The long-term ERA5 data are downscaled from ∼30 km to the local conditions (∼10–100 m) observed at CEAZA's met-stationSDH (Fig. a). Downscaling is performed using artificial neuronal network (ANN) algorithms . The ANNs are solved using 10 hidden layers as well as the Levenberg–Marquardt training algorithm. This process is performed with MATLAB's Neural Net Fitting tool. Air temperature (T), specific humidity (q) and wind speed (U) from the ERA5 dataset are used as input data for training and validation of the ANNs, whereas T, RH, U, WD and Swin collected at met-stationSDH are used as target data. Note that conditions observed at the met-stationSDH (2 m) show the same variabilities and magnitudes as the meteorological observations obtained by the ECwater above the saline lake (1 m, see Fig. b) during the E-DATA field experiment.

As validation, Figs.  and  show the time evolution and orthogonal regression of the ERA5 downscaled and raw variables of T, q and U compared to surface observations of the met-stationSDH and ECwater. In terms of temperature, Fig. a shows that there are significant differences in the diurnal cycle of T between the ERA5raw data and the observations of met-stationSDH and ECwater, especially at lower temperatures. Nonetheless, the temperatures observed above the water are in agreement with the values from ECwater (1 m) and met-stationSDH (2 m). Therefore, we can assume that T above the water and above the land are similar. This similarity allows us to validate ERA5down results on the saline lake using the data observed by the met-stationSDH. Figure a shows a satisfactory correlation between the ERAraw and the met-stationSDH observations (R2=0.95), but with a low slope (m=0.5). This mismatch is overcome when we apply the downscaling, where T increases the correlation coefficient (R2=0.97) and the slope (m=0.92) (Fig. a).

For q, there is more scatter in the met-stationSDH observations, which results in low R2=0.38 (Fig. b). However, similar to temperature, we observe an improvement after the downscaling, where ERA5 data increases the slope in the orthogonal regression from 0.43 to 0.77. Although the agreement between q-ERA5 and observations is lower than T-ERA5, q-ERA5 has a reasonable agreement with observations in the diurnal cycle (Fig. b).

In terms of U, we observe more differences between ECwater and met-stationSDH during the maximum values. The differences are related to the surfaces above which the instruments are installed, i.e., ECwater above the saline lake and met-stationSDH above bare soil (Fig. ). However, our statistical calculations corroborate the benefits of using the downscaling methods: R2 increases from 0.81 to 0.92 and slopes from 0.59 to 0.85 as compared to ERAraw. Finally, although WD is not used to estimate the evaporation and not shown in the plots, the ERA5 data have good agreement with observations.

To estimate the actual evaporation, we employ an adapted version of the equation for open water evaporation , expressed in energy terms LvE. Our approach is to use standard meteorological data of T, q, U and Swin from the downscaled ERA5 dataset, and apply it to the specific conditions of the SDH shallow lake. The adapted version of the Penman equation reads as 1LvE=cice(ss+γcEBNC(Rn-G)︷Radiative+ρacps+γ1ra(es-e)︷Aerodynamic), where s [PaK-1] is the slope of the saturated vapor pressure curve, γ [PaK-1] is the psychrometric constant, Rn [Wm-2] is the net radiation, G [Wm-2] is the ground heat flux, ρ [kgm-3] is the dry-air density, cp [JK-1kg-1] is the air's specific heat at constant pressure, ra [sm-1] is the aerodynamic resistance, esat [Pa] is the saturated vapor pressure, and ea [Pa] is the vapor pressure at measured level. The ice coefficient cice [–] is a correction coefficient that represents the evaporation reduction that occurs when an ice cover is formed above the saline lake , and cEBNC [–] is the energy balance non-closure coefficient, which corrects the available energy (Rn-G) to improve the energy balance closure. Note that Eq. () becomes the equation when cice=cEBNC=1. Appendix A describes the details of the calculation for each term in the Eq. ().

To evaluate the diurnal variability of evaporation, the evaporation estimates are compared with observations using orthogonal regression, where we estimate the error employing the root mean squared error (RMSE), the mean absolute error (MAE), and the correlation (R) and determination (R2) coefficients. The climatology of evaporation estimates and precipitation data obtained from ERA5 are analyzed at seasonal and interannual scales. For seasonal timescales, we use descriptive statistics of mean, maximum, minimums and quantiles (25, 50 and 75) for each averaged month over the entire period (1950–2020). For the interannual timescales, we calculate monthly anomalies as the difference between the 12-month moving average and the mean of the entire period under study. Our reason for using the moving average is to decrease the high scatter that monthly means produce and better evaluate the ENSO and PDO influence on the evaporation and precipitation.

To get an overview of the moisture transport that results in precipitation over the Altiplano region and surrounding areas, we selected the moisture sources of a selected region spanning from 83∘ W to 57∘ E, and from 11∘ N to 27∘ S (Fig. ). For that, we use ERA-Interim data from 1997–2018 to force the Water Accounting Model-2layers WAM-2layers;. The WAM-2layers is an Eulerian offline moisture tracking model which solves the atmospheric water balance for every grid cell. Tracking is performed on two layers in the atmosphere, hence the atmospheric input variables from ERA-Interim are integrated over two layers. Well-mixed conditions are assumed for both layers. More information on the model is given by and . Seasonal averages of moisture sources are evaluated (1997–2018; summer (JFM), autumn (AMJ), winter (JAS) and spring (OND)) together with the direction and intensity of the moisture fluxes.

The long-term water balance in the saline lake is assessed by combining the mass conservation principle with actual evaporation estimates and precipitation data. Evaporation estimates are obtained using data from the downscaled ERA5 and the site-adapted Penman equation (Sect. 2.3.2), whereas precipitation data were obtained from the raw ERA5. The mass balance is evaluated as follows: first, the volume of the lake in a specific month is estimated using the lake's area and assuming a constant lake depth that varied between 0.05 and 0.20 m. Second, we estimate the monthly lake outflow using the actual evaporation values and the lake's area, assuming no groundwater outflow (endorheic basin). Third, we determine the volume reduction of the lake due to evaporation by subtracting the volume of water evaporated in a month from the volume of the lake. Then, the lake area of the next month is computed dividing the lake's volume by its depth. This area is compared to that obtained using remote-sensing data to determine the additional monthly water volume required to achieve the observed lake surface. By associating this additional water input with precipitation, we determine the areal extension of precipitation that contributes to the representation of the observed areas of the lake. Because most of the time there are no surface water inputs (negligible surface runoff is observed), this additional water source must represent groundwater inputs into the lake. The approach followed here is a first order approximation that can be used to understand the key components of the lake's water balance. However, we believe that more precise information is needed to reproduce the seasonal variability of groundwater flow.

Figure a shows the actual evaporation seasonal average from 1950 to 2020 over the saline lake of SDH. In general, seasonal changes of evaporation show their highest monthly values (>90 mm) during the austral summer (JFM) and spring (OND). Within these seasons, October, November and December present the highest monthly evaporation (107–120 mm). Even though the summer also presents high monthly evaporation (90–107 mm), these months also show the highest variability (standard deviation of 13.5–16.5 mm). The variability observed during the summer months is because of the rainy season that usually extends over the summer . Evaporation has its lowest rates during autumn and winter (<78 mm per month). Moreover, within these seasons, the months of June, July and August show the lowest monthly evaporation (∼50 mm) and the lowest variability of the year (standard deviation of 7 mm per month). On the other hand, the seasonal variability of precipitation is shown in Fig. b. Precipitation in the SDH basin shows a very clear seasonal cycle, with the onset of the rainy season in late spring (ND) and the offset at the end of summer (MA). However, this rainy season presents high variability over the years. The rest of the seasons show precipitation values below 25 mm per month, where June and July present a slightly higher variability.

To give a synoptic-scale perspective of the seasonal changes in local evaporation and precipitation presented in the saline lake of SDH, Fig.  shows the seasonally averaged moisture sources of the Altiplano region. Here, we quantify the regions where evaporation occurs which results in precipitation over the Altiplano region (gray box in Fig. ). As most precipitation occurs in the austral summer (Fig. ), the moisture sources are also largest in these seasons. We observe three principal moisture sources that contribute to precipitation in the Altiplano region during the year. The first one comes from the northeast (Amazon basin) and results from the veering of trade winds southwestwardly into the Andes Mountains, associated with the continental low formed by the summery position of the Intertropical Convergence Zone (ITCZ) south of the Equator . This southwestward flux is most pronounced during summer, transporting moisture (∼50 mm) into Altiplano region. This marked moisture flux suddenly decreases towards the autumn and winter (∼10 mm). During these seasons, the trade winds return to their normal westwardly direction (Fig. b and c), resulting in low moisture transport (∼20 mm) into the region. Besides moisture transport into the region from the northeast, there is also recycling of moisture within the region, which can be considered to be a second moisture flux. Especially during summer, evaporation contributes to precipitation within the region, as can be seen by the high moisture source values around the lake in SDH during JFM and OND (Fig. a and d). In addition to the contributions from evaporation over land, there is also a positive moisture source from the Pacific Ocean (south–southwest). In the absence of precipitation over the ocean, we can assume that the evaporation over the ocean contributes to the precipitation over land in the Altiplano region. Finally, this third moisture flux is associated with the subtropical anticyclone and stratocumulus cloud deck . This flux transports a very low but persistent amount of moisture into the Altiplano region (<5 mm) due to the steep topography presented on the western slope of the Andes Mountains, which in combination with the anticyclone, limits the eastward flow up to the mountains. Despite the coarse model resolution, this low-moisture transport has been reported using high-resolution modeling and airborne observations by and .

To unravel the processes involved in the seasonal evaporation, we analyze the seasonal variability of the drivers that control it. Figure  shows the seasonal cycle of the radiative and aerodynamic contribution of the Penman equation and subsequent correlations with monthly evaporation rates.

Figure a shows the seasonality of the radiative contribution to evaporation, with its highest values corresponding to spring and summer, slowly decreasing towards the winter, only to increase again in early spring. The seasonality of the radiative contribution is similar to that of evaporation shown in Fig. b, but it presents two distinctive characteristics. Firstly, from November to March, there is a larger scatter (standard deviation >12 Wm-2), where the radiative contribution to evaporation can be high at 170 Wm-2 or low at 20 Wm-2. This large variability is directly related to the summer rainy season , where the presence of clouds largely modulates the available net radiation . This double feedback that precipitation has over the radiation might explain the large scatter in the radiative contribution to evaporation during spring–summer. Secondly, the small variability (standard deviation of ∼3 Wm-2) of the radiative contribution during the winter months is related to the atmosphere's stability, characterized by the dry weather and cloudless conditions during most of this period. Therefore, there is enough evidence to support the idea that radiative contribution controls evaporation at a seasonal scale (R2=0.91, as shown in Fig. b).

Figure c shows the seasonality of the aerodynamic contribution to evaporation, where the highest and lowest values are observed in early spring (SON) and during summer (JFM), respectively. The variability of the aerodynamic contribution is fairly constant during the whole year (standard deviation of ∼4 Wm-2), which is related to the seasonality of the wind circulation patterns . The wind seasonality also explains the highest and lowest aerodynamic contribution to seasonal evaporation. For example, the thermal contrast between the Pacific Ocean and the Atacama Desert reaches its maximum in November, resulting in the strong regional atmospheric eastward flow, responsible for the onset of diurnal evaporation in the SDH . To the contrary, during summer, predominant westward regional circulation from the Amazon basin counteracts the eastward regional flow , decreasing the wind speed (as described below). Finally, during winter, the lower thermal contrast between the Pacific Ocean and the Andes Altiplano, along with the absence of the summer westward regional flow, results in lower wind speed. Consequently, there is less aerodynamic contribution to evaporation. The scattered seasonality of the aerodynamic contribution to evaporation also results in a low correlation (R2=0.34, as shown in Fig. d).

In summary, at the seasonal timescale, the radiative contribution term contributes significantly more to evaporation than the aerodynamic term, representing 73 % of the energy needed to evaporate the water from the saline lake in SDH. It is important to stress that mechanical turbulence (wind speed) is more relevant at the diurnal scales than available net radiation controlling evaporation .

To complete the seasonal analysis of evaporation in recent decades, we describe the spatial impacts of the evaporation–precipitation variability on the saline lake of SDH. Before showing the results, it is interesting to mention the heterogeneous characteristics of open waters and different types of salty crusts . These salty crusts cover larger areas than open-water surfaces, contributing significantly to the basin's water balance. With respect to the wet/dry salt contribution, the salt crust found in SDH has particularly low evaporation (<50 Wm-2, e.g., see , Fig. 3b). Despite this low rate, it is still relevant since the open-water/wet salt surface proportion has high seasonal variability. Based on our flux data (4.3 mmd-1 for open water and 0.5 mmd-1 for wet salt) and surface areas estimated from satellite remote-sensing observations in November (1.8 km2 for open water and 13.1 km2 for wet salt), we estimate that open-water evaporation is 7.8 against 5.6 m3d-1 for wet salt. In the rest of the section, we focus on the saline lake's water balance as an entity integrating the different contributions. Figure  shows the relationship between the spatial changes of the saline lake and monthly evaporation and precipitation that occurred between 1985 and 2019. The seasonal variability of the lake's area shown in Fig. a reveals that the maximum extension (5 km2) occurs during winter (JJA). During spring, the lake's area decreases rapidly to its minimum extension (∼1.3 km2). The summer season shows high variability in the lake's area (mean: 2 ± 1.8 km2; mean value ± standard deviation). This variability is relatively constant towards winter and decreases during spring, revealing that there is a significant interannual variation over the years, especially between March and July, where precipitation is typically small (Fig. d). Thus, the increase in the lake's area is likely due to groundwater inputs .

Regarding the relationship between the lake's area changes and evaporation, Fig. b shows an orthogonal regression between evaporation and lake extension changes. Here, we find a strong negative correlation (R2=0.92), which reveals the control that evaporation has over the lake discharge. The lowest evaporation rates (winter) coincide with the highest lake extensions, and the highest evaporation rates (spring) coincide with the lowest lake surface. Regarding the relationship between precipitation and the lake's surface associated with the water recharge by precipitation, the relationship is indistinctive (Fig. a). We find a high variability in the onset and offset of precipitation at the seasonal scale, from November to March. This high variability in summer precipitation coincides with the larger variations in the lake's area. As such, it is difficult to find a direct relationship between precipitation and the lake's area. However, analyzing the means (solid lines Fig. a), we observe that high precipitation rates do not directly impact the areal changes of the lake. In February, the lake's surface reaches a first maximum, which might be related to precipitation in the direct proximity of the lake, generating enough surface runoff to enhance the water amount of the lake. However, the highest values of the water-lake surface is reached 4–5 months after the rainy season. For these reasons, the observations suggest that there is another process that modulates the lake recharge. Thus, the alternative that explains lake recharge is groundwater, which is fed by precipitation in the headwaters of the basin.

To unravel the role of groundwater input into the SDH lake, we perform a simple lake water balance test assuming a lake depth between 0.05 and 0.20 m. Our lake mass balance results show that the monthly water required to represent the spatial changes in the lake's surface are in the order of ∼345000 m3permonth (∼0.1 m3s-1). This estimation is reasonable as the only stream flowing in the lake direction has an average flow of 0.13 m3s-1 , which is measured about 1–2 km before the river water completely infiltrates into the ground. If one assumes that ERA5 precipitation is responsible for this water flow, then a lake area of ∼13 km2 is needed to explain it. When varying the water depth between 0.05 and 0.20 m, our results changed less than 1 %. As the mean observed lake's area is ∼2 km2, groundwater is the water source that sustains this habitat. This result agrees with the estimations performed by . They quantified a flow in the range of 0.14 to 0.2 m3s-1 in the springs that discharge water into the lake, which is similar to the flow estimated in our work. It is important to recall that our approach has important limitations. For instance, as the topography in the basin's sink is very flat, there is no hypsometric curve that can relate the lake's volume as a function of depth. Also, the precipitation considered here corresponds to that estimated in the lower part of the basin, whereas higher precipitation values occur at higher elevations in the basin . Hence, most of the groundwater recharge is expected to occur at higher elevations and/or in locations where preferential flow exists, e.g., near the rivers of the basin. Then, this water will flow underground until it upwells into the lake. So, even though this approach has limitations, it allows for a first order approximation that can be used to understand the key components of the lake's water balance.

Evaporation trends in the saline lake of SDH show an indubitable increase from 1950 to 2020. The rate of increase is about 2.1 mmyr-1 (0.2 mm per month), with scattered interannual variability showing a significant increase. Figure a shows a 12-month moving average of monthly total evaporation. For 1950, monthly mean values are approximately 80 mm (950 mmyr-1), whereas in 2020, these values increased to ∼100 mm (1150 mmyr-1). The annual integrated evaporation rates averaged 1075 mm (±74 mm) with a minimum of 862 mm (1993) and a maximum of 1210 mm (2010). This increase in evaporation has a correlation of 0.55 with air temperature (2 m), whose monthly averages increased 3 ∘C (0.04 ∘Cyr-1), from 1950 to 2020 (Fig. a). Likewise, Fig. b shows the precipitation trends in the area of the saline lake from 1950 to 2020. Total precipitation per year is set at 338 mm with a high variability of 248 mmyr-1. Precipitation shows an increasing trend of 0.6 mmyr-1 in the last 70 years. Although this positive trend in precipitation is less significant than evaporation and presents more scatter, it is also in agreement with temperature increase.

Figure  shows the seasonal variability of monthly evaporation in the period 1950–2020 during cool (ONI<-0.5 ∘C), neutral (-0.5∘C<ONI<0.5 ∘C) and warm (ONI>0.5 ∘C) ENSO phases. In general, during cool ENSO phases, evaporation rates are 2 % lower than those observed in the neutral ENSO phases, whereas during warm ENSO phases, evaporation is 15 % higher than that observed in neutral ENSO phases. This variability becomes more significant from October to May, summer (JFM) being the season with the largest variability. During summer, evaporation under cool ENSO phases decreases by 4 % with respect to neutral phases and increases by 14 % under warm conditions. Moreover, summer variability is the highest during warm ENSO phases, showing standard deviations of ∼15 mm per month. The lowest evaporation variability occurs during the neutral phases, with standard deviations of 11 mm per month. In turn, during late autumn and winter seasons, the ENSO phenomenon influences the evaporation less in the saline lake of SDH since evaporation rate differences between cool, neutral and warm phases are lower than 2 %. This analysis suggests that ENSO significantly influences evaporation during summer months, which is in line with other typical meteorological phenomena of the Atacama Desert, such as summer precipitation and coastal fog formation .

The ENSO phases' influence on evaporation observed at the seasonal scale is also present interannually. Figure  shows the relationship between ENSO phases and PDO phenomenon, with evaporation anomalies obtained using the site-adapted Penman equation (Eq. 1) and downscaled ERA 5 data, and precipitation anomalies observed from ERA5 in the last 7 decades in the shallow lake of SDH. Recall that ENSO is a recurrent phenomenon with an ill-defined periodicity, where in the last 70 years, 24 % of the months have been influenced by warm ENSO phases and 26 % by cool ones. However, the frequency of this phenomenon is not constant, neither in intensity nor in time. The ONI varied between 0.5 and 2.6 ∘C during warm phases, and between -2 and -0.5 ∘C during cool phases, with a frequency between 2 and 10 years .

Figure a shows the 12-month moving average evaporation anomalies and the ENSO and PDO phenomena from 1950 to 2020. Positive monthly evaporation anomalies (>5 mm) correlate with warm ENSO phases, whereas negative or non-evaporation anomalies (<0 mm) correlate with cool ENSO phases. The correlation between positive evaporation anomalies and warm ENSO phases is evident during the extreme ENSO events, i.e., events which occurred in 1983, 1997 and 2015. Similarly, the correlation between negative evaporation anomalies and cool ENSO phases is evident in 1988, 1998 and 2010. However, this trend is indistinct when monthly evaporation anomalies are close to 0 mm (e.g., in 1970, 1995 and 2001). Evaporation anomalies also have an interdecadal variability. For instance, between 1950 and 1975, negative evaporation anomalies dominate. On the contrary, between 2000 and 2020, positive evaporation anomalies dominate, but only after a transition period that occurred between 1975 and 2000, where both positive and negative evaporation anomalies are present. Regarding larger macroclimatic phenomena, no significant correlation is found between PDO and evaporation anomalies.

The influence of the ENSO phenomenon also affects precipitation at SDH. Figure b shows the 12-month moving average precipitation anomalies and the ENSO and PDO phenomena from 1950 to 2020. The influence of ENSO on precipitation is the opposite of that observed for evaporation. Here, positive monthly precipitation anomalies (>5 mm) correlate with cool ENSO phases, whereas negative monthly precipitation anomalies (<5 mm) correlate with warm ENSO phases. Contrary to evaporation anomalies, the relationship between precipitation and extreme ENSO events is indistinctive. For example, strong precipitation anomalies observed in 1985 disagree with an extremely cool ENSO phase. The same occurs for the extremely cool ENSO phase that occurred in 1999, where the precipitation anomaly is not correlated with high positive precipitation anomalies. However, the negative correlation trend between ENSO phases and precipitation anomalies is still evident. Precipitation anomalies also have an interdecadal variability that seems to be related to PDO anomalies. For example, between 1950 and 1970, there is a predominance of negative precipitation anomalies, which correlate with negative PDO indices. However, between 1970 and 2000, positive precipitation anomalies predominate along with positive PDO indices. Finally, between 2000 and 2020, negative precipitation anomalies predominate together with negative PDO indices. The negative relationship between precipitation and ENSO phases in the Altiplano region has also been reported by , , and .

To further quantify the opposing trend between evaporation and precipitation, Fig.  shows the relationship between evaporation and precipitation anomalies categorized by ENSO phases. The trend between cool ENSO phases and negative evaporation anomalies is significant (∼-10 mm), although it is weaker during extremely cool phases. Similarly, the trend between warm ENSO phases and positive evaporation anomalies is very clear, even during the most intense warm ENSO phases (>15 mm). Regarding precipitation anomalies, the trend shows a similar pattern, where negative precipitation anomalies (∼-15 mm) are related to extremely warm ENSO phases, and the highest positive precipitation anomalies are related to both cool and neutral ENSO phases.

Opposing behavior of ENSO influences on interannual evaporation and precipitation variability demonstrate the control that global climate phenomena can exert at a local scale in the long term. As shown in Fig. a, air temperature is strongly related to evaporation; thus, atmospherically warmer conditions in the Altiplano region during warm ENSO phases enhance evaporation. This warming intensifies the Pacific anticyclone through the tropospheric thermal stratification , resulting in cloudless conditions during the summer of warm ENSO phases, i.e., an increase in the radiative contribution term of 17 % as compared to the cool phases. Increased radiation also leads to an enhancement of the ocean–land thermal contrast, enhancing the aerodynamic contribution to evaporation by 21 % during warm ENSO phases in summer with respect to cool ENSO phases. This enhanced ocean–land thermal contrast also increases the atmospheric capacity to hold water vapor. Conversely, cool ENSO phases promote higher precipitation rates through the weakening of the Pacific anticyclone and the strengthening of the Bolivian low , which negatively affects the evaporation in two ways. First, the cloudy conditions that result from wet seasons inhibit the available energy required for evaporation (from 120 to 100 Wm-2), mainly affecting the evaporation rates during the summer season (Fig. ) as well as interannual rates (Fig. a) . Second, during cool ENSO phases, a strong rainy season attenuates the characteristic regional atmospheric flow from the Pacific Ocean into the Andes , significantly affecting the aerodynamic contribution to evaporation (Fig. b), decreasing it from 38 to 30 Wm-2.