A Closing Window: Satellite-Observed River-Ice Loss and Peak Water Risks for Sustainable Small-Hydropower Planning in the Tien Shan

Abstract

Sustainable small hydropower (SHP) is central to the clean-energy transition of mountainous Central Asia, yet its long-term reliability depends on a rapidly changing cryosphere. Winter river-ice dynamics—an underappreciated control on run-of-river generation—remain poorly characterized owing to the collapse of in situ hydrometeorological networks since 1991. We use a 112-month Sentinel-1 C-band SAR time series (February 2017–May 2026) over a 5320 km² headwater catchment of the Chu River basin, northern Tien Shan, Kyrgyzstan, to quantify river-ice phenology at 20 m resolution using a per-pixel summer-baseline anomaly approach. Mid-winter (December–February) ice cover declined significantly at −0.51%·yr⁻¹ (p = 0.013; Mann–Kendall p = 0.029), with the 2026 winter recording an unprecedented 2.6–2.8 σ departure from the 2017–2025 climatology. Contrasting the cold 2022 and warm 2026 winters revealed bidirectional climate sensitivity—early breakup versus persistent thin ice—posing distinct SHP hazards. ERA5-Land reanalysis (1992–2026) showed significant winter warming with no precipitation or snowfall trend, indicating thermally forced ice decline. Interpreted within a conceptual Peak Water scenario, this signals a closing window of opportunity for SHP generation, with direct relevance to sustainable water–energy management and the UN Sustainable Development Goals (SDG 7; SDG 13). Our results provide the first decadal, satellite-based evidence of river-ice loss for Central Asian mountain rivers and a transferable monitoring framework to support climate-resilient, sustainable hydropower planning in ungauged basins.

1. Introduction

Achieving the clean-energy and climate-adaptation goals of the UN 2030 Agenda in Central Asia depends critically on the sustainable development of small hydropower (SHP), whose long-term reliability is in turn governed by a rapidly changing mountain cryosphere. Mountain rivers draining the Tien Shan and Pamir ranges supply the majority of Central Asia’s freshwater resources and host a substantial portion of Kyrgyzstan’s installed small hydropower (SHP) capacity. In high-altitude run-of-river systems, winter river ice formation constitutes a critical but underappreciated operational constraint. Frazil and anchor ice can rapidly clog trash racks, block intakes, and force shutdowns, while breakup and ice jams downstream of hydropower operations can cause hanging dams, sharp water-level changes, and ice-run hazards that complicate shutdown/restart strategies [1,2,3,4]. In small, often steep run-of-river plants, ice problems can exceed flood damages, with reduced production capacity due to frazil accumulation at trash racks and ice runs blocking intakes constituting dominant operational issues [5].

The cryosphere of the Tien Shan is undergoing rapid transformation. Glaciers in the northern and central Tien Shan have shrunk by 15–20% in many sectors since the 1960s–1970s [6,7,8], with a broader assessment reporting a 27% glacier mass loss across the entire Tien Shan from 1961–2012, attributed primarily to increased summer melt [9]. Snow-covered days have decreased significantly in central and eastern Tianshan sectors (e.g., −11.9% and −8.0% over 2001–2015), with changes strongly linked to rising air temperatures [10,11]. For Central Asian lakes, including Issyk-Kul, MODIS-based phenology analyses (2002–2020) reveal regional contrasts in freeze-up and breakup patterns, although many lakes do not uniformly follow the canonical “later freeze, earlier breakup” trend [12]. More broadly, Eurasian and Tibetan Plateau lakes exhibit widespread shortening of ice seasons, driven primarily by air temperature and solar radiation [13,14].

Despite this rapid cryospheric change and the strategic importance of SHP for Central Asia’s energy transition, systematic long-term observations of river ice phenology in Kyrgyz mountain basins remain virtually absent due to the post-Soviet collapse of the hydrometeorological network. This data gap is particularly acute for high-mountain rivers, where complex winter ice processes, such as anchor ice dams, water–ice flows, and aufeis, develop under conditions of sharp cooling (air < −20 °C; up to 10 °C·day⁻¹) and rapid subsequent warming [15]. In the northern Tien Shan of southeastern Kazakhstan, anchor ice dams can raise water levels by 1.5–2 m before catastrophic failure, producing destructive water–ice flows with depths up to 5 m, speeds > 10 m·s⁻¹, and discharges up to 300 m³·s⁻¹ [15]. In Ladakh, aufeis fields form layered ice bodies through repeated overflow and freezing, with distinct accumulation (November–April) and melt (May–July) phases and strong interannual variability [16].

Spaceborne C-band synthetic aperture radar (SAR), with its all-weather and day–night acquisition capability, has emerged as the principal tool for river ice monitoring in data-scarce regions. Sentinel-1 (5.405 GHz, 20 m IW mode, 6–12-day revisit) has been successfully used for river ice classification and breakup mapping, with supervised and threshold-based methods achieving overall accuracies of 77–96% across diverse rivers [17,18,19,20]. On the Nemunas and Neris rivers in Lithuania (80–300 m wide), simple VV backscatter threshold models detected ice extent over 525 km with a median VV difference of ~13 dB between sheet ice and open water [17]. However, applications to alpine headwater rivers face significant challenges: narrow channel widths, terrain-induced layover and shadow, and complex geometry require dual-pol data and multi-orbit strategies to retrieve usable information [21,22,23,24]. Dual-polarized C-band SAR has nonetheless shown potential for ice thickness inversion in high-order rivers on the Tibetan Plateau, with RMSEs of 0.11–0.26 m [25], and Random Forest classification of Landsat and Sentinel-2 imagery has successfully mapped 27 recurrent aufeis fields in Ladakh, demonstrating transferability to other high-mountain regions [16].

Beyond ice presence/absence, recent work emphasizes that ice thickness may serve as a leading indicator of climate-driven ice loss. A global synthesis shows that Northern Hemisphere lake ice duration has shortened by 28 days on average over 150 years, with recent decades exhibiting accelerated loss in both duration and thickness [26]. Large-ensemble modeling projects future lake ice loss as reduced maximum thickness and shorter duration, with maximum thickness decreasing by ~0.23 m by 2100 over the Tibetan Plateau and Canadian Arctic [27]. Projections of “safe ice” for transport and recreation indicate dramatic declines in days meeting thicker-ice thresholds, highlighting that thickness is often more sensitive than simple presence/absence [28].

In this study, we present the first decadal-scale satellite SAR record of river ice for a Tien Shan headwater catchment. Our specific objectives are (i) to develop a per-pixel summer-baseline anomaly approach robust to terrain-induced heterogeneity in narrow alpine rivers, (ii) to quantify decadal trends in mid-winter ice cover and intensity over 2017–2026, and (iii) to assess implications of the observed ice loss for SHP operations under the Peak Water hypothesis. The overall four-phase workflow of data acquisition, preprocessing, ice classification, and SHP impact assessment is summarized in Figure 1.

2. Materials and Methods

2.1. Study Area

The study catchment is a 5320 km² headwater basin of the Chu River system, draining the northern slopes of the Kyrgyz Ala-Too range in the northern Tien Shan (centroid ≈ 75.83° E, 42.25° N; elevation 1500–4500 m). The basin is snow-dominated, with negligible direct glacier contribution within a 50 km radius of the proposed SHP site, although the broader area of interest (AOI) contains 515 glaciers totaling 2188 km². The river network spans 106 major segments (HydroSHEDS UPLAND_SKM > 100), with main-stem discharge of 15–20 m³·s⁻¹. The site is representative of mid-elevation Kyrgyz SHP development zones, where post-Soviet hydrometeorological monitoring has been particularly sparse (Figure 2). We note that no in situ discharge or water-level gauging has been maintained in the catchment since the collapse of the Soviet hydrometeorological network in 1991; consequently, our analysis relies on satellite observations and reanalysis, and modeled runoff is used only as a qualitative proxy for streamflow seasonality (Appendix A, Figure A3).

2.2. Sentinel-1 SAR Data

Sentinel-1 C-band Synthetic Aperture Radar (SAR) Ground Range Detected (GRD) imagery was acquired from the European Space Agency (ESA) Copernicus Programme (https://scihub.copernicus.eu; accessed in 1 May 2026) through the Google Earth Engine (GEE) cloud computing platform. We selected all scenes acquired in Interferometric Wide (IW) swath mode with dual-polarization (VV+VH) on descending orbits over the study catchment from February 2017 to May 2026. GEE’s standard Sentinel-1 preprocessing pipeline—comprising GRD border noise removal, thermal noise removal, radiometric calibration to sigma-naught (σ⁰), and terrain correction using the SRTM 30 m DEM—was applied to all scenes. For each calendar month, the per-pixel median backscatter across all available scenes was computed within GEE at 20 m resolution to suppress speckle while preserving the mean radiometric response, yielding 112 monthly composites.

The 112 monthly composites were exported from GEE as GeoTIFF files (EPSG:32643) and subsequently clipped to the 5320 km² catchment boundary on a local workstation using the rasterio (v1.3) and geopandas (v0.14) libraries in Python 3.11. Clipped rasters were stored as LZW-compressed, internally tiled (256 × 256) GeoTIFFs with NoData = 0. This two-stage processing chain (cloud-based compositing followed by local clipping) reduced the raw data volume from 156.1 GB to 11.5 GB (a 92.6% reduction), enabling efficient pixel-wise time-series analysis on local hardware.

2.3. River Mask Generation

A combined river mask was created by merging the JRC Global Surface Water occurrence layer (occurrence ≥ 20%) with HydroSHEDS Free-Flowing Rivers (UPLAND_SKM > 50 km²) buffered by 30–100 m depending on upstream area. The resulting mask covered 88.8 km² (1.67% of the catchment), consistent with the expected 1–2% river coverage in mountain catchments. After resampling to the 20 m SAR grid, 145,531 valid river pixels were retained for analysis.

2.4. Per-Pixel Summer-Baseline Anomaly Approach

To overcome the confounding effects of static terrain backscatter and inter-pixel heterogeneity in mountainous SAR imagery, we adopted a per-pixel summer-baseline anomaly approach rather than a basin-uniform threshold. For each pixel, the mean VV during ice-free months (June–August, n = 27 scenes) was computed as a local reference state representing open-water and adjacent wet-soil conditions, and monthly anomalies were derived as deviations from this baseline. A pixel was classified as frozen when its anomaly fell below −2 dB.

This pixel-wise normalization is conceptually similar to per-pixel change-detection frameworks widely used in flood mapping, soil-moisture retrieval, and permafrost monitoring [29,30,31]. The Percentile-based Freeze–Thaw Identification (PFTI) approach, for instance, uses per-pixel temporal percentiles (e.g., 5th/95th) of VV/VH backscatter to classify frozen, thawed, and transition states rather than applying global thresholds [29]. The −2 dB anomaly threshold matches the order of magnitude of C-band backscatter shifts between water and ice reported in the literature: median VV differences of ~13 dB between sheet ice and open water on temperate rivers [17], and ratio thresholds of 1–2 dB between ice types and water in sea-ice studies [32,33]. Our threshold therefore represents a conservative, literature-consistent empirical choice rather than a universal physical constant and is specifically tuned to suppress slope, vegetation, and roughness contributions while preserving the temporal climate signal in heterogeneous alpine catchments.

To verify that our results are not artifacts of this particular threshold, we performed a sensitivity test in which the frozen-pixel classification was repeated across thresholds ranging from −1.0 to −3.0 dB, using the same river mask and summer baseline throughout. The decadal sign and significance of the mid-winter ice-cover trend were robust across the −1.5 to −3.0 dB range (all p < 0.085), with only the most permissive −1.0 dB threshold yielding a non-significant result; the −2 dB choice falls within the stable central range (Appendix A, Figure A4, and

Table A3. Sensitivity of the mid-winter (DJF) river-ice cover trend to the SAR backscatter-anomaly threshold (−1.0 to −3.0 dB), using a common river mask and summer baseline. The absolute slope increases in magnitude at stricter thresholds because the baseline frozen fraction is correspondingly smaller; however, the sign and statistical significance of the declining trend are stable across the −1.5 to −3.0 dB range, confirming that the decline is not an artifact of the −2 dB choice. Absolute slopes here are computed on the full river mask and differ marginally from the main-analysis value (−0.512%·yr⁻¹; Figure 4, Appendix A, Table A2), which uses the valid-pixel subset; the relative robustness across thresholds is unaffected. * p < 0.05; marginal 0.05 ≤ p < 0.09; n.s., not significant.

Threshold (dB)	OLS Slope (% yr−1)	R2	p	Significance	Mean Cover (%)
−1	−0.213	0.127	0.346	n.s.	62.5
−1.5	−0.353	0.373	0.081	marginal	55.4
−2	−0.468	0.551	0.022	*	48.5
−2.5	−0.543	0.563	0.02	*	41.8
−3	−0.566	0.485	0.037	*	35.6

).

2.5. Statistical Analysis

Decadal trends in mid-winter (December–February) ice cover and backscatter anomaly were assessed using ordinary least-squares (OLS) linear regression, with significance tested at α = 0.05. To address the sensitivity of short time series to individual extreme years, we additionally applied the non-parametric Mann–Kendall trend test and Sen’s slope estimator to all key time series; these results are reported alongside the OLS estimates in Appendix A,

Table A2. Parametric (OLS) and non-parametric (Mann–Kendall, Sen’s slope) trend tests for key cryospheric variables, Site A (2018–2026, n = 9 complete winters). Slopes are reported with explicit sign; p-values < 0.05 indicate statistical significance.

Variable	OLS Slope	R2	p (OLS)	Sen’s Slope	MK Z	p (MK)	n
Mid-winter ice cover (% yr−1)	−0.512	0.607	0.013	−0.480	−2.189	0.029	9
Mid-winter VV anomaly (dB yr−1)	+0.026	0.216	0.207	+0.031	+1.147	0.251	9
Peak-winter ice cover (% yr−1)	−0.397	0.387	0.074	−0.429	−0.938	0.348	9
Mean winter freeze (% yr−1)	−0.512	0.607	0.013	−0.480	−2.189	0.029	9
Freeze-up timing (month yr−1)	+0.067	0.171	0.268	+0.000	+0.730	0.466	9
Break-up index (yr−1)	+0.083	0.104	0.397	+0.000	+0.730	0.466	9
Ice-season length (month yr−1)	+0.017	0.004	0.872	+0.000	+0.417	0.677	9

. Anomaly detection for individual winters was based on z-scores against the 2017–2025 climatology. Welch’s t-tests were applied where appropriate. For z-score-based anomaly tests, one-tailed p-values were derived from the standard normal distribution, reflecting the directional hypothesis that mid-winter ice cover decreases under regional climate warming. Throughout the paper, regression results are reported uniformly as slope, coefficient of determination (R²), and p-value.

2.6. SHP Generation Projection

To project long-term SHP generation under the Peak Water hypothesis, we adopted a purely conceptual trajectory model rather than a calibrated hydrological simulation, because no in situ discharge records exist for the catchment. The model assumes that mountain catchments with glacier/snow contribution exhibit increasing runoff until “Peak Water” before declining; for Tien Shan headwaters, Peak Water is expected between 2040 and 2055 based on regional glacier-runoff modeling [9]. We modeled annual generation for a representative 5 MW run-of-river plant with an assumed capacity factor of 0.55—values typical of mid-elevation Kyrgyz SHP installations—under a rising limb (1990–2048, +30% vs. 1990 baseline) and a declining limb (2048–2080, −25% vs. baseline), with uncertainty bands of ±10–25% expanding post-peak. We emphasize that these parameters are illustrative and intended to bound the qualitative shape of the trajectory; absolute generation values should not be interpreted as site-calibrated forecasts.

3. Results

3.1. Seasonal Ice Climatology

The 2017–2025 climatology revealed a clear annual ice cycle with peak frozen-pixel ratios of 46.7 ± 1.9% in December and 46.4 ± 1.5% in January, declining through spring to a summer minimum of 6.5 ± 2.1% in June (Figure 3). Mid-winter (December–February) values exhibited remarkably low interannual variability (CV < 5%), indicating a stable freezing regime over the first nine winters of the record. The seasonal amplitude (≈40 percentage points) provided a strong basis for trend detection. Mid-winter ice covered an average area of 26.8 km², approximately fivefold higher than the summer-baseline noise (5.5 km²). To allow a full assessment of interannual variability, Figure 3 displays the individual seasonal course of every monitored year (2017–2025) overlaid on the climatological mean ± 1 σ envelope and the 2026 anomaly.

3.2. Decadal Trend in Mid-Winter Ice Cover and Intensity

Mid-winter ice cover declined significantly at −0.51%·yr⁻¹ (R² = 0.61, p = 0.013) over 2017–2026, equivalent to a cumulative reduction of approximately 4.6 percentage points (47.6% → 43.1%) over the decade (Figure 4a). This decline was confirmed by the non-parametric Mann–Kendall test (Sen’s slope −0.48%·yr⁻¹, p = 0.029), indicating that the trend is not an artifact of individual extreme years. The endpoint contrast is illustrated quantitatively in Appendix A,

Table A1. Monthly Sentinel-1 SAR backscatter statistics and derived ice metrics for January–April of the coldest (2018) and warmest (2026) winters in the 2017–2026 record. VV values are mean ± standard deviation over n = 145,531 river pixels (20 m resolution); VH standard deviation and pixel count are omitted for brevity (VH σ ≈ 4.5–5.0 dB throughout). VV − VH denotes the co/cross-polarization difference. Anomaly = monthly VV—per-pixel summer baseline (June–August mean, n = 27 scenes); more negative values indicate a stronger ice signal. Freeze (%) is the fraction of river pixels with anomaly < −2 dB; Area is the corresponding frozen river surface in km². Values for the warmest winter (2026) systematically exceed (i.e., are less negative than) those of the coldest winter (2018), with the largest contrast occurring in February (z = −2.57 vs. climatology).

Year	Month	VV (dB)	VH (dB)	VV − VH (dB)	Anomaly (dB)	Freeze (%)	Area (km2)
2018	Jan	−14.33 ± 4.31	−22.23	7.89	−1.76	47	27.36
2018	Feb	−14.61 ± 4.91	−22.63	8.02	−2.04	49.8	28.98
2018	Mar	−13.90 ± 3.89	−22.43	8.52	−1.33	37.4	21.74
2018	Apr	−13.35 ± 3.59	−22.29	8.94	−0.79	25	14.53
2026	Jan	−14.10 ± 4.07	−23.79	9.7	−1.53	43.1	25.1
2026	Feb	−13.96 ± 4.29	−23.16	9.19	−1.39	39.9	23.25
2026	Mar	−14.15 ± 4.18	−23.52	9.36	−1.59	42.2	24.55
2026	Apr	−13.41 ± 3.90	−22.50	9.09	−0.84	27.8	16.17

Notes. VV values are mean ± standard deviation over n = 145,531 river pixels (20 m resolution). VH std and pixel count omitted for brevity (VH std ≈ 4.5–5.0 dB throughout). Anomaly = VV—summer baseline (Jun–Aug mean, 27 scenes); more negative values indicate stronger ice signal. Freeze (%) = fraction of river pixels with anomaly < −2 dB. Area = corresponding frozen river surface (km²).

, which compares monthly VV backscatter, anomaly, and frozen area between the coldest year of the record (2018; DJF ice cover 48.3%) and the warmest year (2026; DJF 41.5%). Concurrently, the mean winter VV anomaly exhibited a weakening tendency (+0.026 dB·yr⁻¹, R² = 0.22, p = 0.21; Mann–Kendall p = 0.25), suggesting that the ice signal became progressively closer to that of open water; however, this backscatter trend is not statistically significant over the nine-winter record and should be regarded as a qualitative indication rather than conclusive evidence (Figure 4b). Because SAR backscatter responds to a combination of surface roughness, ice structure, and snow loading, it serves as a relative indicator of ice condition rather than a direct measurement of thickness.

By contrast, ice phenology metrics (freeze-up month, breakup month, season length) showed no significant trends (all p > 0.27; Mann–Kendall p > 0.46), with the freeze-up and breakup months remaining at October–November and March–April, respectively. Only peak winter ice cover exhibited a marginal declining trend (−0.40%·yr⁻¹, p = 0.074). This decoupling of stable phenological timing from declining mid-winter ice cover is consistent with the early stage of a “thinning-before-shortening” regime documented for Northern Hemisphere lake ice [26,27,28].

The corresponding long-term climate context, derived from 35 years of ERA5-Land reanalysis (1992–2026), is presented in Figure 5. Mean winter air temperature warmed significantly (+0.50 °C·decade⁻¹, p = 0.015), and cold-season severity, measured by Accumulated Freezing Degree Days (AFDD), declined concomitantly (−128 °C·d·decade⁻¹, p = 0.003). The 2026 winter appears as a significant outlier in both datasets.

3.3. Long-Term Precipitation and Snow Context

Because precipitation and snow cover are significant controls on the formation and disappearance of river ice in mountain basins, we examined their long-term winter trends using the same ERA5-Land record (1992–2026; Appendix A,

Table A4. Long-term winter (DJF) precipitation and snow trends over Site A (ERA5–Land, 1992–2026; n = 34 winters). No variable shows a statistically significant trend (all p > 0.4).

Variable	OLS Slope (per Decade)	R2	p (OLS)	Sen’s Slope (per Decade)	p (MK)
Winter total precipitation (mm)	1	0	0.72	0.8	0.7
Winter total snowfall (mm w.e.)	1	0	0.73	0.9	0.7
Winter mean snow depth (cm)	−0.9	0	0.71	−2	0.46
Winter mean SWE (mm)	−1.4	0	0.77	−2.6	0.46

). Over the 34 complete winters, neither winter total precipitation nor snowfall exhibited a statistically significant trend (OLS +1.0 mm·decade⁻¹, R² ≈ 0.00, p ≈ 0.72; Mann–Kendall p ≈ 0.70). Winter mean snow depth and snow water equivalent (SWE) showed weak, non-significant declines (Sen’s slope −2.0 cm·decade⁻¹ and −2.6 mm·decade⁻¹, respectively; Mann–Kendall p ≈ 0.46). The absence of significant precipitation or snowpack trends, contrasted with the significant winter warming and AFDD decline reported above (Figure 5), indicates that the observed river-ice decline is driven primarily by thermal forcing rather than by changes in precipitation supply or snow loading.

3.4. The 2026 Anomaly

The 2026 winter recorded an unprecedented departure from climatology (Figure 3). The spatial extent of this anomaly is shown in Appendix A, Figure A1 (VV polarization) and Appendix A, Figure A2 (VH polarization), which display SAR backscatter mosaics for January–April of the warmest winter (2026) alongside the coldest winter on record (2018) as a reference for maximum-ice conditions. January and February ice cover dropped to 43.1% and 39.9%, corresponding to z-scores of −2.77 and −2.57 (one-tailed p ≈ 0.003 and 0.005, respectively), more than 2.5 σ below the 2017–2025 climatological mean, which had itself remained within a coefficient of variation of 5%. Mean winter VV anomaly in 2026 (−1.46 dB) was the highest (least frozen) in the record, compared to a 2017 value of −1.91 dB. Conversely, March and May 2026 exhibited above-average ice cover (z = +1.04 and +1.18), suggesting that thinner mid-winter ice persisted later into spring due to the delayed onset of strong warming.

3.5. Bidirectional Climate Sensitivity: 2022 vs. 2026

To characterize the modes of climate impact on river ice in this catchment, we compared the two most diagnostic anomalous winters of the 2017–2026 record (

Table 1. Comparison of the cold (2022) and warm (2026) winter anomalies relative to the 2017–2025 climatology. Ice cover values are percent of river pixels with anomaly < −2 dB; z-scores computed against the nine-winter baseline. Winter VV anomaly (DJF) is in dB relative to the per-pixel summer (June–August) baseline. Season length is the number of months with ice cover ≥ 20%.

Variable	2022 (Cold)	2026 (Warm)	Climatology
Jan ice cover (%)	46.4 (z = −0.04)	43.1 (z = −2.77)	46.4 ± 1.2
Feb ice cover (%)	45.3 (z = −0.21)	39.9 (z = −2.57)	45.8 ± 2.3
Mar ice cover (%)	25.2 (z = −5.69)	42.2 (z = +1.57)	38.5 ± 2.3
Winter VV anomaly (dB)	−1.71	−1.46	−1.74 ± 0.13
Season length (≥20%)	7 months	5 months (incomplete)	7.3 ± 0.7

). We emphasize that the 2022 and 2026 winters were not selected as the absolute temperature extremes of the record—that role is held by 2018, the coldest winter in terms of ice extent (DJF ice cover 48.3%), which is retained as the maximum-ice reference in Appendix A, Figure A1 and Table A1. Rather, 2022 and 2026 were selected because they represent two fundamentally different modes of departure from the climatological norm, each carrying distinct implications for small hydropower (SHP) operations.

The 2022 winter exemplifies the “thick-ice, early-breakup” cold mode. Mid-winter ice cover remained near the 2017–2025 baseline (z ≈ 0 in both January and February; Table 1), yet the winter mean VV anomaly was the most negative in the record (−1.71 dB), consistent with unusually thick and/or hydraulically rough ice. This anomalously robust ice cover collapsed abruptly the following spring: ice cover fell from 45.3% in February to 25.2% in March 2022, a record-breaking departure of z = −5.69 from the March climatology (38.5 ± 2.3%).

The 2026 winter exemplifies the opposite “thin-ice, persistent-melt” warm mode. Mid-winter ice cover dropped to record-low values of 43.1% in January and 39.9% in February (z = −2.77 and −2.57, respectively; Table 1), accompanied by a markedly less negative winter mean VV anomaly (−1.46 dB)—the highest (least frozen) value in the record. Unlike the 2022 case, however, the thinner mid-winter ice in 2026 did not trigger an early breakup; instead, March and April ice cover remained above the climatological mean (z = +1.57 and +1.04, respectively), indicating a delayed and prolonged melt-out. Appendix A, Table A1 quantifies this asymmetry: the VV − VH co/cross-polarization difference increased from 7.89–8.94 dB in 2018 to 9.09–9.70 dB in 2026, which is consistent with smoother and thinner ice rather than reduced ice extent alone.

These two contrasting modes—a cold-but-fragile regime (thick ice followed by abrupt breakup, exemplified by 2022) and a warm-but-persistent regime (thin discontinuous ice with delayed melt, exemplified by 2026)—correspond to fundamentally different SHP operational hazards. The cold-mode pathway generates ice-run and ice-jam risks consistent with documented water–ice flow events in the northern Tien Shan, where rapid warming following sharp cooling produces destructive discharges [15]. The warm-mode pathway, by contrast, promotes frazil and slush formation, which is responsible for the majority of trash-rack clogging events at high-altitude SHP plants [4,5]. Importantly, the existence of both modes within a single ten-year record demonstrates that climate-driven SHP risk cannot be reduced to a unidirectional “warmer = safer” or “colder = safer” framing, a point developed further in Section 4.3.

3.6. Implications for Hydropower Generation Under the Peak Water Hypothesis

Translating the observed ice loss into hydropower terms within the conceptual framework described in Section 2.6, our 2017–2026 trend is consistent with the early rising limb of the Peak Water trajectory (Figure 6a). We caution that, in the absence of in situ discharge records, this translation is illustrative rather than a calibrated generation estimate; the modeled runoff proxy derived from ERA5-Land (Appendix A, Figure A3) captures the expected snowmelt-dominated seasonality but cannot substitute for gauged streamflow. Extrapolation to mid-century under the Peak Water framework projects a transient generation surplus reaching approximately +30% above the 1990 baseline by ~2048, followed by a decline toward −25% by 2080 as glacier and snow reserves are progressively exhausted (Figure 6b). These projections should be read as a conceptual scenario bounding the plausible trajectory shape rather than as a site-specific forecast.

4. Discussion

4.1. Thermodynamic Regime Shift and Its Detection by SAR

The significant decline of mid-winter ice cover (−0.51%·yr⁻¹), together with a weakening tendency in the VV backscatter anomaly, in the absence of a statistically significant phenological shift (freeze-up, breakup, and season length all p > 0.27; Section 3.2), is directionally consistent with the “thinning-before-shortening” framework documented for Northern Hemisphere lake ice [26,27,28]. Under this framework, warming winters first erode ice thickness and surface quality before substantially altering the freeze/thaw calendar. We stress, however, that the backscatter weakening is not statistically significant over our nine-winter record and that SAR intensity is a relative indicator; the thinning interpretation is therefore advanced as a hypothesis consistent with our observations rather than as a directly measured result. Large-ensemble climate projections support this sequencing: future lake ice loss is expected to manifest as both reduced maximum thickness and shorter duration, with maximum thickness projected to decrease by ~0.23 m by 2100 over the Tibetan Plateau [27].

Two independent but suggestive lines of evidence in our data point toward a thickness-related change. First, direct coupling of SAR-derived mid-winter ice cover with cumulative cold (AFDD) reveals an inverse relationship (Figure 7a; R² = 0.31, p = 0.096, marginal at α = 0.05), and the Stefan-equation reconstruction (Figure 7b) yields a tentative thinning trajectory across the observation period, with the 2026 winter producing the thinnest estimated ice in the record (≈45 cm). Because no in situ ice-thickness measurements exist for the catchment, this reconstruction is presented with an explicit parametric uncertainty band rather than as a validated thickness retrieval. Second, the polarization geometry of the SAR signal is consistent with a thickness- rather than extent-driven change: the VV − VH co/cross-polarization difference increased from 7.89–8.94 dB in 2018 to 9.09–9.70 dB in 2026. Because VH is more sensitive to volume scattering from rough or layered ice while VV is dominated by specular returns from smooth surfaces, this widening contrast is suggestive of ice that is both smoother and thinner rather than simply less extensive. We emphasize that, in the absence of field validation, these signatures are interpreted qualitatively, and confirmation would require ground-based ice-thickness measurements.

This regime shift is consistent with the broader pattern of cryospheric retreat in the region. Tien Shan glaciers have lost 15–20% of their area since the 1960s–1970s [6,7,8,9], and snow-covered days have decreased by 8–14 days per decade in several Tianshan sectors [10,11]. Although the study catchment is snow-dominated with negligible direct glacier input near the SHP site, the river-ice signal we observe is part of an interconnected cryospheric response in which glacier ablation, snowpack evolution, and river-ice thinning are driven by the same warming forcing. Regionally, accelerating glacier mass loss and earlier snowmelt advance the timing and increase the volume of spring runoff [9], which in turn modulates the thermal and hydraulic regime governing river-ice growth and decay; conversely, thinner and shorter-lived river ice feeds back on channel heat exchange. Although a full attribution of these synergistic linkages is beyond the scope of our SAR record, our results indicate that river ice responds sensitively to the same climatic forcing as glaciers and snow and can therefore serve as a complementary indicator within an integrated cryospheric monitoring framework.

Finally, the successful detection of these trends in narrow, steep mountain rivers addresses a methodological challenge highlighted in the SAR literature: monitoring high-order, narrow rivers with SAR is difficult due to complex topography and small channel width [25], with terrain-induced layover and shadow being major limitations [21,22,23]. Our per-pixel summer-baseline anomaly approach explicitly addresses these challenges by normalizing static terrain effects and isolating the temporal climate signal, building on established per-pixel change-detection methods in flood mapping and permafrost monitoring [29,30,31].

4.2. The 2026 Anomaly: Tipping Point or Interannual Variability?

The 2026 winter’s ~2.7–2.8 σ deviation in January–February ice cover is the largest single-winter departure in the nine-year record. Mid-winter ice cover in 2017–2025 exhibited remarkably low variability (CV < 5%), making a departure of this magnitude (|z| > 2.5) a priori improbable (one-tailed p ≈ 0.003 for January). The contrast between 2022 (cold, thick ice) and 2026 (warm, thin ice) further demonstrates bidirectional sensitivity, in which both cold and warm extremes can produce hazardous SHP conditions through different mechanisms.

Nonetheless, a nine-year observational record is insufficient to conclusively distinguish a regime shift from interannual variability. To partially address this short-record limitation, we placed the SAR observations within the longer 1992–2026 ERA5-Land reanalysis context (Figure 5 and Appendix A, Figure A3), which confirms that 2026 ranks among the warmest and least severe winters of the 35-year record. Continued monitoring through the late 2020s, combined with longer climate reanalyses and CMIP6-informed projections, will be necessary to establish whether 2026 marks the onset of more frequent warm-winter anomalies or a transient extreme.

4.3. SHP Operations: Beyond Linear Risk Reduction

The bidirectional climate sensitivity revealed in our analysis carries direct implications for SHP operational planning. A central finding in the broader hydropower literature is that climate-driven changes in ice regimes do not translate into proportional reductions in operational problems: shortened ice seasons are offset by more mid-winter thaws and breakups, expanding the time window for frazil and jam events [5,34,35,36]. For small, often steep run-of-river plants in Alpine and Norwegian settings, ice problems exceed flood damages, with frazil accumulation and ice-run intake blockages constituting dominant issues [5].

Frazil and drifting ice can rapidly clog trash racks, block intakes, and force shutdowns; frazil is identified as a major and frequent cause of immediate blockage at hydropower intakes [3,37,38]. Breakup and ice jams can cause hanging dams, sharp water-level changes, and ice-run hazards that complicate shutdown/restart strategies [3,39,40].

These considerations argue against a simplistic interpretation of declining ice as uniformly beneficial for SHP. Instead, our results suggest that the Kyrgyz SHP sector may currently be experiencing a transient productivity gain from meltwater surge, but with rising operational variability that may erode net benefits. For the specific conditions of high-altitude Tien Shan run-of-river plants, we recommend a set of concrete adaptive measures: (i) installing heated or sloped trash racks and air-bubbler or warm-water diffuser systems at intakes to mitigate frazil and anchor-ice blockage during cold-mode winters [37,38]; (ii) adopting flexible winter operating rules—including controlled discharge to encourage a stable continuous ice cover where hydraulically beneficial, and pre-emptive drawdown ahead of forecast warm spells to limit ice-jam surge risk during warm-mode breakups [3,39]; (iii) integrating near-real-time satellite ice monitoring of the type demonstrated here into operational decision support, providing early warning of both early-breakup (cold-mode) and persistent-thin-ice (warm-mode) conditions [36,41,42,43]; and (iv) scheduling maintenance and inspection windows around the October–November freeze-up and March–April breakup periods identified as phenologically stable in this record. Because the dominant hazard differs between cold and warm winters, these measures should be deployed as a flexible portfolio rather than a fixed seasonal protocol.

4.4. The Peak Water Hypothesis and Long-Term SHP Planning

Our observed ice loss is consistent with the rising limb of the Peak Water trajectory, in which deglaciating mountain catchments experience temporarily increased runoff before reserves are exhausted [9]. For Tien Shan SHP planning, this implies a window of opportunity—likely closing by 2040–2055—during which winter generation may benefit from reduced ice blockage and earlier meltwater pulses. We note, however, that the ERA5-Land modeled runoff for our catchment does not itself show a rising trend over 1992–2026 (Appendix A, Figure A3); given the known limitations of reanalysis runoff in complex mountainous terrain and the snow-dominated character of this specific basin, we interpret this cautiously and emphasize that catchment-scale runoff projections require dedicated, calibrated hydrological modeling. Beyond Peak Water, generation is generally projected to decline as glacier and snow contributions diminish.

This framing has policy implications. Current SHP investment decisions in Kyrgyzstan, often based on recent (post-2000) hydrological records, may overestimate long-term plant productivity by capturing the rising limb without accounting for the subsequent decline. Incorporating Peak Water projections into project financing, water-rights allocation, and grid-integration planning could improve the long-term resilience and sustainability of Central Asian SHP development, ensuring that the region’s clean-energy transition remains robust to cryospheric change and aligned with the water–energy–climate objectives of the UN Sustainable Development Goals.

4.5. Method Transferability and Limitations

The per-pixel summer-baseline anomaly approach developed here is transferable to other data-scarce mountain regions. The combination of Sentinel-1 SAR with HydroSHEDS-based river masks and JRC permanent-water layers requires only freely available global datasets, and the entire workflow can be executed within Google Earth Engine.

Several limitations warrant acknowledgment. First, the −2 dB anomaly threshold is empirically derived; although our sensitivity test (Section 2.4, Appendix A, Figure A4) shows that the decadal trend is robust across −1.5 to −3.0 dB, recalibration may be required for catchments with different ice types or terrain. Second, monthly median composites smooth over sub-monthly events such as freeze-up and breakup transitions, limiting temporal resolution for ice-jam forecasting. Third, the absence of in situ ice-thickness and discharge validation in our basin means that the SAR-derived intensity weakening and the Stefan-equation thickness reconstruction can only be interpreted as relative indicators, not absolute measurements; field campaigns and gauged streamflow would be required for quantitative validation. Fourth, in narrow mountain channels where the river width approaches or falls below the 20 m SAR pixel size, mixed pixels combining river ice with adjacent bank, vegetation, or wet-soil returns introduce classification uncertainty that may bias the absolute frozen-fraction estimates, although the per-pixel anomaly normalization mitigates this effect for relative temporal comparisons. Fifth, our nine-year record, while one of the longest continuous SAR-based river-ice records published for Central Asian mountain rivers, remains short for detecting regime shifts, and continued monitoring together with CMIP6-informed projections will be essential.

To support reproducibility, Appendix A, Figure A1 and Figure A2 provide the spatial SAR backscatter mosaics underlying the catchment-aggregated trends, and Appendix A, Table A1 reports the underlying monthly pixel-level statistics, allowing readers to independently inspect the 2018 (maximum-ice) and 2026 (minimum-ice) endpoints of our ten-year record.

5. Conclusions

This study presents the first decadal-scale Sentinel-1 SAR record of river ice for a Tien Shan headwater catchment (5320 km², 2017–2026, n = 112 monthly composites). Using a per-pixel summer-baseline anomaly approach, we documented a significant decline in mid-winter ice cover at −0.51%·yr⁻¹ (p = 0.013; Mann–Kendall p = 0.029), accompanied by a weakening tendency in backscatter anomalies (+0.026 dB·yr⁻¹, p = 0.21) that, while not statistically significant over the nine-winter record, is directionally consistent with a thermodynamically driven “thinning-before-shortening” regime shift. A threshold-sensitivity test confirmed that the ice-cover trend is robust to the choice of anomaly threshold, and non-parametric tests corroborated the parametric results. Long-term ERA5-Land reanalysis showed significant winter warming with no significant change in precipitation or snowfall, indicating that the decline is primarily thermally forced. The 2026 winter recorded an unprecedented 2.6–2.8 σ deviation from the 2017–2025 climatology. Bidirectional climate sensitivity—with cold winters (2022) producing rapid breakups and warm winters (2026) producing thin persistent ice—argues against a simplistic view of declining ice as uniformly beneficial for SHP operations.

Under the Peak Water hypothesis, interpreted here as a conceptual scenario in the absence of calibrated hydrological data, observed ice loss is associated with a transient SHP generation surplus that may peak around 2045–2050 before declining. Adaptive operational strategies—including frazil mitigation at intakes, flexible winter operating rules, and satellite-based real-time monitoring—will be required to maintain SHP reliability across this transition. Our results provide the first decadal, satellite-based evidence of river-ice loss for Central Asian mountain rivers and offer a transferable monitoring framework for ungauged SHP basins worldwide—supporting evidence-based, climate-resilient planning that advances the sustainable development of mountain hydropower (SDG 7) under accelerating climate change (SDG 13).