2.1 Materials

This study integrates observational, reanalysis, and model simulation data sets. For observational data, we compute monthly means from the daily Gatún Lake water level measurements made by the ACP for the period 1965–2023. For precipitation data, we use a subset of the precipitation gage network data collected by the ACP that are situated within the Gatún Lake watershed (Table S1 in Supporting Information S1)). Precipitation data are available for the entire observational period of lake level measurements (1965–2023). Evaporation measurements (ETgages) are also available from one site (Barro Colorado Island), an island at the northern end of Gatún Lake, continuously from 1994 to 2019.

We use six variables from the ERA5-Land reanalysis (Muñoz Sabater et al., 2021) to derive monthly reservoir evaporation: 2 m air temperature (t2m), 10 m u-wind speed (u10), 10 m v-wind speed (v10), surface long-wave downward radiation (strd), surface short-wave downward radiation (ssrd), and 2 m dewpoint temperature (d2m). We extracted 9 grid cells from an area covering the Gatún Lake watershed (8.75°–9.5°N and 80.25°–79.5°W). We use these variables to compute reservoir evaporation using the method of Zhao and Gao (2019), who modified the Penman equation for open water evaporation to account for heat storage in a reservoir and the effect of fetch distance (see Supporting Information S1)). Our estimates of monthly reservoir evaporation are similar in magnitude to observed evaporation from BCI, with slight overestimation in winter months and underestimation in spring months (Figure S1 in Supporting Information S1)).

We extracted six variables from 27 models of the NEX-GDDP-CMIP6 data set (Thrasher et al., 2022): precipitation rate (pr), near-surface air temperature (tas), 10 m wind speed (sfcWind), longwave downward radiation (rlds), shortwave downward radiation (rsds), and relative humidity (hurs). Daily data were extracted for all four scenarios (SSP1-2.6, SSP2-4.5, SSP3-7.0, and SSP5-8.5) for the period 2015–2099 and for the historical period (1950–2014) from the same area of interest as the ERA5 data set, and used to compute monthly sums (pr) and means (all other variables). Of the 35 models available in the NEX-GDDP-CMIP6 data set, 8 were excluded because they did not include all required variables and/or scenarios. The NEX-GDDP-CMIP6 data set uses statistical downscaling (0.25° resolution) and bias correction to CMIP6 models while preserving the relative spatial patterns of climate variables across the domain (Thrasher et al., 2022).

Prior work using downscaled CMIP5 model simulations in Panama (Nakaegawa & Mizuta, 2023) demonstrates that a 20 km resolution (similar to the resolution of NEX-GDDP-CMIP6) captures spatial patterns of precipitation over the Gatún Lake watershed, which are controlled primarily by physiography (i.e., topography and distance to ocean). Precipitation and reservoir evaporation estimates from the NEX-GDDP-CMIP6 data set capture the seasonality and magnitudes of observations well (Figure S1 in Supporting Information S1)), with low RMSEs (precipitation RMSE = 39 mm, or 1.7% of total annual precipitation over the period 1965–2014; evaporation RMSE = 29 mm, or 3.0% of total annual evaporation over the period 1994–2014). The inter-model spread of precipitation (σ = 10 mm; 4.5% of annual mean) and evaporation (σ = 2 mm; 2.3% of annual mean) during the observation period is small due to bias correction (Thrasher et al., 2022).

2.2 Lake Level Modeling

To model Gatún Lake water levels we use a Generalized Additive Model (GAM) implemented in RStudio 2024.09.1 using the gam() function in the mgcv package, with monthly lake level as the response variable and monthly precipitation, monthly evaporation, the sum of the prior 12 months of precipitation, and the sum of the prior 12 months of evaporation as predictor variables. The GAM was fitted using observed lake levels, observed precipitation, and reservoir evaporation estimated from ERA5 inputs using the Restricted Maximum Likelihood method with smoothing (k = 20). To emphasize extreme lake levels, we use 3x weights on the upper (0.9) and lower (0.1) quantiles.

We opted for a relatively simple and flexible statistical model that could be applied to downscaled CMIP6 projections. Our model does not explicitly include reservoir management decisions because these operational decisions are not publicly documented for the calibration period and are unknown for the future, although the non-linear dynamics in the GAM implicitly include average operational decisions over the calibration period. To assess generalization, we performed a leave-one-year-out cross-validation focused on the lowest 10% of low extreme years; for each iteration, the GAM was trained on all years except the target year, which was then used as a test set (Table S4 in Supporting Information S1)). The GAM was then applied to the NEX-GDDP-CMIP6 data set, predicting lake levels at monthly steps for each model and each scenario. For plotting multi-model means and distributions, we use bootstrapping using the boot() function in RStudio with 10,000 iterations. Finally, we perform EVAs of annual lake level minima using GEV distributions with maximum likelihood fitting using the gev.fit() function in the ismev package in RStudio.

3.1 Climatic Controls on Low Gatún Lake Water Levels

The lowest Gatún Lake water levels within the observational record tend to be preceded by El Niño conditions, which are associated with anomalously low precipitation and enhanced evaporation beginning 1 year prior to the lowest water levels (Figure 2). Limited latitudinal migration of the ITCZ during El Niño conditions reduces wet season precipitation (Hatvani et al., 2025) and increases evaporation via warmer surface temperatures and stronger northeasterly winds that enhance the vapor pressure gradient over Gatún Lake (Wang & Georgakakos, 2007). Of the 14 low water level extremes analyzed here, 79% (n = 11) are associated with El Niño conditions (defined as Niño3.4 index >0.5) in the preceding year, and 84% (n = 12) are associated with negative precipitation and positive evaporation anomalies in the preceding year (Table S3 in Supporting Information S1)). The most extreme low water levels occur when negative water balance anomalies (i.e., precipitation minus evaporation) persist over multiple wet seasons. The response of ENSO to greenhouse forcing remains uncertain (Alizadeh, 2024; Brotons et al., 2024; Simpson et al., 2025), but these findings imply that precipitation and evaporation of the preceding 12 months exert a strong influence on extreme low water levels of Gatún Lake.

Given the role of precipitation and evaporation in regulating Gatún Lake water levels, a GAM based only on these climatic predictors performs well (adjusted R² = 0.57, D² = 58.7%, RMSE = 0.41; Figure S2 in Supporting Information S1)), effectively capturing the observed seasonal cycle of lake level and extreme low water levels (Figure 3). We note that GAM predictions of extremes are conservative (i.e., predicted extreme lows tend to be higher than observations), which is evident when performing leave-one-year-out cross-validation for extreme low water level years (Table S4 in Supporting Information S1); RMSE = 0.46 m) including 2023—although observations still generally fall within 95% prediction intervals (Figure S3 in Supporting Information S1)). The conservative behavior of GAM predictions is also evident when performing an extreme value analysis of annual minima using a GEV distribution (Figure 3b). For example, the estimated return period for the observed annual minimum lake level in 2023 (24.2 m) is 22 years (11–70 years; 5%–95% confidence intervals), but the GAM-predicted value (24.6 m) has a return period of 28 years (12–79 years). For context, the lowest observed mean monthly lake level occurred in May 2016 (23.9 m) and has a return period of 40 years (18–132 years). These results demonstrate that a relatively simple statistical model based on precipitation and evaporation patterns explains the majority of variance in Gatún Lake water levels, and that this model provides robust and conservative estimates of extreme low water levels relative to observations. Validation of this GAM allows us to evaluate how lake levels respond under different emissions scenarios using simulated precipitation and reservoir evaporation derived from the NEX-GDDP-CMIP6.

3.2 Gatún Lake Water Level Projections

Projections of Gatún Lake water levels based on a multi-model analysis of downscaled and bias-corrected simulations show that annual lake level minima decrease over the 21st century under all SSP scenarios, but that lake levels drop more under higher emissions scenarios (Figure 4). Comparing the end of the 20th (1980–1999) and 21st (2080–2099) centuries, the multi-model mean of annual minima decreases by 0.4 m (1.5%) under the SSP5-8.5 scenario, by 0.3 m (1.2%) under SSP3-7.0, by ∼0.1 m (0.4%) under SSP2-4.5 and SSP1-2.6 (Figure 4a); decreases under higher emissions scenarios (SSP5-8.5 and SSP3-7.0) are highly significant (p < 0.001) based on Wilcoxon signed-rank tests that compare the decrease in lake level across models between late 20th and 21st centuries. Under the higher emissions scenarios (SSP5-8.5 and SSP3-7.0) end of 21st century mean annual minima (25.2 and 25.3 m, respectively) fall below the range of historic variability (25.5–25.8 m), but under lower emissions scenarios (SSP2-4.5 and SSP1-2.6) annual minima (25.5 m) are generally lower but remain within the historical range of variability. At the individual model level, annual lake level minima decreases in 89% (n = 24) of models under the SSP5-8.5 scenario and in 100% (n = 27) of models under the SSP3-7.0 scenario when comparing the late 20th to late 21st centuries. These findings signal that Gatún Lake water levels undergo significant declines in their annual minima under high (SSP5-8.5) and moderately high (SSP3-7.0) emissions scenarios, with more moderate changes under lower emissions scenarios (SSP2-4.5 and SSP1-2.6).

An extreme value analysis of lake level minima across all models exhibits a similar bifurcation among higher (SSP5-8.5 and SSP3-7.0) and lower (SSP2-4.5 and SSP1-2.6) emissions scenarios, with significant drops in the return periods of low water level extremes in warmer climates (Figure 4b). For example, lake level minima with 10-year return periods drop 1.0 m (0.9–1.2 m, 95% confidence intervals) under SSP5-8.5, 0.8 m (0.7–1.0 m) under SSP3-7.0, and 0.3 m (0.4–0.2 m) under SSP2-3.5 and SSP1-2.6 when comparing the late 20th century (1980–1999) to the late 21st century (2080–2099). As described above, the GAM-predicted lake-levels are conservative (Figure 3), but projected return periods (2080–2099) under higher emissions scenarios still drop significantly when compared to those derived from the full observational record (1965–2023) (Figure S4 in Supporting Information S1)). For example, the return period of the observed 2023 lake level minimum drops in half from 22 years (11–70 years) to 11 years (9–15 years) under SSP5-8.5, and by a third to 15 years (12–21 years) under SSP3-7.0 by the end of the 21st century. Put another way, the annual exceedance probability for minima (i.e., the probability of occurrence in any year) for the lowest monthly lake level observed (2016) doubles from 2.5% (1%–6%) to 5.0% (3%–7%) by the end of the 21st century under a high emissions scenario. Given the disruption to Panama Canal transits generated by low water levels, the projected increase in the probabilities of extreme lows portends significant challenges to canal operations by the end of the century under moderately high to high emissions scenarios assuming no change in operations.

3.3 Mechanisms of Projected Lake Level Decline

From a mechanistic perspective, lower Gatún Lake water levels projected for the end of the 21st century are primarily associated with large precipitation deficits during the wet season that mirror conditions during observed low water level extremes (Figure 5). Under higher emissions scenarios (SSP5-8.5 and SSP3-7.0), multi-model precipitation anomalies of the late 21st century exhibit precipitation deficits in all months, but these deficits are most pronounced in May–August, reaching −50 mm/month (−70 to −25 mm/month; 95% confidence intervals) in May (Figure 5a). Under lower emissions scenarios (SSP2-4.5 and SSP1-2.6), late 21st century reductions in wet season precipitation are moderate relative to the late 20th century, with a decrease in May of 10 mm/month (−25 to +5 mm/month) under SSP2-4.5 and an increase of 12 mm/month (0 to +25 mm/month) in May under SSP1-2.6. The large wet-season precipitation deficits projected under high emissions scenarios are similar in timing and magnitude to those associated with the lowest 10% of observed lake levels, where El Niño conditions suppress the latitudinal migration of the ITCZ to generate subsidence over Panama (Hatvani et al., 2025). Projected precipitation patterns in earth system model simulations exhibit a similar drying of Central America over the 21st century that is induced by warming of the eastern tropical Pacific and resembles El Niño conditions, similarly driving subsidence and a reduction in wet season precipitation (Almazroui et al., 2021; Brotons et al., 2024; Hidalgo et al., 2013; Nakaegawa & Mizuta, 2023; Vichot-Llano et al., 2021). It is worth noting that these precipitation projections exhibit a large inter-model spread—evident in the wide bootstrapped confidence intervals in our analysis—that reflects the ongoing challenge of accurately simulating precipitation variability in and teleconnections to Central America (Brotons et al., 2024; Simpson et al., 2025).

Reservoir evaporation represents a secondary control on the reduced lake levels under strong greenhouse forcing (Figure 5b). Late 21st century evaporation anomalies are larger under higher emissions scenarios, with enhanced evaporation during both the dry season (January–March) and wet season (May–September), but evaporation anomalies remain small relative to precipitation anomalies. For example, in the SSP5-8.5 scenario evaporation anomalies reach only +8 mm/month (+4–10 mm/month) while precipitation anomalies for this scenario reach −50 mm/month (−70 to −25 mm/month). Nevertheless, evaporation anomalies under high emissions scenarios are, in most months, greater than reservoir evaporation associated with observed low lake level extremes as a result of warming surface temperatures. Taken together, these findings imply that under high emissions scenarios, an average year at the end of the 21st century resembles years of extremely low observed Gatún Lake water levels in terms of precipitation and evaporation.