Introduction

Climate change driven by anthropogenic activities is intensifying extreme heat events, increasing their frequency, duration, and severity worldwide1,2,3. Escalating atmospheric warming and persistent circulation anomalies have increased the frequency and duration of extreme atmospheric heatwaves (AHWs)4,5. Extreme heat events across South Asia, Europe and North America underscore the growing societal vulnerability to heatwaves and the urgent need for adaptation amid escalating climate extremes6,7,8. Subsequently, growing evidence suggests that heatwaves in aquatic ecosystems such as lakes, are also intensifying with increasing frequency and magnitude9,10.

Lakes play a crucial role in freshwater ecosystems by providing habitats for diverse aquatic species, regulating local climates, and supporting ecosystem services such as drinking water supply and water for irrigation. However, they are highly vulnerable to extreme heat events such as lake heatwaves (LHWs)11,12, defined as periods where water temperatures exceed the climatological threshold for an extended period. Recent research has increasingly focused on extreme variations in lake surface water temperatures, which have been shown to have substantial ecological and biogeochemical consequences13. Among these extremes, LHWs are particularly severe, as they alter thermal stratification, increase water-column stability14, reduce vertical mixing, promote harmful algal blooms15, and ultimately threaten water quality and ecosystem health.

Beyond their individual effects, the co-occurrence of heatwaves with other environmental stressors is increasingly recognized as a key mechanism by which interacting drivers and non-linear responses intensify risks to natural systems16,17. Observations show that surface wind speed decreased by 5–15% across much of the northern mid-latitudes from 1979 to around 201018,19. In lakes, this decline may intensify stratification, reduce oxygen levels, and promote eutrophication20, modulating and potentially amplifying the ecological impacts of heatwaves, which also depend on lake morphology21,22. Nevertheless, it remains unclear whether these wind-related changes systematically exacerbate LHWs on a global scale.

Besides the above gaps, to our knowledge, there is currently no specific research comparing the differences between lake and atmospheric heatwaves. This question is crucial, since the distinct thermodynamic properties of water and air result in fundamentally different heatwave characteristics, leading to divergent ecological impacts. Firstly, AHWs typically manifest as short-duration temperature extremes, inducing acute physiological stress in terrestrial and aerial organisms. Conversely, LHWs, driven by the high specific heat capacity and low thermal conductivity of water, tend to generate long-term and cumulative effects. LHWs also exacerbate metabolic stress in aquatic species through dissolved oxygen depletion. Secondly, while atmospheric heat transfer occurs predominantly via rapid convection and radiation, aquatic heat dynamics are influenced by turbulence, mixed-layer depth, and air-water coupling processes. Consequently, these systems exhibit markedly different spatiotemporal scales and ecological response patterns. Finally, thermal tolerance thresholds and adaptation strategies vary considerably across ecosystems. Terrestrial organisms mitigate thermal damage through behavioural migration and physiological regulation23, whereas prolonged heat exposure renders many aquatic species vulnerable to community degradation, and even ecosystem collapse24.

AHWs have been extensively documented due to the abundance of air temperature data. However, studying lake heatwaves presents unique challenges due to the scarcity of continuous, long-term lake temperature records. In-situ measurements, while precise, are limited in terms of their spatial and temporal coverage25. Satellite-based thermal infrared observations have provided valuable large-scale insights, but cloud cover often disrupts data collection26,27. To address these gaps, recent advances in numerical modelling and machine learning techniques have enabled more comprehensive analyses of lake temperature trends28,29.

In this study we combine remote sensing, machine learning, and in-situ observations to construct global daily lake temperature records for 2000–2022. We test two hypotheses: (1) LHWs differ markedly from AHWs, with lakes potentially experiencing more severe heatwaves because of their high specific heat capacity; and (2) decreasing wind speed exacerbates LHWs by further stabilizing stratification and limiting heat dissipation. By clarifying the contrasts and couplings between AHWs and LHWs, including the characteristics of atmosphere–lake coupled heatwave events (CHWs), our work provides a foundation for improved forecasting and for the development of effective freshwater-ecosystem adaptation strategies.

Results

Comparison of LHWs and AHWs

During the study period, LHWs typically occurred more than five times per year, differing from AHWs which occurred, on average, four times per year (t test: p < 0.05). The average accumulated heatwave days for LHWs were 29.7 ± 0.2 days for all lakes, being significantly longer than the 18.0 ± 0.3 days calculated for AHWs (t test: p < 0.05). Compared with LHWs, AHWs experienced more pronounced temperature anomalies (t test: p < 0.05), with a global average intensity of 6.1 °C ± 0.2 °C, while the average of LHWs was 3.0 °C± 0.1 °C. Additionally, the global mean reoccurrence interval days for AHWs were 121.4 ± 2.3 days, whereas those of LHWs were 86.9 ± 1.6 days (t test: p < 0.05). These results indicate that LHWs persist longer and have shorter reoccurrence intervals. The cumulative heatwave indices (CHI) and the cumulative heat index among climatic zones (CHIP) (as defined in the methods) of LHWs were greater than those of AHWs (CHI: LHW at 14.9 ± 0.1 days·°C, AHW at 12.7 ± 0.2 days·°C; CHIP: LHW at 5.6 ± 0.1 days·°C, AHW at 5.0 ± 0.2 days·°C), indicating that lake surface waters are likely more sensitive to long-term thermal stress.

Regionally, there were marked spatial variations across continents (Fig. 1). South America and Oceania exhibited similar long-term trends in both LHW and AHW indicators. In contrast, Africa exhibited the weakest heatwaves among all regions for both AHWs and LHWs, with the smallest differences in heatwave intensity between the two. This suggests that heatwaves in Africa are characterized by milder temperature anomalies and more similar LHWs and AHW characteristics. Europe, on the other hand, had a greater frequency and longer accumulated days of AHWs than the other regions. However, LHWs in Europe were less pronounced, leading to smaller differences between LHW and AHW indicators. Interestingly, the specific indices for Europe suggest that AHWs might be more severe, as reflected in the differences between LHWs and AHWs in terms of CHI and CHIP, values of −1.8 days·°C and −0.7 days·°C (95% CI: CHI −2.1–1.3 days·°C; CHIP −0.7– −0.5 days·°C), respectively.

Fig. 1: Global distributions of annual average heatwave metrics and differences for lake heatwaves (LHWs) and atmospheric heatwaves (AHWs).
Fig. 1: Global distributions of annual average heatwave metrics and differences for lake heatwaves (LHWs) and atmospheric heatwaves (AHWs).
Full size image

a Annual averages of heatwave frequency, b accumulated heatwave days, c heatwave intensity, d reoccurrence period, e cumulative heatwave indices (CHI), and f cumulative heat index among climatic zones (CHIP). The maps show the differences between LHWs and AHWs, and the circular charts display each metric for LHWs and AHWs on each continent. Blue and light blue show annual means; error bars represent the upper and lower bounds of 95% confidence intervals (AF: Africa, NA: North America, SA: South America, AS: Asia, EU: Europe, OC: Oceania).

Long-term trends in LHW and AHW differences

Globally, approximately 71% of the lakes exhibited a rising trend in the frequency of LHWs by, on average, +1 events decade-1. In Asia, approximately 50% of the lakes showed an increase of +2 events decade-1. Additionally, the number of accumulated LHW days increased by approximately +10 days per decade in 74% of lakes. Moreover, the reoccurrence period of LHWs notably decreased, by −30 days per decade in 74% of lakes. CHI and CHIP also exhibited rapidly increasing trends (in 71 and 68% of lakes, respectively), particularly in Asia and North America, where these trends were most pronounced (Supplementary Fig. 1). In contrast, AHWs were similar but experienced less pronounced trends relative to LHWs. AHW frequency and accumulated heatwave days increased more gradually, but the decline in the reoccurrence period was larger, with a decrease of 34 days decade-1. The increase rates of CHI and CHIP were also lower than those for LHWs (Fig. 2).

Fig. 2: Mann-Kendall and Theil-Sen slope trend results for the CHI and CHIP of LHWs and AHWs.
Fig. 2: Mann-Kendall and Theil-Sen slope trend results for the CHI and CHIP of LHWs and AHWs.
Full size image

a CHI trends of LHWs and b CHI trends of AHWs, c CHIP trends for LHWs and d CHIP trends for AHWs. All trends were subjected to significance testing, with significance levels indicated in the legend at the bottom of the figure.

Atmosphere-lake coupled heatwaves

Compared with isolated heatwave events, CHWs exhibit greater intensity and longer durations, with the CHI and CHIP showing notable differences (Fig. 3a). Our analysis indicated that the proportion of CHWs increased annually, with the probability of occurrence rising significantly over time: 0.30% per year relative to LHWs, 0.23% per year relative to AHWs, and 0.17% per year relative to CHWs (p < 0.05) (Fig. 3b). In Europe and the coastal areas of East Asia, the annual frequency of CHW occurrence exceeded three events per year, corresponding to higher frequencies of both LHWs and AHWs in these areas, suggesting a substantially increased likelihood of overlap (Fig. 3c).

Fig. 3: Characteristics, occurrence probability, and spatial distribution of atmosphere-lake coupled heatwaves (CHWs).
Fig. 3: Characteristics, occurrence probability, and spatial distribution of atmosphere-lake coupled heatwaves (CHWs).
Full size image

a The metrics of CHWs (CL for lake surface water and CA for the atmosphere) are compared with those of independent heatwave events (DL and DA, the error bars represent the standard deviation). b Trend in the CHW occurrence probability; c represents the annual frequency of CHWs, and (b) and (a) are the frequencies of AHWs and LHWs, respectively. The joint probability (c) is calculated as c/(a + b), with the vertical axis showing the probability (ranging from 0 to 1). c Annual spatial distribution of CHWs across the study lakes and their continental breakdown. Continental abbreviations follow those used in Fig. 1.

As indicated by CHWs captured using a five-day search window (i.e., the interval between LHW and AHW is less than 5 days), AHWs tended to begin earlier than LHWs, while LHWs exhibited longer total durations. At the global scale, the mean time-lag response (TLR, > 0 indicates that AHW occurred first) was 0.32 ± 0.03 days, and the coupling-time difference (CTD, > 0 indicates that LHW lasted longer) was 0.89 ± 0.04 days (Fig. 4). Notably, both TLR and CTD still exhibited pronounced spatial heterogeneity: LHWs tended to occur earlier during CHWs in North America and parts of East Asia, whereas in regions such as South America, the western coast of Africa, and parts of South Asia, AHWs demonstrated longer total durations.

Fig. 4: Time-lag response (TLR) and coupling-time difference (CTD) of CHWs across global lakes.
Fig. 4: Time-lag response (TLR) and coupling-time difference (CTD) of CHWs across global lakes.
Full size image

a Annual average TLR, positive values indicating earlier occurrence of AHWs than of LHWs, and negative values the opposite. The bar chart shows annual average TLR and standard errors for each continent. b Annual average CTD, positive values indicating longer duration of LHWs than of AHWs and negative values the opposite. The bar chart illustrates annual average CTD and standard errors for each continent.

Furthermore, we analysed the potential impacts of lake morphology on CHWs (e.g., depth, area, etc.), but no statistically significant results were found in our database. Lakes with depths greater than the median depth exhibited a larger lag than shallower lakes (TLR = 0.39 d vs. 0.29 d, t test: p > 0.05), while their durations were marginally similar (CTD ≈ 0.84 vs. 0.83 d). Similarly, lakes with surface areas larger than median showed earlier TLRs compared to smaller lakes (0.39 and 0.29 d, respectively, p > 0.05). High-altitude lakes ( ≥ 2000 m) recorded the longest mean TLR (0.83 d), but their CTD (−0.10 d) indicates that AHWs and LHWs durations were nearly. Low- and mid-altitude lakes displayed intermediate lags but the greatest CTD (0.88 d).

Factors affecting the differences between LHWs and AHWs

Because extreme heatwave events can arise from increases in mean temperature or changes in temperature variance, we analysed both mean and variance of lake and air temperatures during the study period to identify and compare the mechanisms driving these events. From 2000 to 2022, both near-surface air temperature above the study lakes and lake surface water temperature showed pronounced warming. Of the 265 lakes, 211 (79.6%) exhibited upward trends in annual-mean air temperature and 202 (76.2%) in annual-mean lake-surface temperature, with the strongest warming in East Asia, Europe, and the west coast of North America (Supplementary Fig. 2a, c). Interannual variance changed far less: only 39 lakes (14.7%) displayed significant trends in air-temperature variance and only 30 lakes (11.3%) in water temperature variance (Mann–Kendall, p ≤ 0.05; Supplementary Fig. 2b, d). In 76.2% of lakes, the mean air temperature change exceeded that of the variance, while the opposite appeared for the remaining 23.4%. In 83.8% of the lakes, mean water temperature increased faster than the variance.

Attribution analysis showed that CHI and CHIP trends for AHW coincided with rising annual-mean air temperature in 204 and 205 lakes. Moreover, CHI in 58 lakes and CHIP in 55 lakes were attributed to changes in temperature variance, while for the remaining lakes (3 for CHI and 5 for CHIP), trends of CHI and CHIP could not be explained by temperature variance alone, indicating an impact of other factors (Supplementary Fig. 2e, g). For LHWs, CHI and CHIP trends aligned with temperature in 223 and 225 lakes, respectively, while only 41 (CHI) and 37 (CHIP) lakes could be explained by temperature variance (Supplementary Fig. 2g). Collectively, these results confirm that, for the majority of lakes, rising mean temperature, rather than changes in variance, was the primary driver of the intensifying lake and atmospheric heatwaves as well as the growing disparity between LHWs and AHWs.

To investigate the contributions of key factors to the differences between LHWs and AHWs, a Random-Forest analysis was applied to each lake. At the annual scale, an attribution of environmental drivers (wind speed, precipitation, sensible-heat flux, latent-heat flux, long-wave and short-wave radiation, and atmospheric pressure) to the differences in CHI and CHIP showed that, besides temperature, wind speed and precipitation dominated in 53 and 48 lakes, respectively. Wind speed exhibited the highest mean weighted importance (ΔCHI = 29.2 ± 11.6%; ΔCHIP = 28.7 ± 10.4%, ranking first in 48 lakes), followed by precipitation, sensible-heat flux, and long-wave radiation (Fig. 5a, b). Wind speed was also an important factor – during LHWs, 205 out of 265 lakes presented a mean wind speed reduction of -0.18 m s-1, and during AHWs, 126 out of 265 lakes revealed a mean wind speed increase of 0.02 m s-1. Regarding CHWs, 182 out of 265 lakes exhibited a mean wind speed reduction of -0.15 m s-1 (Fig. 5c, d). The pronounced influence of wind speed underscores its pivotal role in modulating heatwave severity across lake–atmosphere systems. During CHW events, long-wave radiation was the leading contributor (ΔCHI = 21.0 ± 12.8%, 95 lakes ranked first; ΔCHIP = 18.8 ± 12.0%, 78 lakes ranked first), followed by latent-heat flux and wind speed (ΔCHI = 15.5 ± 10.2% and 12.2 ± 7.6%; ΔCHIP = 13.4 ± 8.2% and 12.7 ± 7.3%). These factors collectively amplified the lake–air heatwave contrast during coupled events.

Fig. 5: Contributions of meteorological factors to heatwave differences and wind speed differences before and during heatwaves.
Fig. 5: Contributions of meteorological factors to heatwave differences and wind speed differences before and during heatwaves.
Full size image

a, b Number of lakes where each meteorological factor (LR: long-wave radiation, SR: short-wave radiation, LH: latent-heat flux, SH: sensible-heat flux, RA: rainfall, PR: pressure, WS: wind speed) is the primary contributor to CHI and CHIP heatwave differences along with mean and standard deviation of the maximum contributions. c, d Annual average differences in wind speed (units: m s-1) before and during AHWs and LHWs, respectively.

Structural Equation Modelling was further used to analyse the influencing factors of spatial distribution patterns on the differences between LHWs and AHWs. The results revealed that geographic factors contributed most to the differences, with a path coefficient of -0.32, underscoring the dominant role of lake location—especially latitude—in determining the disparity between lake and atmospheric heatwaves (Supplementary Fig. 3). The path coefficient for lake characteristics was 0.07, with depth (0.53) and Normalized Difference Vegetation Index (NDVI, 0.79) contributing most strongly to this latent variable, suggesting that deeper lakes possess greater buffering capacity against heatwave effects. Meteorological factors and lake trophic state had relatively minor contributions to the heatwave indicator differences, with path coefficients of -0.04 for both. Generally, lower wind speeds, increased rainfall, and reduced chlorophyll content in lakes were associated with greater disparities in composite heatwave indicators.

Projected differences in heatwaves under future climate change

LHWs and AHWs were projected to intensify throughout the twenty-first century, as indicated by the results based on a fixed-baseline from 2000 to 2022 (Fig. 6). Although the specific values varied among scenarios, the future trends were consistent. In scenario Shared Socio-economic Pathway (SSP) 585, the frequency of LHWs rose by +1.4 events decade-1 (SSP370 + 1.1; SSP126 + 0.7; p < 0.01), accumulated heatwave days by +26.9 days decade-1 (SSP370 + 21.1; SSP126 + 7.7; p < 0.01), and mean intensity by +0.14 °C decade-1 (SSP370 + 0.10; SSP126 + 0.04; p < 0.01). CHI and CHIP increased by +5.1 days·°C decade-1 and +2.3 days·°C decade-1, respectively (SSP370: +3.8/ + 1.6; SSP126: +1.9/ + 0.8; p < 0.01), while the reoccurrence period shortened by -7.8 days decade-1 (SSP370 -7.5; SSP126 -5.2; p < 0.01).

Fig. 6: Annual trends in heatwave metrics for AHWs and LHWs under future emission scenarios (Shared Socio-economic Pathway (SSP): SSP126, SSP370, SSP585), calculated with a fixed 2000–2022 baseline.
Fig. 6: Annual trends in heatwave metrics for AHWs and LHWs under future emission scenarios (Shared Socio-economic Pathway (SSP): SSP126, SSP370, SSP585), calculated with a fixed 2000–2022 baseline.
Full size image

The left panel depicts the AHW trends, and the right panel shows the corresponding LHW trends. Hist conditions (2000–2022) are shown as a black line; SSP126, SSP370, and SSP585 for 2023–2100 are shown as green, blue, and red lines, respectively. Shading denotes the 25–75th percentile range. From top to bottom, the metrics are a, b accumulated heatwave days, c, d mean intensity, e, f reoccurrence period, g, h CHI, i,j CHIP, and k, l heatwave frequency.

AHWs accelerated even more rapidly than LHWs, as indicated by frequency and intensity. In the SSP585 scenario, their frequency increased by +1.2 events decade-1 (SSP370 + 1.0; SSP126 + 0.7; p < 0.01), and accumulated heatwave days rose by +34.2 days decade-1 (SSP370 + 28.5; p < 0.01) and intensity by +0.30 °C decade-1 (SSP370 + 0.20; p < 0.01). CHI and CHIP increased by +6.2 and +3.6 days·°C decade-1, respectively, (SSP370: +4.9/ + 2.7; SSP126: +2.8/ + 1.2; p < 0.01), whereas the reoccurrence period decreased by -11.9 days decade-1 (SSP370 -9.7; SSP126 -7.0; p < 0.01).

Consequently, the difference between LHWs and AHWs narrowed, especially in scenarios SSP370 and SSP585. By following a fixed-baseline approach, the accumulated heatwave days and reoccurrence days for LHWs and AHWs nearly converged by the 2040 s, while the cumulative severity (CHI, CHIP) of AHWs exceeded those for LHWs in the early 2030 s.

Because the choice of baseline may affect heatwave evaluation, we also tested a sliding baseline method, as emphasized in the heatwave studies30,31. Across the three scenarios, heatwaves identified with the sliding baseline still intensified for both LHWs and AHWs (Supplementary Fig. 4). Nevertheless, by the end of the twenty-first century, the integrated intensity and cumulative exposure remained higher for LHWs than for AHWs.

Discussion

Spatiotemporal differences between LHWs and AHWs

Our study suggests that heatwaves in lakes have been more severe than those in the atmosphere during the period 2000–2022. Specifically, frequency, duration, and accumulated heat were greater in LHWs than in AHWs. Moreover, the long-term trends of LHWs increased more rapidly than those of AHWs from 2000 to 2022. Among the study lakes, the trend of atmosphere-lake coupled heatwave events was consistent with previous studies32,33, with peak frequencies observed globally around ~2010 and 2015. Furthermore, our results indicate, based on CHWs, that LHWs start earlier in North America and parts of East Asia, last longer, and accumulate more heat. This suggests that aquatic organisms may also face greater risks than those exposed to atmospheric heatwaves alone. To the best of our knowledge, this is the first study to compare the differences between heatwaves in lakes and the overlying atmosphere. Knowing these differences is crucial for understanding the risks that heatwaves pose to different ecosystems and for developing corresponding strategies. For instance, terrestrial species can alleviate heat stress via transpiration, or behavioural avoidance (e.g., seeking shade or by nocturnal activity), whereas aquatic biota mainly rely on passive heat conduction and limited horizontal movement within a thermally stratified water column. In shallow lakes, the biological exposure risk might be higher due to relatively uniform thermodynamic characteristics and modest night-time cooling, offering little relief34, especially for temperature-sensitive species35.

The fundamental reason for these differences between LHW and AHW characteristics and their long-term variations may be partly driven by differences in the heat capacity. Notably, water has a specific heat capacity of ≈ 4.2 kJ kg⁻¹ K⁻¹, almost three times higher than that of air ( ≈ 1.4 kJ kg⁻¹ K⁻¹)36, so lake temperatures respond more slowly to external forcing, storing heat and imparting a long-term thermal memory that sustains and amplifies LHWs. This lagged response is further enhanced under calm, warming conditions that suppress turbulent mixing and strengthen stratification, allowing heat to accumulate in the upper layer37. Spatially, climate background variability is an important factor explaining differences between LHWs and AHWs in lakes worldwide. Differences between LHWs and AHWs show the smallest differences in Europe, followed by Asia. In terms of heatwave intensity, these differences are minimal in African lakes and most pronounced in North America. Besides the reasons mentioned above, although not statistically significant in our study, lake depth and area largely determine the evolution of a lake’s thermal structure38, thereby potentially mediating the frequency and intensity of lake heatwaves and ultimately influencing the magnitude of LHW–AHW differences.

Drivers of long-term variability in lake–air heatwave differences

Because the lake surface is in continuous thermal contact with the overlying air, LHWs cannot be regarded as independent phenomena. In most cases they are initiated by the same persistent synoptic anomalies that trigger AHWs, such as high-pressure, clear-sky, low-wind conditions. In both the air and the lakes, increasing average temperature is the primary driver of changes in both AHWs and LHWs39,40.

Further analysis of the role of wind speed in heatwave processes in the study lakes suggested that during heatwave events, wind speeds were lower than in the weeks before and after the heatwaves, especially regarding LHWs and CHWs. Lower wind speeds typically promote the occurrence of LHWs, thereby increasing the probability of CHWs. Hence, changes in wind speed also played an important role in modulating LHWs, especially in North America and Eurasia (Supplementary Fig. 5), likely related to the more pronounced decline in wind speed observed here. Rainfall may also affect heatwaves, but its contribution is relatively small. The observed behaviour in lakes closely parallels the wind signature reported for coastal marine heatwaves, where early-event quiescent winds suppress latent-heat loss, while subsequent wind rebounds enhance vertical mixing and help terminate the anomaly41,42.

As wind speeds decline, near-surface air circulation slows, reducing the efficiency of heat exchange between the surface and the atmosphere. This diminished dissipation leads to heat accumulation near the surface, increasing air temperatures. For lakes, reduced wind speeds decrease vertical mixing, leading to enhanced thermal stratification, warmer lake surface waters and, in turn, more intense and prolonged LHWs24,43. This effect is particularly pronounced in shallow lakes, where diminished mixing increases surface temperatures and exacerbates stratification. Conversely, deeper lakes, owing to their greater thermal inertia, are less affected by declines in wind speed44.

In limnological studies, the significance of wind as a driving factor has historically been underestimated when compared to temperature. Initial understanding primarily centred on wind-wave disturbance. Subsequently, its role in cyanobacterial bloom migration was recognized45, followed by its impacts on nutrient dynamics20, and underwater light fields46. Our research further demonstrates that wind speed was a key factor, exerting a substantial influence on lake thermodynamic processes and lake temperature extremes. This finding calls for close attention to wind conditions in future lake heatwave management, particularly regarding predictive modelling of future scenarios.

Future projections of heatwave characteristics

Our results based on the fixed climatological baseline showed that, in both SSP370 and SSP585, the frequency of LHWs and AHWs, the accumulated heatwave days, CHI and CHIP increased markedly, whereas the reoccurrence period of the two heatwave types shortened synchronously. Whether in terrestrial or aquatic ecosystems, the risk of heatwaves in the future will further increase. However, as the rate of air temperature increase will be higher than that of water temperature, the differences between LHW–AHW decrease progressively. Previous studies have highlighted that, owing to the greater thermal inertia of water bodies, the lake surface response to global warming might be slow.

Organisms have the capacity to adapt to climate change, as demonstrated in several studies47,48,49. Extreme events can also drive (micro) evolutionary changes over geological and contemporary timescales50, affecting both microorganisms51 and larger organisms48,50. The effects of future temperature variations on performance or tolerance may depend on previous thermal history52,53. This suggests that the impact of heatwaves on organisms may be partially alleviated through adaptation or evolution48,50,54,55. Therefore, the impacts of heatwaves on organisms differ, even with similar heatwave patterns, as communities may respond rapidly to unexpectedly impactful events56.

Considering this, we have reorganized the future heatwave trends based on a sliding baseline. The results show that the risks of heatwaves increase slightly before 2030 and then stabilise until 2090, after which there will be a more pronounced increase until 2100 (Supplementary Fig. 4). However, the risks faced by aquatic ecosystems are still slightly higher than those faced by terrestrial ecosystems as indicated by the CHIP and CHI. The divergence between the two baselines arises primarily because atmospheric warming in many regions has been observed to outpace lake-surface warming57,58. Consequently, under a fixed-baseline, the increase in LHWs intensity will be smaller than that of AHWs. In contrast, the sliding baseline dynamically raises the threshold in tandem with warming lake-water and air temperatures. Simultaneously, this approach preserves intrinsic lake-specific physical attributes and regional heterogeneity. Therefore, LHWs must remain a priority in future assessments of ecological impacts and water-resource risks.

Methods

Study sites

To ensure data reliability and accuracy, we selected global lakes with a surface area exceeding 100 km² that have at least one weather station included in the World Meteorological Organization dataset (see Data Availability) within a 50 km radius. A total of 460 lakes met these criteria. For each lake we extracted the 0.1° ERA5-Land grid cell(s) whose centroids intersect the lake polygon and retrieved daily sensible-heat flux, latent-heat flux, long-wave and short-wave radiation, wind speed and 2 m air temperature data for 2000–2022 (see Data Availability). To verify ERA5-Land 2 m air temperature data56, we collocated observations from World Meteorological Organization stations located ≤ 50 km from the shoreline. Only stations with ≥ 80% complete 2 m-temperature records for 2000–2022 were retained. Daily-scale Pearson correlations (r) between station 2 m-temperature and the overlapping grid 2 m air temperature data were computed. Lakes with r ≥ 0.80 were deemed to have reliable reanalysis forcing. Ultimately, 265 lakes exhibited relatively complete time series data, with minimal discrepancies compared to the ERA5 reanalysis dataset. These lakes exhibit substantial variability in their physical characteristics, with surface areas ranging from 100.1 to 65,441.8 km² and average depths from 1.0 m to 378.6 m, spanning latitudes between 54.6°S and 71.0°N (Supplementary Fig. 6). Detailed information on lake morphology, surface area, and geographical coordinates is available in the GLAKES dataset59. Additionally, we extracted annual mean chlorophyll-a concentrations (mg m-3) from the MODIS Aqua L3SMI product using the Google Earth Engine platform. We also derived the annual mean normalized NDVI for the non-lake areas surrounding the lakes from the MOD13A1 v6 product. Lake transparency data were obtained from the 2019 global lake transparency distribution dataset (covering lakes larger than 1 hectare), provided by the National Earth System Science Data Center of China (see Data Availability).

Daily water temperature

Daily surface water temperature data for the 265 selected lakes were extracted from the MOD11A1 v6 product, which uses MODIS aboard the Terra satellite. MODIS provides global daily land surface temperature observations, and although long-term Terra operation is affected by orbit drift and a gradual increase in the solar β-angle, factors that can shift local sampling time and marginally reduce on-board calibration accuracy60, the product has consistently demonstrated sufficient reliability for extracting lake-surface temperatures in large-scale studies61,62.

To remove the impact of intra-lake heterogeneity in lake response to climate change63, surface observations for each lake were averaged, resulting in a time series of daily mean surface temperatures for each lake. Data extraction and processing were primarily conducted using the Google Earth Engine platform, where time series data for lake-averaged surface temperatures were generated. The process started with spectral band calculations to distinguish water bodies from surrounding areas, creating water masks to identify the extent of each lake. To reduce the potential impact of shoreline fluctuations and regional factors, pixels within 0.5–3 kilometers of the lake boundaries were excluded, depending on each lake’s surface area. Quality control was applied to the daily lake surface water temperature data, and only pixels with an LST error of less than 1 K were retained, specifically those with quality control values of 0, 1, 5, 17, and 2164. For lakes with ice cover periods, potential anomalies in remote sensing data (such as sensors detecting cold air temperatures above the ice surface or being influenced by ice surface characteristics) were also flagged. Data corresponding to these anomalies were excluded from the model training, ensuring that only reliable data were used in the analysis. This approach helps improve the model’s accuracy and avoids interference from ice period data during the training process65.

Missing temperature data due to cloud cover or quality issues was estimated by machine learning models. In the modelling process, the physical characteristics of the lakes (such as depth, latitude, longitude, and surface area), Julian day information, and meteorological data (including air temperature, pressure, wind speed, precipitation, cloud cover, and total radiation) were integrated. These data were sourced from the global lake dataset GLAKES, meteorological observations from the World Meteorological Organization, and the ERA5 reanalysis dataset. To select the most accurate model for predicting water temperature on missing days, various models were compared, including tree-based ensembles (LightGBM, Random Forest, Extra Trees) and a genetic algorithm-optimised backpropagation neural network. Multicollinearity analysis was conducted to assess potential multicollinearity issues among the variables. The data were split into training, validation, and test sets at a ratio of 8:1:1. Hyperparameter tuning, cross-validation, and learning rate adjustment strategies were employed to enhance model convergence. To assess the performance of the various machine learning models employed in the study, the root-mean-square error (RMSE) and coefficient of determination (R²) were used as the primary evaluation metrics66,67. The results for each model are shown in Supplementary Table 1. After the performance of the different models was compared, the genetic algorithm-optimised backpropagation neural network model emerged as the best performer, achieving an RMSE of 1.601 °C and an R² of 0.96, demonstrating high accuracy and predictive power in estimating lake surface temperatures.

To project future trajectories of lake and atmospheric heatwaves, we selected the model with bias-adjusted outputs from five CMIP6 Earth-system models, namely GFDL-ESM4, IPSL-CM6A-LR, MPI-ESM1-2-HR, MRI-ESM2-0, and UKESM1-0-LL, each represented by the r1i1p1f1 ensemble member68,69. This five-ESM subset mirrors the configuration adopted in the IPCC AR6 WG II Technical Summary and spans an equilibrium-climate-sensitivity range of approximately 2.4 °C–5 °C. Daily fields of near-surface air temperature, 10 m wind speed, surface short-wave and long-wave radiation were retrieved from the ISIMIP3 bias-adjusted archive at 0.5° × 0.5° resolution and bilinearly downscaled to the 0.1° ERA5-Land grid cells centred on each lake, ensuring consistency with the historical analysis. Heatwave metrics were subsequently computed under both a fixed climatological baseline and a sliding baseline for three Shared Socio-economic Pathways: SSP126 (low forcing), SSP370 (intermediate), and SSP585 (high forcing).

Definition and indicators of heatwaves

This study defined heatwave events on the basis of the widely adopted criteria for global surface water and the atmosphere70. Though they may deviate slightly from the stricter definitions typically applied to lake heatwaves, both LHWs and AHWs were defined as events when temperatures exceeded the 90th percentile of the climatological baseline for at least three consecutive days (Supplementary Fig. 7) to ensure a like-for-like comparison between the two media. When focusing solely on LHWs, the field convention is to require a 5-day duration at the 90th-percentile threshold of the seasonally varying climatology of daily mean lake surface water temperature, and to use the World Meteorological Organization-recommended 30-year climate normal as the historical baseline; using a shorter period requires additional robustness tests to demonstrate consistency with the 30-year baseline. Including sensitivity to threshold percentiles, temperature baselines, and baseline-period length (Supplementary Figs. 8–10). The threshold was calculated for each calendar day using observations from all available years. To reduce noise associated with shorter data records and enhance statistical robustness, we applied a 31-day moving average to smooth the data9,71. Lake surface water and atmospheric heatwave events were detected via the Python package (marineHeatWaves), which processes daily simulated temperature data (https://www.github.com/ecjoliver/marineHeatWaves).

The selected indicators for assessing heatwave events included the frequency of occurrence, mean intensity, and accumulated days of heatwaves. Additionally, we introduced other indicators, such as the reoccurrence period, which represents the interval (number of days) between the occurrence of two heatwave events. This metric is useful for exploring relationships between different types of heatwaves37. To provide a more comprehensive comparison between LHW and AHW severity, we developed two cumulative heatwave indices, the cumulative heat index (CHI) and the cumulative heat index among climatic zones (CHIP), which is a scaled heatwave index accounting for temperature variation along latitude. These two indices were formulated on the basis of two considerations: (1) Global warming is expected to have substantial ecological impacts, partly due to the dependence of metabolism on temperature. We aimed to incorporate the relationship between temperature and metabolism into heatwave accumulation, described by a simplified Boltzmann‒Arrhenius equation, which applies a logarithmic relationship between temperature and metabolic rates72. Not only the cumulative intensity of heatwaves but also their occurrence intervals can have a substantial impact on ecosystems. Therefore, we also incorporated the time interval of heatwaves into the index73,74. (2) The intensity of lake heatwaves, similar to marine heatwaves, is influenced by the range of annual surface water temperature variability75. Regions with greater annual surface temperature variation exhibit stronger heatwaves, whereas areas with lower variability (e.g., tropical lakes) experience less intense heatwaves. The spatial variation in heatwave intensity across different regions results in diverse impacts. These indices provide a two-dimensional perspective on heatwave severity by considering both absolute threshold exceedances and daily temperature fluctuations.

$${{{{\rm{CHI}}}}}_{i}={{{\mathrm{ln}}}}\left({{{\mbox{mean}}}}_{i}\times {{\mbox{dur}}}_{i}+\frac{1}{{\,{{\mbox{Reoc}}}}_{i-1}}\times {{{{\rm{CHI}}}}}_{i-1}+1\right)$$
(1)
$${{{{\rm{CHIP}}}}}_{i}={{{\mathrm{ln}}}}\left(\frac{{{\mbox{mean}}}_{i}\times {{\mbox{dur}}}_{i}}{| {\,{\mbox{tem}}}_{i}| }+\frac{1}{\,{{\mbox{Reoc}}}_{i-1}}\times {{{{\rm{CHIP}}}}}_{i-1}+1\right)$$
(2)

where i represents the i-th calculated heatwave event, and \({{\mbox{dur}}}_{i}\) denotes the duration of the i-th heatwave, \({{\mbox{mean}}}_{i}\) denotes the average heatwave intensity of the i-th heatwave, and \({{\mbox{Reoc}}}_{i}\) represents the time interval of the i-th heatwave. Specifically, CHI and CHIP were set to 0 when i = 1.

To further explore the relationship and coupling between LHWs and AHWs, atmosphere-lake coupled heatwave events (CHWs) were analysed. For each LHW event, a search window was established to identify spatiotemporally overlapping AHW events. The length of the search window extends from n days prior to the onset of each LHW event to its conclusion, where n is a user-defined parameter (Supplementary Fig. 7). In this study, n was set to five days. Additionally, we quantitatively analysed the coupling-time difference (CTD) and the time-lag response (TLR) of both heatwaves before and after their occurrence. Specifically, CTD is defined as the difference between the total duration of LHW events and the total duration of AHW events within a CHW event, whereas TLR is defined as the difference in days between the onset of the LHW event and the onset of the first AHW event captured within the same CHW event.

Statistical analyses

To assess the statistical significance of trends in various heatwave indicators over time, the Mann-Kendall trend test was applied. Widely used in climatology, this test determines whether a time series exhibits a monotonic upwards or downwards trend and whether that trend is statistically significant76. To quantify the long-term trends in heatwave events, the Theil-Sen estimator was used to calculate the rates of both temperature and wind speed. This nonparametric method is particularly useful in environmental studies as it provides a slope estimate that is resistant to outliers and suitable for skewed distributions77,78.

To further investigate the contributions of various factors to heatwave variability, we employed structural equation modelling via the partial least squares path modelling approach. This method allows us to quantify the relationships between observed (manifest) variables and latent constructs, providing insights into the weight of each observed variable in contributing to the latent variables that drive heatwave dynamics79,80. The variables were grouped into four latent variables: geographic factors (latitude and longitude), lake characteristics (surface area, depth, and surrounding vegetation, as indicated by the NDVI), meteorological factors (wind speed and precipitation; temperature was not included here because it was highly related to geographic factors), and trophic levels (water transparency and chlorophyll-a concentration).

All statistical analyses were conducted in Python. NumPy and SciPy were used for descriptive statistics and hypothesis testing, Statsmodels was used for linear regression and attribution analyses, and scikit-learn was used for Random-Forest importance evaluation. Specifically, we calculated global and continental means, standard errors, and 95% confidence intervals (derived via bootstrap resampling) for heatwaves. Random-Forest was also applied to quantify the relative contributions of multiple drivers to the LHW–AHW differences for each lake. In addition, a regression-based fingerprint attribution approach was used to separate the effects of changes in air-temperature mean and variance on the interannual trends of CHI and CHIP, identifying the primary drivers and their relative shares according to significance and driver trends81. Statistical significance was assessed at a two-tailed threshold of p ≤ 0.05.