 20 days. We also remove MXPGR dates that are flagged as unconstrained (when there are multiple candidate dates) to limit the impact of unreliable data on our analysis.

MXPGR is trending earlier (negative trend) in HMA outside of a small band running from the Karakoram through the interior Tibetan Plateau (Fig. a). In another snowmelt study leveraging SSMI and QuickSCAT data in HMA, find a similar distribution of positive and negative snowmelt onset trends. However, a direct comparison with their results is difficult due to differences in the temporal and spatial resolution of source data, filtering methods, and statistical treatment of SWE trends. Negative snowmelt onset trends have also been previously observed in Central Asia , the Himalaya , and the Tibetan Plateau .

A complex pattern of regionally increasing and decreasing spring snow depth in the Tibetan Plateau has been observed since the 1970s , which could help account for the mixed MXPGR trends observed in the Tibetan Interior. High-elevation zones in the upper Indus catchment, running from the Karakorum in a southeastward direction, have seen increased precipitation over the past decades due to increases in the strength of the WWD .

Temperatures in HMA are increasing faster than the global average and are likely the primary driver of the almost universal earlier snowmelt end dates as seen in Fig. b. Increased temperatures have likely both reduced overall SWE amounts, by causing more precipitation to fall as rain, and decreased SWE persistence into the spring and summer months. These changes have helped drive a 2–8 days decade-1 earlier end to the snowmelt season (Fig. b).

The length of the snowmelt season is shortening in much of HMA, with the exception of small areas in the Pamir, Tian Shan, and Karakoram regions (Fig. c). We attribute this to a combination of increased WWD storm intensity and increases in late-season storms, which could help extend the snowmelt season slightly later into the year . In general, however, the snowmelt season is shortening throughout HMA. Intensification of the spring runoff regime in HMA has been observed in both model and empirical data.

Not only are changes in the snowmelt regime spatially complex (e.g., Fig. ), but they exhibit distinct temporal heterogeneity as well.

Changes in MXPGR do not have a bias towards early or late onset snow regimes (Fig. a). The end of the snowmelt season is almost universally negative (earlier), excepting a few isolated areas in the Kunlun Shan (see Fig. ). The majority of locations show negative (shorter) trends in snowmelt period. Strong negative changes in the snowmelt period are biased towards areas with long melt seasons (120 days or more). This implies that high-elevation areas, such as the Tibetan Plateau, and high-SWE areas, such as the Karakoram, will see a relatively stronger compression in the length of the snowmelt season. While changes in the MXPGR date are partially responsible, the main driver of shorter snowmelt periods is the earlier end of the snowmelt season across most of HMA.

Several-decade-long trends conceal short-term fluctuations in the snowmelt regime of HMA. To assess the impact of the analysis time frame on our regression results, we analyzed trends with window sizes ranging from 4 to 28 years, across all possible start-year and window-size combinations, averaged over the entire study area (Fig. ).

Trends are universally negative for the MXPGR and the end of the snowmelt season, as well as for the snowmelt period, between 1988 and 1995, regardless of the time frame over which the regression is performed. While there were some short positive trends in snowmelt end date (5–10 years) starting in the mid 1990s, trends in end dates and snowmelt period are generally negative. Although long-term trends in MXPGR date (longer than 20 years) are negative, recent trends (after 2002) are positive when considered at time frames of 5–10 years. This implies that while the 3-decade trend in MXPGR dates has been negative, the trend has become more variable in the past decade.

It is clear that decadal trends (see Fig. ) are not consistent throughout the entire study period (see Fig. ). When trends in the first half (1988–2002) and second half (2002–2016) of the data are compared, distinct regional patterns are apparent (Fig. ).

The lack of statistically significant trends limits some interpretations, particularly with regards to changes in the snowmelt period. Nowhere in HMA are MXPGR trends consistent in both analysis periods. While many snowmelt end dates have remained negative in both time periods, trends in parts of the Pamir and Karakoram have moved from negative to positive, and those in the Tian Shan have become less negative (see Fig. S8). A similar story is apparent when MXPGR dates are considered, where the Tian Shan and parts of the Pamir have moved from negative to positive trends. Unfortunately, due to the climatically short nature of the dataset, it is not clear whether this change represents interannual variability or a reversal of a long-term trend.

The spatially and topographically complex changes in MXPGR, snowmelt end, and snowmelt period make interpretation of downstream impacts difficult. The long-term trend in HMA of a shortened and earlier melt season will impact downstream populations who rely on the consistent timing and volume of spring and summer runoff . Already the impacts of precipitation intensification and shifts in the snowmelt season have been felt in many regions . These trends are likely to continue as temperatures rise across HMA, and each major catchment will feel the impacts of a shortened snowmelt season regardless of changes in the start and end dates of melt.

Many regions rely on glaciers as their only water source between the end of snowmelt and the beginning of major precipitation systems . This important water reserve is certain to be impacted by, and reflect changes in, the snowmelt regime of HMA, as the timing of precipitation has been shown to be an important factor in the response of glaciers to climate change . While many regions have seen rapid glacier retreat , there exist regions of glacier stability and even growth, such as the Karakoram and Kunlun Shan . Our results (see Fig. ) show longer snowmelt periods in parts of the Pamir, later snowmelt end dates in parts of the Karakoram and Kunlun Shan, and relatively less negative trends in snowmelt end in the Pamir when compared with the rest of HMA. These regions overlap with both the “Karakoram anomaly” and positive glacier mass balances in parts of the Kunlun Shan and Pamir, implying that changes in the timing of the snowmelt season could be partially responsible for regional heterogeneity in glacier change.

The majority of HMA, however, exhibits a 3-decade-long trend towards an earlier end of the snowmelt season. Earlier snow clearance increases the absorption of solar radiation and thus stores more heat at high elevations and generates a positive feedback . As seasonal snow is removed earlier from glacier regions, glacier melt will accelerate. In general, glaciers in HMA are decreasing in volume and shrinking, which fits with the observed long-term decrease in snowmelt end dates (see Figs. –), despite clear spatial and temporal heterogeneity in these trends (see Fig. ).

Our algorithm-derived snowmelt end dates and those derived from the independent MOD10C1 product show close alignment, indicating that the algorithm is well suited to identifying the end of the snowmelt season (see Fig. ). The identification and interpretation of the MXPGR, however, is more difficult. While previous work has used the XPGR to identify the presence of liquid water in snowpack, this relationship has not been confirmed with in situ data in HMA. While it is likely that XPGR peaks are linked to melt–refreeze cycles, liquid water, or other snowpack metamorphism, these conclusions lack true in situ controls. This uncertainty, combined with the periods of melt and refreeze and late-season storms in much of HMA, makes linking MXPGR to the onset of the snowmelt season difficult. The multiple peaks and troughs in the XPGR data also hamper the identification of a single strong peak to classify as the MXPGR. Without rigorous measurements of surface air temperature or in situ monitoring of snowmelt, the efficacy of our algorithm for identifying snowmelt onset cannot be directly confirmed.

Despite these drawbacks, MXPGR dates are correlated with the day of year that HAR temperatures first start to increase and MODIS SCA is maximal (Fig. S2) and provide a single consistent proxy for the onset of the snowmelt season in this vast and largely unmonitored area. Furthermore, MXPGR is associated with days with a high temperature range and on-average positive daytime surface temperatures (see Fig. ). This implies that rising daytime temperatures, in conjunction with solar radiation, are linked to MXPGR dates in our study area. However, as we lack a direct control dataset for snowmelt onset, and there is a high degree of variance in the HAR surface temperature–MXPGR relationship, MXPGR dates and trends therein should be considered as less reliable than trends in snowmelt end dates.