Skip to Content
You are currently on the new version of our website. Access the old version .
Remote SensingRemote Sensing
  • Article
  • Open Access

13 January 2026

26 Pages

Spatiotemporal Dynamics and Drivers of Potential Winter Ice Resources in China (1990–2020) Using Multi-Source Remote Sensing and Machine Learning

Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, Beijing 100101, China

Highlights

What are the main findings?
  • This paper developed a simple, transferable “surface water + LST” framework on Google Earth Engine to map river and lake ice across China (1990–2020), avoiding complex SAR modeling and regional thresholds and enabling consistent, large-scale, long-term monitoring of freshwater ice across diverse climate zones and surface conditions.
  • This paper revealed that, although the freezing-zone boundary moved markedly northward (over 400 km in the Northeast Plain and ~13 km/year along the eastern coast), the total ice-covered area still increased (~1.1%/year) while the average ice season slightly shortened, showing asynchronous temporal–spatial responses to warming.
What are the implications of the main findings?
  • This paper identified a persistent “Northwest–Northeast dual-core” pattern of ice resources with strong spatial clustering (Global Moran’s I ≈ 0.56–0.57), with increasing ice in Tibet, Qinghai, Xinjiang, Inner Mongolia, and Northeast China, but decreasing ice limited to Beijing and Yunnan under strong urbanization and low-latitude warming.
  • Using a Random Forest model, this paper showed that water area fraction, nighttime lights, built-up area, altitude, and water–heat indices are the main controls on potential winter ice area, highlighting the combined effects of hydrological–thermal conditions and urbanization on ice patterns under climate change.

Abstract

River and lake ice are sensitive indicators of climate change and important components of hydrological and ecological systems in cold regions. In this study, we develop a simple and transferable “surface water + land surface temperature (LST)” framework on Google Earth Engine to map potential winter ice area across China from 1990 to 2020. The framework enables consistent, large-scale, long-term monitoring without relying on complex remote sensing models or region-specific thresholds. Our results show that, despite a pronounced northwestward shift in the freezing-zone boundary, more than 400 km in the Northeast Plain and about 13 km per year along the eastern coast, the total ice-covered area increased by approximately 1.1% per year. At the same time, the average ice season became slightly shorter. This indicates asynchronous spatial and temporal responses of potential winter ice to warming. We identify a persistent “Northwest–Northeast dual-core” spatial pattern with strong positive spatial autocorrelation, characterized by increasing ice cover in Tibet, Qinghai, Xinjiang, Inner Mongolia, and Northeast China, and decreasing ice cover mainly in Beijing and Yunnan, where intense urbanization and low-latitude warming dominate. Random Forest modeling further shows that water area fraction, nighttime lights, built-up area, altitude, and water–heat indices are the main controls on potential winter ice. These findings highlight the combined influence of hydrological and thermal conditions and urbanization in reshaping potential winter ice patterns under climate change.

1. Introduction

Lake and river ice are an important component of the cryosphere and play a critical role in the climate system. The Global Climate Observing System (GCOS) of the World Meteorological Organization (WMO) has identified lake ice as one of the Essential Climate Variables (ECVs) [1]. Recent estimates indicate that, at the time of maximum ice thickness, freshwater ice in the Northern Hemisphere covers about 1.7 × 106 km2 with a volume of roughly 1.6 × 103 km3 [2]. The areal extent is comparable to that of the Greenland Ice Sheet, and the volume is of the same order of magnitude as seasonal snow on land, underscoring the importance of lake and river ice within the global cryosphere [3].
Seasonally frozen water bodies are widespread at high latitudes. Ice-covered lakes occupy about 2% of global land area, and up to 40–50% in some polar regions [4,5,6]. Approximately 29% of global river length develops substantial seasonal ice cover, and 58% of river reaches are affected by ice for at least part of the year [7]. River-ice duration and melt regime, together with lake-ice phenology, jointly regulate water and energy budgets and runoff, reshape fluvial morphology and ecological patterns, and influence flood risk [8,9]. Understanding how these processes evolve under ongoing global warming has become a key scientific challenge.
Research on lake ice has mainly focused on ice concentration, spatial extent, and phenology [10,11]. Additional topics include discrimination of ice types, estimation of ice thickness and overlying snow depth, and retrieval of snow/ice surface temperature, in order to characterize lake water–heat exchanges and their response to climate warming [12,13]. In contrast, river-ice research, beyond identifying whether a reach is ice-covered and the timing of freeze-up and break-up, places greater emphasis on classification of ice types [14], detection of ice jams and their inundation extent [3,15], estimation of ice thickness [16,17], and derivation of surface flow velocity from ice-floe motion [18]. These parameters are crucial for flood-risk assessment and for coupled hydrodynamic–ice jam modeling. Overall, lake-ice and river-ice studies differ in their specific targets and metrics but share a common objective to improve understanding of how freshwater ice controls water and energy fluxes, river–lake morphological evolution, and associated ecological and hazard impacts.
Methodologically, a comprehensive monitoring framework based on multi-source remote sensing has been established. Optical sensors such as AVHRR, MODIS, and operational products such as NOAA IMS and ice charts from the Canadian Ice Service are primarily used to derive lake and river ice extent and phenology through thresholding, decision trees, and manual or semi-automated interpretation [19,20,21,22,23]. These data are well suited for large-scale, long-term analyses, but are constrained by cloud cover and polar night. Passive microwave radiometers and scatterometers use time series of brightness temperature and backscatter to automatically detect key phenological events on large lakes and major rivers, including freeze onset, melt onset, and complete ice-on/ice-off dates [24,25]. When combined with radar altimetry, they can retrieve the dates of initial ice formation, stable ice-cover establishment, and first open water at sub-weekly timescales, partly compensating for the scarcity of in situ observations [26,27]. Synthetic Aperture Radar (SAR), with its high spatial resolution and polarimetric capabilities, enables detailed mapping of multi-year residual ice, open-water patterns during thaw, and river-ice types and ice jams [28]. Automated processing chains such as IceMap-R can directly output multiple ice classes and convert them into roughness zones required by hydrodynamic models [29]; polarimetric SAR has also been used to develop empirical models for ice-thickness retrieval [30]. Airborne thermal infrared imaging, combined with one-dimensional heat-conduction modeling, can be used to estimate thin-ice thickness and extract floe geometries, demonstrating potential for future satellite-based thin-ice monitoring [31]. High-resolution stereo optical imagery can treat ice floes as tracer particles, whose displacement is used to derive surface flow velocity, providing new data to validate and constrain river hydrodynamic models [32]. As a result, freshwater-ice research is shifting from single-indicator, single-sensor monitoring toward integrated observation and quantitative inversion of ice phenology, ice type, ice thickness, and ice dynamics within multi-source data-fusion frameworks.
Although existing freshwater-ice monitoring studies indicate that the technical toolbox for fine-scale mapping and process-based analyses is relatively mature, particularly polarimetric SAR approaches for ice-type discrimination and thin-ice information retrieval, airborne thermal infrared observations that capture ice–water temperature contrasts, and high-resolution stereo imagery that characterizes ice-surface morphology and dynamic features, these studies simultaneously highlight practical bottlenecks to developing national-scale, operational products. Long-term observations and statistical analyses show that river- and lake-ice phenology is highly sensitive to air-temperature change and large-scale circulation teleconnections, exhibiting pronounced interannual to decadal variability and clear long-term trends. Ice cover also modulates air–water heat exchange and thereby influences regional climate coupling, underscoring the need for monitoring systems with broad coverage and long-term continuity [33,34,35,36]. Nevertheless, these fine-scale approaches generally entail high technical barriers, large data volumes, and substantial computational and financial costs, making them better suited to intensive studies and engineering applications in representative lakes or priority river reaches.
For nationwide inventories of river- and lake-ice resources, existing efforts remain insufficient in spatial and temporal coverage, long-term continuity, and cross-regional transferability, limiting the development of stable and portable baseline datasets at macroscales. On the one hand, optical time series can be used to track breakup/freeze-up processes and phenology, but persistent cloud cover, polar night/low solar elevation, and mixed pixels make it difficult to produce robust, large-area, consistent long-term products [37,38]. On the other hand, SAR provides all-weather capability and polarimetric information can enhance ice–water and river-ice type discrimination; however, backscatter is highly sensitive to incidence angle, snow cover, liquid-water content, and surface roughness, resulting in limited cross-regional transferability of thresholds and feature representations [32,39,40]. Moreover, although retrievals of ice thickness, bottom-fast ice, and ice-regime characteristics have been demonstrated in specific regions, they depend strongly on prior information and scene conditions, and contrasting ice regimes, such as bottom-fast versus floating ice, alter remote-sensing responses and inference pathways, further complicating the development of a “nationally universal” model [41,42]. Meanwhile, conventional ice–water classification methods based on a single optical spectral index or a single SAR backscatter metric typically target individual water bodies and often require repeated parameter tuning and algorithmic optimization across regions, making them ill-suited for generating nationally uniform, comparable long-term products. Collectively, these limitations indicate that key technical barriers remain for establishing a unified, continuous, and reusable national-scale monitoring and assessment framework for freshwater-ice resources.
To address these gaps, this study proposes an integrated approach for extracting potential winter ice resources based on the joint use of satellite-derived surface water products and land surface temperature (LST), targeting national-scale, long-term freshwater-ice monitoring. The method builds on publicly available, globally consistent, and regularly updated water and temperature products. Long-term water-body datasets are first used to delineate the spatial extent of potentially ice-bearing water bodies. The freezing state of these water bodies is then inferred from the LST field, thereby directly deriving the spatiotemporal distribution and evolution of potential winter ice on a common water-body baseline. This strategy reduces dependence on complex scattering models, texture analysis, and polarimetric decomposition, resulting in a streamlined method with transparent parameterization that can be transferred across different climate zones and surface types in China, and lowering technical and computational barriers for large-scale applications. In addition, superimposing temperature-based freeze/thaw discrimination on a unified water mask facilitates the construction of temporally consistent multi-year time series of potential winter ice, enabling systematic quantification of key “ice resource” indicators such as ice-covered area, freeze frequency, and ice-season length. This represents a shift from “local process monitoring” toward building a “national-scale baseline of freshwater-ice resources.”

2. Materials and Methods

2.1. Identification of Potential Winter Ice Resources Using Remote Sensing Data

In cold-region environmental research, identifying frozen environments using near-surface temperature thresholds and, in parallel, analyzing the long-term distribution of water bodies to delineate “ice carriers”, key hydrological units that support ice formation and persistence, such as lake margins and river reaches, constitutes a well-established and widely adopted methodological framework in remote sensing and hydrology [3,23]. The core principle is to retrieve land surface temperature (LST) and apply a threshold of 0 °C or lower to infer potentially frozen land surfaces and water bodies, thereby characterizing freeze–thaw dynamics [43]. Building on this foundation, to further resolve the spatial distribution, timing, and evolution of potential winter ice, multispectral remote sensing such as Landsat and Sentinel 1–3 is commonly used for finer ice–water classification and long-term time-series analyses, with indices such as the Normalized Difference Snow Index (NDSI) or lake-ice indices employed to discriminate ice-covered surfaces from open water [44,45]. In this study, “ice resources” are defined as the potential environments and physical carriers that can support the formation of stable water-body ice cover at a given spatiotemporal scale. Specifically, a freezing-zone threshold is used to characterize the potential atmospheric freezing environment associated with sustained cold-season conditions, thereby reducing the influence of short-lived cold-air events on boundary delineation and producing a more stable, national-scale freezing-zone boundary. Permanent water bodies are treated as the physical carriers capable of generating potential winter ice, and are used to constrain the spatial extent of water bodies where potential ice cover may occur.
This study integrates multi-source remote sensing datasets on the Google Earth Engine (GEE) platform to identify river and lake ice resources. Air-temperature information is obtained from the MOD11A1.061 Terra Land Surface Temperature and Emissivity Daily Global 1 km dataset (temporal resolution: daily; spatial resolution: 1 km) developed by NASA LP DAAC at the USGS EROS Center [46], and the Climate Hazards Center InfraRed Temperature with Stations (CHIRTS) daily temperature data product (temporal resolution: daily; spatial resolution: 0.05°) provided by the University of California [47]. We define freezing zones as areas in which, within any 7-day window, there are at least 5 consecutive days with either minimum air temperature below 0 °C or maximum air temperature below 0 °C. The boundary between freezing and non-freezing areas is delineated as the “freezing line,” representing the potential spatial extent of ice-bearing environments. Surface-water information is derived from the Global Surface Water Explorer (JRC Monthly Water History) dataset (temporal resolution: monthly; spatial resolution: 30 m) provided by Google [48], which characterizes monthly variations in global surface water from 1984 to 2021. Permanent water bodies extracted from this dataset are treated as the physical carriers of potential lake or river ice resources. To ensure consistency, all raster datasets are resampled and harmonized to the 30 m spatial resolution of the JRC product.
The JRC “Permanent Water” constraint implicitly assumes persistent, continuously present surface water, and therefore cannot capture the pronounced hydrological variability that characterizes intermittent and seasonal/ephemeral river systems [49]. Nevertheless, in the context of this study on winter rivers in China, the primary value of the JRC Global Surface Water (GSW) dataset lies not in representing such variability per se, but in providing a stable, nationally consistent multi-year baseline and a reproducible framework for spatiotemporal statistics. Specifically, GSW offers long-term, methodologically consistent information at the national scale that can serve as background constraints and prior information for winter water-surface delineation and change analysis. This reference is particularly robust over open-water settings, including wide river reaches, main stems, and hydraulically influenced areas such as lake/reservoir–river confluences and backwater zones.
Building on these datasets, we overlay the freezing zones with the permanent-water layer to identify water bodies that lie entirely within persistently freezing areas; these are defined as lake or river ice resources (Figure 1). Taking Qinghai Lake, the largest inland saline lake in China, as a representative example, we first extract its permanent-water boundary from the JRC Global Surface Water product, obtaining a clear shoreline morphology. It should be noted that the freezing threshold defined in this study is used to characterize the potential atmospheric freezing environment that can support the formation of ice cover, thereby delineating the spatial domain for identifying river and lake ice resources, rather than to directly infer the actual freeze–thaw dynamics of an individual water body. To derive a more robust national-scale freezing-zone boundary, we further employ a moving temporal window and consecutive-day criteria to suppress noise from short-lived cold-air outbreaks.
Figure 1. Schematic diagram of river and lake ice resource extraction.
Fourteen well-known lakes distributed across northern China, Northeast China, Northwest China, and the North China Plain were selected (Table 1). Relevant statistical records were consulted, and the lake areas reported in official statistics were treated as reference values. These values were compared with the lake water surface areas extracted in this study to assess the accuracy of the surface-water extraction results. Overall, the extracted areas showed good agreement with the statistical records, yielding an area-weighted overall accuracy of approximately 89.19%. For several lakes, including Qinghai Lake and Tianchi (Changbai Mountain), the errors were small, with relative errors of 0.75% and 2.67%, respectively, indicating that the proposed method performs robustly for large water bodies with relatively clear boundaries. However, for lakes/reservoirs with complex shorelines or strong sensitivity to water-level and seasonal variability, such as the Miyun Reservoir, Jingyue Lake, and the Panjiakou Reservoir, the relative errors increased to 30–38%. In summary, the method can reliably delineate the extent of major lakes in the study area, while further improvements are still needed to enhance extraction accuracy for small- to medium-sized and more dynamically varying water bodies.
Table 1. Accuracy assessment of surface-water extraction for typical lakes.

2.2. Trend Analysis and Classification

The time series analyzed in this study may exhibit nonlinearity, regime, and mechanism shifts, and nonstationarity, under which traditional parametric approaches like ARIMA typically rely on strong stationarity assumptions and require substantial manual specification and tuning, making it difficult to consistently characterize complex dynamics across heterogeneous samples. By contrast, an LSTM autoencoder can learn long-range dependencies and nonlinear temporal structure directly from the data and, by minimizing reconstruction error, perform unsupervised representation learning to produce a low-dimensional latent embedding that summarizes each sequence’s overall evolutionary pattern. This embedding is well suited to K-means clustering driven by pattern similarity, while reducing reliance on ad hoc feature engineering and restrictive model-form assumptions. Meanwhile, to ensure interpretability and statistical transparency, trend quantification is conducted using robust, interpretable estimators, linear regression and the Theil–Sen slope, on the original series or cluster-aggregated series, so that the deep learning component primarily supports representation learning and pattern discovery, whereas trend magnitude and direction remain clear and statistically testable.
The LSTM-Autoencoder + K-means + trend-analysis framework is an integrated time-series approach that combines deep learning-based feature extraction with traditional statistical methods. A Long Short-Term Memory Autoencoder (LSTM-Autoencoder) is first applied to time series for automatic feature learning [50]. The encoder captures long-term dependencies and nonlinear variability in the sequences and produces a low-dimensional latent representation at the bottleneck layer, thereby compressing information on complex dynamics and extracting the most salient features.
These latent features are then input to a K-means clustering algorithm to identify groups of samples with similar temporal evolution patterns. Finally, linear regression and Theil–Sen slope estimation are used to quantify the trends of individual samples or clusters and to classify them according to their temporal change characteristics (Figure 2).
Figure 2. LSTM-Autoencoder + K-means + trend-analysis framework.

2.3. Random Forest Regression and Analysis of Influencing Factors

2.3.1. Selection of Influencing Factors

The formation and evolution of winter river and lake ice are jointly controlled by a range of natural and anthropogenic factors. Demographic, economic, land-use, and eco-environmental conditions shape regional energy balance and hydrological processes, and thus directly or indirectly affect ice formation, melt, and stability [8,51,52].
Socioeconomic indicators such as permanent population and GDP reflect the intensity of human activities, which influence ice mainly through thermal disturbances and pressure on water resources [53]. Topographic factors, including slope and elevation, regulate water flow and the spatial pattern of air temperature, thereby modifying conditions for ice accumulation [54]. Ecological indicators such as NDVI, NPP, and forest and grassland cover alter local climate and water-retention capacity, and thus help support the persistence of ice. Precipitation and land surface temperature, as primary climatic controls, determine both the material input to ice and its surface energy balance [55,56]. In addition, indicators such as nighttime lights, built-up area, and surface-water area characterize urbanization and environmental modification, which strongly influence the spatial distribution and stability of ice [57]. To more comprehensively represent these relationships, this study also constructs composite indices that integrate hydrothermal, ecological, and socioeconomic information, with the aim of elucidating the combined mechanisms through which multidimensional factors control winter river and lake ice (Table 2).
Table 2. Primary influencing factors.
Based on multicollinearity diagnostics, using VIF < 10 as the retention criterion, two original predictors, precipitation (Precip) and the proportion of cropland area (Cropland), were excluded due to multicollinearity concerns. To avoid reintroducing collinear information via derived terms, all precipitation-related composite features were also removed, specifically the hydrothermal product (Precip times LST) and the simple hydrothermal ratio (Precip div LST). All remaining indicators met the VIF < 10 threshold and were retained for subsequent modeling.

2.3.2. Construction of the Random Forest Model

The training of the random forest model is based on the Bagging (Bootstrap Aggregating) approach. The main steps are as follows:
Sample selection: Subsamples are randomly drawn with replacement from the training set and used to construct each individual decision tree.
Feature selection: At each node split, a random subset of predictors is sampled, typically of size equal to the square root of the total number of features, from which the best splitting variable is selected. This increases model diversity.
Decision-tree construction: Multiple trees are built using a standard decision-tree algorithm, and each tree makes predictions independently.
A trimming procedure was applied to the raw data by removing the top 1% of extreme observations. A two-stage random hold-out sampling design (hold-out = 0.3) was then implemented. After feature construction and z-score standardization of both predictors and the response, the full dataset was randomly partitioned into training and test subsets at a 70/30 ratio, and an initial bagging regression model was fitted to the training subset. Finally, a second random 70/30 hold-out split was conducted for model training and independent test evaluation, and optimal hyperparameters were selected via grid search with five-fold cross-validation on the training data.

2.3.3. Model Parameter Optimization

The model hyper parameters, namely the number of trees ( n _ t r e e s ) and the minimum leaf size ( m i n L e a f S i z e ), are optimized using grid search combined with five-fold cross-validation. The cross-validation procedure provides an evaluation of the model’s generalization performance and helps to avoid overfitting.

2.3.4. Assessment of Variable Importance

Variable importance is evaluated using the Out-of-Bag (OOB) approach combined with feature randomization. For each predictor, its importance is quantified according to the following formula:
I m p o r t a n c e ( j )   = M S E o r i g i n a l − M S E p e r m u t e d
where M S E o r i g i n a l is the mean squared error computed when the model uses the original feature set, and M S E p e r m u t e d is the mean squared error recalculated after the given feature has been randomly permuted. This procedure makes it possible to identify those predictors that exert the greatest influence on the dependent variable.

2.3.5. Partial Dependence Plots (PDP)

To further understand how key features affect the target variable, we generated Partial Dependence Plots (PDPs). The PDP function visualizes the influence of a specific feature on the model’s predicted output by varying the value of that feature while holding all other features fixed. In this way, it reveals potential nonlinear relationships between predictors and the response.
P D P ( X j ) = 1 N ∑ i = 1 N f ( X j , X − j ( i ) )
where f denotes the random forest model, X j s the specific feature of interest, X − j ( i ) is the feature vector of the i-th sample with that feature excluded, N is the total number of samples.

3. Results

3.1. Analysis of Changes in River and Lake Ice Resources in China over the Past 30 Years

3.1.1. Northwestward Shift in the Freezing-Zone Boundary

Using time-series land surface temperature data, we constructed a freezing-area identification model and extracted the freezing-zone boundary across China to characterize freezing regions over the past 30 years. The results indicate a pronounced northwestward shift in the freezing-zone boundary, with the magnitude of this shift gradually increasing from west to east. The maximum shift displacement exceeds 400 km and occurs mainly in the central and southern parts of the Northeast Plain. Along the eastern coast, the freezing-zone boundary has shifted at an average rate of about 13 km per year.
In 2020, a total of 3277 large-scale outdoor natural ice resources were identified, covering 65,518 km2 and distributed across northeastern, northern, eastern, northwestern, and southwestern China. These resources are predominantly small, with individual areas of 1–2 km2 accounting for approximately 38.79% of the total area, whereas resources larger than 50 km2 represent about 6.35%. The spatial distribution of ice resources shows a clear clustering pattern. Large clustering zones, here termed primary clusters, are located in Northeast China and Tibet. In addition, three smaller clustering zones, termed secondary clusters, are found in Xinjiang, Inner Mongolia, and Beijing (Figure 3).
Figure 3. Spatial distribution of river and lake ice resources in China.

3.1.2. Changes in Ice Resource Area and Regional Classification

Using annual remote sensing data on permanent surface water bodies, we extracted river and lake ice resources in China over the past 30 years. At the national scale, the area of ice resources has increased markedly during this period. The total ice-resource area was 50,016.77 km2 in 1991 and 68,955.81 km2 in 2020, corresponding to an average annual growth rate of about 1.1%. At the provincial level, the ice-resource areas, ranked from largest to smallest, are as follows (Figure 4): Tibet > Qinghai > Xinjiang > Inner Mongolia > Heilongjiang > Jilin > Liaoning > Gansu > Hebei > Tianjin > Sichuan > Beijing > Shanxi > Shandong > Ningxia > Shaanxi > Yunnan.
Figure 4. Temporal variation in ice resource area nationwide and by province.
To better understand the temporal dynamics of river and lake ice resources over the past 30 years, we use an LSTM–Autoencoder deep learning model to extract latent features from the time series (Figure 5). Combined with K-means clustering and trend-estimation methods, including linear regression and Theil–Sen slope, we quantitatively characterize changes in river and lake ice area across major provinces in China.
Figure 5. Model training results.
The network has eight layers in total (including input and output), with 20 hidden units in the recurrent layer (hidden size = 20). The latent space is represented by a 6-dimensional vector (latentDim = 6). The model is trained for up to 250 epochs (MaxEpochs = 250), with 500 iterations and a mini-batch size of 8 (MiniBatchSize = 8). Mean Squared Error (MSE), implemented via a regression layer, is used as the loss function to measure the reconstruction error between the input and reconstructed output sequences (Table 3).
Table 3. Key Parameters of Trend and Classification Analysis Models.
The results indicate that over the past 30 years, river and lake ice resources in China have generally increased, with overall stable growth and localized decreases in some regions.
The significance tests indicate that, among the 18 provinces, 10 provinces exhibit trend-analysis p-values below 0.01, demonstrating statistical significance at the 99% confidence level. The remaining provinces have p-values below 0.05, indicating statistical significance at the 95% confidence level.
The ice-rich provinces in China can be divided into two categories (Figure 6). The first category (Cluster 1) is the increasing type, including Tibet, Qinghai, Xinjiang, Inner Mongolia, Heilongjiang, Jilin, Liaoning, Gansu, Hebei, Tianjin, Sichuan, Shanxi, Shandong, Ningxia, and Shaanxi. In these provinces, the ice-covered area shows an overall increasing trend, with both the linear and Theil–Sen slopes being positive, indicating a general expansion of ice resources under climatic forcing over the past 30 years. The increase is most pronounced in high-cold and arid regions. For example, the linear slope for Tibet Autonomous Region is 3.82 × 102 km2/yr, Theil–Sen estimate is 3.24 × 102 km2/yr, for Qinghai Province it is about 1.46 × 102 km2/yr, Xinjiang and Inner Mongolia are approximately 87 km2/yr and 47 km2/yr, respectively, both showing a sustained growth trend. In mid-latitude regions such as Gansu, Ningxia, Shanxi, Hebei, and Sichuan, the increase is relatively modest (about 3–6 km2/yr), but still dominated by a steady upward trend. Although the northeastern provinces (Heilongjiang, Jilin, Liaoning) exhibit some fluctuations, their overall trend is also upward. The second category (Cluster 2) is the “decreasing type,” consisting of Beijing and Yunnan. In Beijing, the linear slope is −2.03 km2/yr, and the Theil–Sen estimate is −2.24 km2/yr, indicating a continuous decline in ice-covered area over the past three decades. In Yunnan Province, the linear slope is about −0.0074 km2/yr; although the magnitude of change is smaller, the overall trend is consistent.
Figure 6. Classification of provinces based on trends in river and lake ice resource area.

3.1.3. Changes in the Ice-Covered Period over the Past 30 Years

At the national scale, the average ice-covered period shows a slight decreasing trend of about −0.005 month/year, equivalent to a reduction of roughly 0.05 month per decade. In the early 1990s, the ice-covered period was generally stable across regions. From the 2000s to the 2010s, however, it shortened markedly in most areas, with particularly pronounced decreases in North China and Southwest China. Since the late 2010s, a fluctuating “warming-back” signal has appeared in parts of the Qinghai–Tibet Plateau and some northeastern provinces, suggesting that, despite overall climate warming, short-term climate variability and regionally differentiated adjustments persist in parts of China (Figure 7).
Figure 7. Changes in the ice duration by province.
The range of changes among provinces lies between −0.020 and +0.035 month/year. On the Qinghai–Tibet Plateau, the ice-covered period remains around 7 months overall, showing a slight increasing trend. In Tibet, the average ice-covered period increased from 6.94 to 7.03 months, indicating a modest upward trend of +0.003 month/year; in Qinghai Province, it rose from 6.97 to 7.10 months, with a slightly higher rate of increase about +0.004 month/year. In Northeast China and along the northern boundary of North China, the ice-covered period generally shows slight shortening or fluctuating stability. In Heilongjiang Province, the overall trend is relatively stable but with pronounced fluctuations, with a maximum variation of about 1.3 months. Jilin Province exhibits clear periodic fluctuations with a slight increase in the middle and later stages, whereas Liaoning Province and Inner Mongolia both show a persistent decreasing trend of −0.006 month/year. In Northwest China, the ice-covered period as a whole shows a pattern of “early increase followed by later stabilization”. In Xinjiang, the ice-covered period lengthened during the middle of the study period and then slightly declined in the later stage, reflecting distinct phase-dependent variability. In Gansu Province and Ningxia, the ice-covered period increased slightly by 0.1–0.2 month, forming a weak, nonlinear upward trend. In the southern and eastern provinces, the ice-covered period generally shows a pronounced shortening trend. Sichuan Province is the most notable, with a linear decreasing trend of about −0.008 month/year. In Yunnan Province, the ice-covered period remains around 5 months with relatively stable fluctuations, but with a slight increase in the past decade. In Henan, Hebei, Shandong, and Shanxi, the ice-covered period is stable or slightly decreasing, with long-term averages of about 3–5 months.

3.2. Spatial Distribution Characteristics of River and Lake Ice Resources in China over the Past 30 Years

3.2.1. Spatial Distribution Patterns and Their Changes

From 1990 to 2020, the spatial distribution of China’s lake and river resources remained broadly stable, with localized areas of contraction. The northwest and northeast have consistently functioned as the country’s core cluster regions for ice resources, forming a relatively stable “northwest–northeast dual-core” spatial structure (Figure 8). In contrast to this overall stable macro-pattern, parts of Shandong, Hebei, and Inner Mongolia experienced a notable reduction in ice-resource areas.
Figure 8. Spatial distribution of river and lake ice resources in China over the past 30 years.

3.2.2. Spatial Autocorrelation Characteristics and the Changes

In 1990 and 2020, the Global Moran’s I values for ice resource area were 0.5647 and 0.5677, respectively, both clearly higher than the theoretical expected value (about −0.000007). The corresponding z-values reached 479.3633 and 476.2561, with p-values both less than 0.001 (Table 4). These results indicate that in both periods, the ice-covered area exhibited a significant positive spatial autocorrelation, meaning that high-value areas tend to cluster spatially with other high-value areas, and low-value areas with other low-value areas, rather than being randomly distributed.
Table 4. Spatial autocorrelation statistics.
Meanwhile, Moran’s I increased only slightly from 0.5647 to 0.5677 between 1990 and 2020, suggesting that the overall intensity of spatial clustering of ice-covered area remained basically stable over the three decades, with a marginal strengthening trend. This reflects a strong spatial lock-in effect and persistent regional agglomeration in China’s ice resource pattern at the macro scale.

3.3. Analysis of Influencing Mechanisms

3.3.1. Model Performance Evaluation

Using a spatially explicit evaluation strategy, the study assessed model performance under repeated spatial block cross-validation to reflect spatial transferability. Specifically, samples were partitioned into spatial blocks based on longitude and latitude (quantile binning, 6 × 6); in each repetition, approximately 30% of blocks were held out as an external test set, while the remaining blocks were used for model training. Within the training data, 5-fold spatial block CV was further applied for hyperparameter tuning of the random forest, number of trees n T r e e s n T r e e s , minimum leaf size M i n L e a f S i z e , and the number of predictors sampled at each split m t r y . Because the dependent variable is non-negative and highly skewed, modeling was conducted in the transformed space Y ’ = l o g ( 1 + Y ) , and predictions were back-transformed to the original units for reporting.
Across 20 repetitions, the model achieved a mean test R 2 of 0.417 in the original Y scale (median = 0.444, P5–P95 = 0.115–0.813), with a median test RMSE of 5.24 × 108 (P5–P95: 1.63 × 108–3.23 × 109). In the transformed log-standardized space, the mean test R2 was 0.603 (median = 0.638, P5–P95 = 0.397–0.734), indicating improved stability under spatial extrapolation. The most frequently selected hyperparameter combinations concentrated on M i n L e a f S i z e = 2, m t r y = 18, and n T r e e s = 300 or 800. Overall, the spatial CV results suggest that the model provides moderate predictive skill with substantial spatial heterogeneity, which is expected for coupled natural–socioeconomic systems.

3.3.2. Analysis of Variable Importance Characteristics

Permutation importance indicates that the spatial variability of winter river- and lake-ice area is associated with a combination of hydroclimatic conditions, topography, ecosystem attributes, and human-activity proxies. Table 5 lists the ten most influential predictors ranked by permutation importance.
Table 5. Top 10 primary driving factors.
It is worth noting that proportion of surface water ranks first. This result is expected because ice area is physically constrained by the presence and extent of open water, that is, water acts as a baseline factor that bounds the maximum ice area within a grid cell. Therefore, beyond this necessary condition, the subsequent predictors such as temperature–moisture regime, topography, and human-activity indicators, provide additional explanatory power for the remaining variability in ice area.
Among these (Figure 9), human-activity proxies, NTL (0.0331), NTL × Built (0.0197), and Built (0.0040), rank prominently, indicating that stronger urbanization signals are associated with reduced winter ice area, plausibly through enhanced anthropogenic heat release and urban heat-island effects that increase surface warming and accelerate ice ablation. Topographic controls, especially elevation (Elev) (0.0086) and topographic exposure (Slope × Elev) (0.0030), also contribute, which is consistent with colder conditions at higher elevations and increased runoff/energy exposure in more rugged terrain that is unfavorable for ice accumulation. Ecosystem indicators show smaller but non-negligible effects: NDVI (0.0055) and NPP (0.0044) suggest modest buffering via moisture retention and regulation of local water–heat balance, while NDVI_times_Forest (NDVI × Forest) (0.0029) captures the combined extent and condition of forest cover, where greater canopy shading and moisture conservation can help maintain cooler, more humid microclimates that support ice stability. In addition, LST (0.0032) indicates that higher land-surface temperatures increase available heat, accelerating melt and shortening the ice season.
Figure 9. Importance of influencing factors.
It should be emphasized that permutation importance reflects contributions to predictive performance rather than causal effects. In particular, correlated predictors such as NTL, Built, and NTL × Built may share overlapping information; their rankings are therefore better interpreted as joint associations with ice area rather than independent effect sizes.

4. Discussion

This study systematically characterizes the spatiotemporal evolution of river and lake ice resources in China over the past 30 years at the national scale and analyzes the main driving mechanisms from the dual perspectives of the natural environment and human activities. Overall, winter river and lake ice resources in China show four key features: a marked northwestward shift in the freezing-zone boundary, a general increase in ice area, a slight shortening of the freezing period, and a spatial pattern that remains broadly stable but undergoes local reorganization. These findings both respond to the broader context of global warming and reflect the complex coupling among regional water-body configurations, human activities, and topographic–ecological conditions.

4.1. Spatiotemporal Trends in Changes in River and Lake Ice Resources in China

From the perspective of the overall freezing environment, the freezing-zone model based on land surface temperature shows that over the past 30 years China’s freezing-zone boundary has clearly shifted northwestward, with the magnitude of northward movement gradually increasing from west to northeast. In the central and southern parts of the Northeast Plain, the maximum northward shift exceeds 400 km, and along the eastern coast the northward migration rate is about 13 km/year. This is generally consistent with previous studies on the northward shift in winter isotherms and the reduction in freezing days in China [58,59,60]. However, despite the overall northwestward movement of the freezing-zone boundary, the total area of river and lake ice resources nationwide increased from about 50,000 km2 to nearly 69,000 km2 between 1991 and 2020, with an average annual growth rate of around 1.1%. Under ongoing global warming, an increase in China’s surface water area is supported by evidence from multi-scale observations and mapping products [61,62]. National wetland remote-sensing inventories indicate that total wetland area in China underwent substantial contraction prior to 2015, followed by a rebound during 2015–2020, alongside an expansion of open-water extent of approximately 14,000 km2 [61]. Consistent with these findings, high-resolution surface-water mask analyses for inland basins show that the total surface-water area across 17 inland basins in China increased by 38.8% over 1987–2020 [62]. With respect to underlying drivers, existing studies identify a clear pattern of regionally differentiated attribution. In western basins, such as the inland river basins of northern Tibet, the upper Yellow River, the upper Yangtze River, and inland rivers in Xinjiang, climate change is the dominant contributor to surface-water expansion, approximately 51–61%. This attribution aligns with the well-documented climatic “warming and humidification” observed in Northwest China over recent decades, characterized by persistent warming, increasing precipitation, declining aridity, and a humidification trend that exhibits stage-wise, nonlinear intensification and eastward expansion [62,63,64,65]. This apparent paradox of “spatial expansion and temporal compression” suggests that the response of ice resources to climate warming is not synchronized in space and time. It further suggests that under China’s warming-and-wetting climate trend, the expansion of water bodies driven by increased humidity contributes more to the growth of winter river and lake ice area than the loss of ice caused by the warming-induced shrinkage of freezing regions. Note that a larger ice area does not necessarily mean better ice resources; under overall warming, the duration and thickness of ice are still declining [66,67]. In this study, winter ice area is essentially defined as the spatial intersection between temperature-based freezing conditions and permanent surface water bodies. Accordingly, the observed increase in ice area primarily indicates an expansion of potential ice-bearing surfaces over stable lakes/reservoirs and persistently wetted river reaches. At the same time, this approach likely underestimates contributions from seasonal or episodic water surfaces that can freeze but are excluded by the permanent-water constraint such as seasonal floodplain inundation, temporary tributaries or channels, and short-lived shoreline expansions during wetter years.
At the level of spatial configuration, China’s river and lake ice resources have long maintained a macro-structure characterized by a “northwest–northeast dual core”: the northwestern plateau and arid regions, together with the cold-temperate zone in Northeast China, form the two major high-value clusters of ice resources. The Global Moran’s I values for ice area in 1990 and 2020 are 0.5647 and 0.5677, respectively, both well above the random expectation and only slightly higher after 30 years. This indicates that, although local ice area and distribution have adjusted, the large-scale spatial clustering pattern has remained broadly stable, showing a pronounced “spatial lock-in effect.” This stability mainly arises from long-term constraints of topography and climatic zoning: the high elevation, high latitude, and cold–dry climate of the Tibetan Plateau and Northeast China provide a robust physical backdrop for ice formation [54,68,69], while the abundance of lakes, reservoirs, and river networks in these regions supplies sufficient water-body carriers [70,71]. By contrast, provinces with moderate ice resources located in climatic transition zones and regions of rapid urbanization, including Shandong, Hebei, and parts of Inner Mongolia, have experienced some contraction in ice area.
From the perspective of temporal trajectories and regional classification, the time-series clustering method that combines an LSTM–Autoencoder with K-means divides the evolution of provincial-level ice resources into two categories: increasing type and decreasing type. Most ice-rich provinces and regions (such as Tibet, Qinghai, Xinjiang, Inner Mongolia, and the three northeastern provinces) fall into the increasing type, with ice area continuing to grow between 1990 and 2020, especially on the Tibetan Plateau and in the arid northwest, where the increase is most pronounced. This may be related, first, to lake expansion, reservoir construction, and changes in the stability of seasonal water bodies under the Elevation-Dependent Warming (EDW) effect of global warming in high-altitude regions [72,73]; and second, to the relatively intact natural ecosystems and low urbanization levels in these regions, which weaken the heat-island effect and anthropogenic heat emissions so that rising temperatures have not yet translated into substantial reductions in ice area. It should be emphasized that an increase in area does not necessarily imply thicker or safer ice; ice thickness, structure, and load-bearing capacity may still be declining.
In contrast, Beijing and Yunnan are classified as decreasing-type provinces, with ice area showing an overall decline over the past 30 years. Beijing, as a highly urbanized region, has experienced rapid expansion of built-up areas and intensified nighttime lighting, which have substantially altered the surface energy balance. The urban heat-island effect, together with the increase in impervious surfaces, has led to a rapid retreat of the freezing environment. Yunnan, located in a low-latitude, high-radiation zone, was already close to the climatic threshold for freezing; under the combined influence of climate warming and economic development, its marginal ice resources are more sensitive to temperature increases and therefore exhibit continuous shrinkage.

4.2. Important Influencing Factor Analysis of River and Lake Ice Resources

The permutation-based importance analysis suggests that winter ice area is associated with multiple interacting factors. The high importance of Water primarily reflects a physical constraint: within a given spatial unit, ice area cannot exceed the available open-water extent. Accordingly, we interpret Water as a bounding variable rather than a standalone “dominant driver” of interannual change.
After accounting for water availability, the relatively high contributions of LST-related and hydro-thermal coupling predictors such as Precip × LST are consistent with the role of temperature and moisture conditions in shaping the likelihood and persistence of ice cover. Topographic variables including Elev and Slope × Elev likely act through elevation-dependent climate gradients and exposure-related energy and wind regimes, which can modulate local freezing conditions.
The prominent ranking of urbanization proxies including NTL, Built, and NTL × Built indicates that anthropogenic land-surface modification and associated warming signals are systematically linked to reduced winter ice area. However, because these predictors are partly correlated and may also co-vary with other geographic factors such as regional climate and water distribution, the present importance analysis should be interpreted as evidence of association rather than direct causation. Further analyses controlling for spatial dependence and predictor redundancy including spatial cross-validation, correlation screening, and alternative importance measures, would help strengthen causal interpretation. Ecosystem indicators including NDVI and NPP show lower but measurable contributions, suggesting that vegetation and ecosystem productivity may covary with local hydro-thermal conditions and thus indirectly relate to ice persistence.

4.3. Limitations and Future Research Directions

At the methodological level, the simplified framework proposed in this study, combining long-term surface-water products with LST-based ice detection, shows good applicability and transferability for national-scale, long-term monitoring of ice resources. Compared with fine-resolution approaches that rely on multi-polarization SAR, airborne thermal infrared data, or high-resolution stereo imagery, our method has a clearer structure, more transparent parameters, and uses more easily accessible datasets, making it suitable for integration into national- or global-scale ice assessments. The combined use of an LSTM-Autoencoder and K-means clustering allows automatic extraction of nonlinear and phase-type change patterns from complex time series, rather than being limited to simple linear-trend analysis, while the random forest model offers an explanatory framework that captures nonlinearities and interactions among multi-source, multi-scale natural and socio-economic drivers.
Nonetheless, several limitations and uncertainties remain. First, the JRC product is available only at a monthly temporal resolution, which hampers the detection of short-term changes in ice extent and is insufficient for accurately characterizing thin ice or small ice bodies. Second, freezing thresholds derived from air temperature are essentially climatological approximations and do not explicitly account for local conditions such as water depth, flow velocity, wind speed, and salinity. Meanwhile, the importance analysis provides a data-driven overview of predictors linked to ice-area patterns, while the underlying mechanisms require corroboration with additional process-based or causal analyses. Third, this study focuses on the contributions of key influencing factors to changes in ice-resource area and provides a qualitative discussion of possible mechanisms; however, it did not further conduct deeper quantitative simulations to explicitly model these complex, nonlinear mechanisms. Within the existing national-scale framework, future research should incorporate higher-resolution observations of ice phenology and ice thickness in representative regions, and conduct multi-source cross-validation of ice-resource maps by integrating ground-based monitoring with physically based process models.

5. Conclusions

This study examines river and lake ice resources in China from 1990 to 2020 and proposes a national-scale framework for ice identification and analysis based on permanent surface water and land surface temperature (LST). Using LSTM-Autoencoder, K-means clustering, and random forest regression, the study draws the following key conclusions:
(1)
China’s freezing environment has warmed substantially, driving a marked northwestward shift in freezing-zone boundary.
Freezing-zone identification based on LST reveals a pronounced displacement of the freezing-zone boundary over the past 30 years, with larger shifts from western China toward to the Northeast. In the central–southern Northeast Plain, the maximum northward shift exceeds 400 km, and along the eastern coast the mean migration rate is about 13 km/year indicating a strong warming-driven reorganization of winter freezing boundaries.
(2)
Winter ice area has expanded overall, while the freezing period has slightly shortened, and the two changes are not spatially synchronized.
National ice area increased from 50,016.77 km2 (1990) to 68,955.81 km2 (2020), with a growth rate of 1.1% every year. Meanwhile, the national mean freezing period shows a small decreasing trend (about −0.005 months/year), provincial trends −0.020 to +0.035 months/year. High-cold, high-altitude regions like the Tibetan Plateau are generally stable or slightly longer, whereas North China, Southwest China, and several low latitude provinces exhibit more pronounced shortening. Under the warming-and-wetting trend, water-body expansion contributes more to ice area growth than warming-induced losses from shrinking freezing regions.
(3)
Ice resources exhibit a stable, strongly clustered “northwest–northeast” dual core spatial structure.
Ice resources remain concentrated in the northwestern plateau or arid regions and the cold-temperate Northeast. Global Moran’s I is 0.5647 (1990) and 0.5677 (2020), indicating persistent strong positive spatial autocorrelation and a little change in clustering intensity, consistent with combined controls of topography and climatic zoning.
(4)
Time-series trajectories are dominated by increasing patterns; decreases are localized to urbanized and marginal freezing areas.
LSTM-Autoencoder + K-means results show increasing-type trajectories across most ice-rich regions, including Tibet, Qinghai, Xinjiang, Inner Mongolia, Northeast China, Gansu, and Ningxia. Only Beijing and Yunnan show a decreasing-type trajectory with persistent ice-area decline, suggesting heightened sensitivity in highly urbanized regions and low-latitude marginal freezing zones.
(5)
Surface water availability, the thermal regime, and urbanization are the three primary controls on winter river–lake ice resources.
Random forest permutation importance identifies water-body proportion as the leading predictor of ice-area patterns, followed by NTL, NTL × Built, Elevation, and other urbanization/topographic terms; LST contributes additionally. Ecosystem variables, such as NDVI and NPP, play secondary but positive roles in supporting ice persistence and stability.

Funding

This research was supported by the General Program of the National Natural Science Foundation of China (Grant No. 42301349).

Data Availability Statement

All data supporting the findings of this study have been uploaded as part of the submitted manuscript.

Acknowledgments

The author confirms that no generative artificial intelligence (AI) tools were used in this manuscript. All research design, data analysis, figures, and scientific interpretations were conceived, conducted, and written by the author, without any AI-based assistance in generating original content or ideas.

Conflicts of Interest

The author declares no conflicts of interest.

References

  1. Murfitt, J.; Duguay, C.R. 50 Years of Lake Ice Research from Active Microwave Remote Sensing: Progress and Prospects. Remote Sens. Environ. 2021, 264, 112616. [Google Scholar] [CrossRef] [Scilit]
  2. Brooks, R.N.; Prowse, T.D.; O’Connell, I.J. Quantifying Northern Hemisphere Freshwater Ice. Geophys. Res. Lett. 2013, 40, 1128–1131. [Google Scholar] [CrossRef] [Scilit]
  3. Duguay, C.R.; Bernier, M.; Gauthier, Y.; Kouraev, A. Remote Sensing of Lake and River Ice. In Remote Sensing of the Cryosphere; John Wiley & Sons, Ltd.: Chichester, UK, 2014; pp. 273–306. ISBN 9781118368909. [Google Scholar]
  4. Leppäranta, M. Freezing of Lakes and the Evolution of Their Ice Cover; Springer International Publishing: Cham, Switzerland, 2023; ISBN 9783031256042. [Google Scholar]
  5. Sharma, S.; Blagrave, K.; Magnuson, J.J.; O’Reilly, C.M.; Oliver, S.; Batt, R.D.; Magee, M.R.; Straile, D.; Weyhenmeyer, G.A.; Winslow, L.; et al. Widespread Loss of Lake Ice around the Northern Hemisphere in a Warming World. Nat. Clim. Change 2019, 9, 227–231. [Google Scholar] [CrossRef] [Scilit]
  6. Rouse, W.R.; Binyamin, J.; Blanken, P.D.; Bussières, N.; Duguay, C.R.; Oswald, C.J.; Schertzer, W.M.; Spence, C. The Influence of Lakes on the Regional Energy and Water Balance of the Central Mackenzie River Basin. In Cold Region Atmospheric and Hydrologic Studies. The Mackenzie GEWEX Experience; Springer: Berlin/Heidelberg, Germany, 2008; pp. 309–325. ISBN 9783540739357. [Google Scholar]
  7. Prowse, T.D.; Bonsal, B.R.; Duguay, C.R.; Lacroix, M.P. River-Ice Break-Up/freeze-up: A Review of Climatic Drivers, Historical Trends and Future Predictions. Ann. Glaciol. 2007, 46, 443–451. [Google Scholar] [CrossRef] [Scilit]
  8. Prowse, T.D.; Brown, K. Hydro-Ecological Effects of Changing Arctic River and Lake Ice Covers: A Review. Hydrol. Res. 2010, 41, 454–461. [Google Scholar] [CrossRef] [Scilit]
  9. Burrell, B.C.; Beltaos, S.; Turcotte, B. Effects of Climate Change on River-Ice Processes and Ice Jams. Int. J. River Basin Manag. 2023, 21, 421–441. [Google Scholar] [CrossRef] [Scilit]
  10. Yang, Z.; Zong, J.; Zhu, Y.; Liu, X.; Tao, R.; Yu, Y. River Ice Regime Recognition Based on Deep Learning: Ice Concentration, Area, and Velocity. Water 2023, 16, 58. [Google Scholar] [CrossRef] [Scilit]
  11. Higgins, S.N.; Desjardins, C.M.; Drouin, H.; Hrenchuk, L.E.; van der Sanden, J.J. The Role of Climate and Lake Size in Regulating the Ice Phenology of Boreal Lakes. J. Geophys. Res. Biogeosci. 2021, 126. [Google Scholar] [CrossRef] [Scilit]
  12. Aslamov, I.A.; Kozlov, V.V.; Kirillin, G.B.; Mizandrontsev, I.B.; Kucher, K.M.; Makarov, M.M.; Gornov, A.Y.; Granin, N.G. Ice–water Heat Exchange during Ice Growth in Lake Baikal. J. Great Lakes Res. 2014, 40, 599–607. [Google Scholar] [CrossRef] [Scilit]
  13. Palshin, N.I.; Zdorovennova, G.E.; Efremova, T.V.; Bogdanov, S.R.; Zdorovennov, R.E.; Smirnov, S.I.; Terzhevik, A.Y. Heat Exchange at the Water–ice Interface and Ice Growth in a Small Boreal Lake. Water Resour. 2024, 51, S231–S244. [Google Scholar] [CrossRef] [Scilit]
  14. Barrette, P.D.; Lindenschmidt, K.-E. Frazil Ice Events: Assessing What to Expect in the Future. Hydrol. Res. 2023, 54, 770–781. [Google Scholar] [CrossRef] [Scilit]
  15. Kovachis, N.; Burrell, B.C.; Huokuna, M.; Beltaos, S.; Turcotte, B.; Jasek, M. Ice-Jam Flood Delineation: Challenges and Research Needs. Can. Water Resourc. J./Rev. Can. Ressour. Hydr. 2017, 42, 258–268. [Google Scholar] [CrossRef] [Scilit]
  16. Murfitt, J.C.; Brown, L.C.; Howell, S.E.L. Estimating Lake Ice Thickness in Central Ontario. PLoS ONE 2018, 13, e0208519. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  17. Li, X.; Long, D.; Huang, Q.; Zhao, F. The State and Fate of Lake Ice Thickness in the Northern Hemisphere. Sci. Bull. 2022, 67, 537–546. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  18. Stubblefield, A.G.; Creyts, T.T.; Kingslake, J.; Siegfried, M.R.; Spiegelman, M. Surface Expression and Apparent Timing of Subglacial Lake Oscillations Controlled by Viscous Ice Flow. Geophys. Res. Lett. 2021, 48, e2021GL094658. [Google Scholar] [CrossRef] [Scilit]
  19. Brown, L.C.; Duguay, C.R. Modelling Lake Ice Phenology with an Examination of Satellite-Detected Subgrid Cell Variability. Adv. Meteorol. 2012, 2012, 2719–2729. [Google Scholar] [CrossRef] [Scilit]
  20. Weber, H.; Riffler, M.; Nõges, T.; Wunderle, S. Lake Ice Phenology from AVHRR Data for European Lakes: An Automated Two-Step Extraction Method. Remote Sens. Environ. 2016, 174, 329–340. [Google Scholar] [CrossRef] [Scilit]
  21. Liu, B.; Wan, W.; Xie, H.; Li, H.; Zhu, S.; Zhang, G.; Wen, L.; Hong, Y. A Long-Term Dataset of Lake Surface Water Temperature over the Tibetan Plateau Derived from AVHRR 1981–2015. Sci. Data 2019, 6, 48. [Google Scholar] [CrossRef] [Scilit]
  22. Du, J.; Kimball, J.S.; Duguay, C.; Kim, Y.; Watts, J.D. Satellite Microwave Assessment of Northern Hemisphere Lake Ice Phenology from 2002 to 2015. Cryosphere 2017, 11, 47–63. [Google Scholar] [CrossRef] [Scilit]
  23. Zhang, H.; Yao, X.; Wei, Q.; Duan, H.; Zhang, Y. Comparison of Lake Ice Extraction Methods Based on MODIS Images. Remote Sens. 2022, 14, 4740. [Google Scholar] [CrossRef] [Scilit]
  24. Hall, D.K.; Foster, J.L.; Chang, A.T.C.; Rango, A. Freshwater Ice Thickness Observations Using Passive Microwave Sensors. IEEE Trans. Geosci. Remote Sens. 1981, GE-19, 189–193. [Google Scholar] [CrossRef]
  25. Cai, Y.; Duguay, C.R.; Ke, C.-Q. A 41-Year (1979–2019) Passive-Microwave-Derived Lake Ice Phenology Data Record of the Northern Hemisphere. Earth Syst. Sci. Data 2022, 14, 3329–3347. [Google Scholar] [CrossRef] [Scilit]
  26. Korhonen, A. A Method for Estimating Ice Phenology of Finnish Waterbodies Using Sentinel-1 IW Mode Data and in-Situ Wind Observations. Master’s Thesis, Aalto University, Espoo, Finland, 2018. [Google Scholar]
  27. Gaudreau, A. The Influence of Bubbles on the Seasonal SAR Backscatter Response of Perennially Ice-Covered Lakes, Antarctica. Master’s Thesis, University of Ottawa, Ottawa, ON, Canada, 2023. [Google Scholar]
  28. Brown, R.S.; Duguay, C.R.; Mueller, R.P.; Moulton, L.L.; Doucette, P.J.; Tagestad, J.D. Use of Synthetic Aperture Radar (SAR) to Identify and Characterize Overwintering Areas of Fish in Ice-covered Arctic Rivers: A Demonstration with Broad Whitefish and Their Habitats in the Sagavanirktok River, Alaska. Trans. Am. Fish. Soc. 2010, 139, 1711–1722. [Google Scholar] [CrossRef] [Scilit]
  29. Gauthier, Y.; Poulin, J.; Bernier, M. Ice Mapping Automated Procedure from Radar Data; Institut National de la Recherche Scientifique: Québec, QC, Canada, 2016. [Google Scholar]
  30. Dabboor, M.; Shokr, M. Sensitivity of Compact Polarimetric SAR Parameters to Modeled Lake Ice Growth. IEEE Trans. Geosci. Remote Sens. 2021, 59, 9953–9967. [Google Scholar] [CrossRef] [Scilit]
  31. Rudjord, Ø.; Solberg, R.; Spreen, G.; Gerland, S. Estimating Thin Ice Thickness around Svalbard Using MODIS Satellite Imagery. Geogr. Ann. Ser. A Phys. Geogr. 2022, 104, 127–149. [Google Scholar] [CrossRef] [Scilit]
  32. Zakharov, I.; Puestow, T.; Khan, A.A.; Briggs, R.; Barrette, P. Review of River Ice Observation and Data Analysis Technologies. Hydrology 2024, 11, 126. [Google Scholar] [CrossRef] [Scilit]
  33. Huang, L.; Timmermann, A.; Lee, S.-S.; Rodgers, K.B.; Yamaguchi, R.; Chung, E.-S. Emerging Unprecedented Lake Ice Loss in Climate Change Projections. Nat. Commun. 2022, 13, 5798. [Google Scholar] [CrossRef] [Scilit]
  34. Walsh, C.R.; Patterson, R.T. Regional Impact of Large-Scale Climate Oscillations on Ice out Variability in New Brunswick and Maine. PeerJ 2022, 10, e13741. [Google Scholar] [CrossRef] [Scilit]
  35. Nedbalová, L.; Mihál, M.; Procházková, L.; Kvíderová, J.; Lukavský, J. The First Phylogenetic and Ecophysiological Characterisation of Ankistrodesmus Antarcticus CCAP 202/25, an Antarctic Green Alga Isolated from Freshwater Ice. Czech Polar Rep. 2025, 14, 305–322. [Google Scholar] [CrossRef] [Scilit]
  36. Rouse, W.R.; Blanken, P.D.; Duguay, C.R.; Oswald, C.J. Chapter 8: Climate-Lake Interactions. In Cold Region Atmospheric and Hydrologic Studies, the Mackenzie GEWEX Experience; Ming-ko, W., Ed.; Springer: Berlin/Heidelberg, Germany, 2008; Volume 2, pp. 139–160. [Google Scholar]
  37. Pavelsky, T.M.; Smith, L.C. Spatial and Temporal Patterns in Arctic River Ice Breakup Observed with MODIS and AVHRR Time Series. Remote Sens. Environ. 2004, 93, 328–338. [Google Scholar] [CrossRef] [Scilit]
  38. Kropáček, J.; Maussion, F.; Chen, F.; Hoerz, S.; Hochschild, V. Analysis of Ice Phenology of Lakes on the Tibetan Plateau from MODIS Data. Cryosphere 2013, 7, 287–301. [Google Scholar] [CrossRef] [Scilit]
  39. Sobiech, J.; Dierking, W. Observing Lake- and River-Ice Decay with SAR: Advantages and Limitations of the Unsupervised k-Means Classification Approach. Ann. Glaciol. 2013, 54, 65–72. [Google Scholar] [CrossRef] [Scilit]
  40. Bevainis, L.; Bielinis, M.; Česnulevičius, A.; Bautrėnas, A. Lithuanian River Ice Detection and Automated Classification Using Machine-Learning Methods. Baltica 2023, 36, 1–12. [Google Scholar] [CrossRef] [Scilit]
  41. Jeffries, M.O.; Morris, K.; Liston, G.E. A Method to Determine Lake Depth and Water Availability on the North Slope of Alaska with Spaceborne Imaging Radar and Numerical Ice Growth Modelling. Arctic 1996, 49, 367–374. [Google Scholar] [CrossRef] [Scilit]
  42. Stolpmann, L.; Nitze, I.; Bussmann, I.; Jones, B.M.; Lenz, J.; Meyer, H.; Wolter, J.; Grosse, G. Thermokarst Lake Change and Lake Hydrochemistry: A Snapshot from the Arctic Coastal Plain of Alaska. EGUsphere 2024, 1–34. [Google Scholar] [CrossRef] [Scilit]
  43. Pour, H.K.; Duguay, C.R.; Martynov, A.; Brown, L.C. Simulation of Surface Temperature and Ice Cover of Large Northern Lakes with 1-D Models: A Comparison with MODIS Satellite Data and in Situ Measurements. Tellus A 2012, 64, 17614. [Google Scholar] [CrossRef] [Scilit]
  44. Tang, F.; Chen, P.; An, Z.; Xiong, M.; Chen, H.; Qiu, L. A Dual-Threshold Algorithm for Ice-Covered Lake Water Level Retrieval Using Sentinel-3 SAR Altimetry Waveforms. Sensors 2023, 23, 9724. [Google Scholar] [CrossRef] [Scilit]
  45. Stonevicius, E.; Uselis, G.; Grendaite, D. Ice Detection with Sentinel-1 SAR Backscatter Threshold in Long Sections of Temperate Climate Rivers. Remote Sens. 2022, 14, 1627. [Google Scholar] [CrossRef] [Scilit]
  46. MODIS/Terra Land Surface Temperature/Emissivity Daily L3 Global 1 km SIN Grid V061. Available online: https://www.earthdata.nasa.gov/data/catalog/lpcloud-mod11a1-061 (accessed on 10 December 2025).
  47. Climate Hazards Center—UC Santa Barbara. Available online: https://chc.ucsb.edu/data/chirtsdaily (accessed on 10 December 2025).
  48. Global Surface Water Explorer. Available online: https://global-surface-water.appspot.com/ (accessed on 10 December 2025).
  49. Stubbington, R.; Chadd, R.; Cid, N.; Csabai, Z.; Miliša, M.; Morais, M.; Munné, A.; Pařil, P.; Pešić, V.; Tziortzis, I.; et al. Biomonitoring of intermittent rivers and ephemeral streams in Europe: Current practice and priorities to enhance ecological status assessments. Sci. Total Environ. 2018, 618, 1096–1113. [Google Scholar] [CrossRef] [Scilit]
  50. Zafar, A.; Che, Y.; Ahmed, M.; Sarfraz, M.; Ahmad, A.; Alibakhshikenari, M. Enhancing Power Generation Forecasting in Smart Grids Using Hybrid Autoencoder Long Short-Term Memory Machine Learning Model. IEEE Access 2023, 11, 118521–118537. [Google Scholar] [CrossRef] [Scilit]
  51. Bekele, B.; Wu, W.; Legesse, A.; Temesgen, H.; Yirsaw, E. Socio-Environmental Impacts of Land Use/cover Change in Ethiopian Central Rift Valley Lakes Region, East Africa. Appl. Ecol. Environ. Res. 2018, 16, 6607–6632. [Google Scholar] [CrossRef] [Scilit]
  52. Lindenschmidt, K.; Baulch, H.; Cavaliere, E. River and Lake Ice Processes—Impacts of Freshwater Ice on Aquatic Ecosystems in a Changing Globe. Water 2018, 10, 1586. [Google Scholar] [CrossRef] [Scilit]
  53. Filazzola, A.; Imrit, M.A.; Fleck, A.; Woolway, R.I.; Sharma, S. Declining Lake Ice in Response to Climate Change Can Impact Spending for Local Communities. PLoS ONE 2024, 19, e0299937. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  54. Li, H.; Li, H.; Wang, J.; Hao, X. Monitoring High-Altitude River Ice Distribution at the Basin Scale in the Northeastern Tibetan Plateau from a Landsat Time-Series Spanning 1999–2018. Remote Sens. Environ. 2020, 247, 111915. [Google Scholar] [CrossRef] [Scilit]
  55. Fujisaki-Manome, A.; Anderson, E.J.; Kessler, J.A.; Chu, P.Y.; Wang, J.; Gronewold, A.D. Simulating Impacts of Precipitation on Ice Cover and Surface Water Temperature across Large Lakes. J. Geophys. Res. Ocean. 2020, 125, e2019JC015950. [Google Scholar] [CrossRef] [Scilit]
  56. Xie, C.; Zhang, X.; Zhuang, L.; Zhu, R.; Guo, J. Analysis of Surface Temperature Variation of Lakes in China Using MODIS Land Surface Temperature Data. Sci. Rep. 2022, 12, 2415. [Google Scholar] [CrossRef] [Scilit]
  57. Prowse, T.; Alfredsen, K.; Beltaos, S.; Bonsal, B.R.; Bowden, W.B.; Duguay, C.R.; Korhola, A.; McNamara, J.; Vincent, W.F.; Vuglinsky, V.; et al. Effects of Changes in Arctic Lake and River Ice. Ambio 2011, 40, 63–74. [Google Scholar] [CrossRef] [Scilit]
  58. Chen, F.; Xie, T.; Yang, Y.; Chen, S.; Chen, F.; Huang, W.; Chen, J. Discussion of the “warming and Wetting” Trend and Its Future Variation in the Drylands of Northwest China under Global Warming. Sci. China Earth Sci. 2023, 66, 1241–1257. [Google Scholar] [CrossRef] [Scilit]
  59. Li, B.; Liu, D.; Yu, E.; Wang, L. Warming-and-Wetting Trend over the China’s Drylands: Observational Evidence and Future Projection. Glob. Environ. Change 2024, 86, 102826. [Google Scholar] [CrossRef] [Scilit]
  60. Ding, Y.; Yanju, L.; Ying, X.; Ping, W.; Tong, X.; Wang, J.; Ying, S.; Zhang, Y.; Song, Y.; Wang, P. Regional Responses to Global Climate Change: Progress and Prospects for Trend, Causes, and Projection of Climatic Warming-Wetting in Northwest China. Adv. Earth Sci. 2023, 38, 551–562. [Google Scholar]
  61. Mao, D.; Wang, M.; Wang, Y.; Jiang, M.; Yuan, W.; Luo, L.; Feng, K.; Wang, D.; Xiang, H.; Ren, Y.; et al. The trajectory of wetland change in China between 1980 and 2020: Hidden losses and restoration effects. Sci. Bull. 2025, 70, 587–596. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  62. Zhu, Y.; Ye, A.; Zhang, Y. Changes of total and artificial water bodies in inland China over the past three decades. J. Hydrol. 2022, 613, 128344. [Google Scholar] [CrossRef] [Scilit]
  63. Zhang, Q.; Yang, J.; Wang, W.; Ma, P.; Lu, G.; Liu, X.; Yu, H.; Fang, F. Climatic warming and humidification in the arid region of Northwest China: Multi-scale characteristics and impacts on ecological vegetation. J. Meteorol. Res. 2021, 35, 113–127. [Google Scholar] [CrossRef] [Scilit]
  64. Zhang, Q.; Yang, J.; Duan, X.; Ma, P.; Lu, G.; Zhu, B.; Liu, X.; Yue, P.; Wang, Y.; Liu, W. The eastward expansion of the climate humidification trend in northwest China and the synergistic influences on the circulation mechanism. Clim. Dyn. 2022, 59, 2481–2497. [Google Scholar] [CrossRef] [Scilit]
  65. Qiang, Z.; Jinhu, Y.; Pengli, M.; Ping, Y.; Haipeng, Y.; Zesu, Y.; Pengling, W.; Xinyu, D.; Xiaoyun, L.; Biao, Z.; et al. The enhancement and eastward expansion of climate warming and humidification, formation mechanism and important environmental impacts in Northwest China. J. Arid Meteorol. 2023, 41, 351. [Google Scholar]
  66. Yang, X.; Pavelsky, T.M.; Allen, G.H. The Past and Future of Global River Ice. Nature 2020, 577, 69–73. [Google Scholar] [CrossRef] [Scilit]
  67. Lin, Y.; Lindenschmidt, K.-E.; Lü, H.; Zhu, Y.; Liu, M.; Chen, T.; Xu, Y. Modeling Climate Change Effects on River Ice Thickness in the Northern Hemisphere. J. Hydrol. 2025, 660, 133268. [Google Scholar] [CrossRef] [Scilit]
  68. Wang, Z.; Ding, Y.; Zhou, B.; Chen, L. Comparison of Two Severe Low-temperature Snowstorm and Ice Freezing Events in China: Role of Eurasian Mid-high Latitude Circulation Patterns. Int. J. Climatol. 2020, 40, 3436–3450. [Google Scholar] [CrossRef] [Scilit]
  69. Qi, W.; Feng, L.; Yang, H.; Liu, J. Warming Winter, Drying Spring and Shifting Hydrological Regimes in Northeast China under Climate Change. J. Hydrol. 2022, 606, 127390. [Google Scholar] [CrossRef] [Scilit]
  70. Liu, Y.; Xie, X.; Tursun, A.; Wang, Y.; Jiang, F.; Zheng, B. Surface Water Expansion due to Increasing Water Demand on the Loess Plateau. J. Hydrol. Reg. Stud. 2023, 49, 101485. [Google Scholar] [CrossRef] [Scilit]
  71. Wang, S.; Zhang, Z. Effects of Climate Change on Water Resources in China. Clim. Res. 2011, 47, 77–82. [Google Scholar] [CrossRef] [Scilit]
  72. Xie, A.; Zhu, J.; Qin, X.; Wang, S.; Xu, B.; Wang, Y. Surface Warming from Altitudinal and Latitudinal Amplification over Antarctica since the International Geophysical Year. Sci. Rep. 2023, 13, 9536. [Google Scholar] [CrossRef] [Scilit]
  73. Wu, F.; You, Q.; Cai, Z.; Sun, G.; Normatov, I.; Shrestha, S. Significant Elevation Dependent Warming over the Tibetan Plateau after Removing Longitude and Latitude Factors. Atmos. Res. 2023, 284, 106603. [Google Scholar] [CrossRef] [Scilit]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Article Metrics

Citations

Article Access Statistics

Multiple requests from the same IP address are counted as one view.