Based on the classification scheme described in Sect. 3.1.1, we identified 72 pluvial and 34 snow-dominated basins. Some of their main characteristics are presented in Fig. 2. Most of the snow-dominated basins are located in central Chile (30–35∘ S), given the higher elevation of the Andes at these latitudes (Fig. 2a). These basins feature mean annual precipitation values between 258 and 1882 mm (mean of 779 mm), with a marked concentration of precipitation during the autumn and winter months (April to September, Fig. 2b), while most of the streamflow is released during spring and summer months, when snowmelt occurs (October to January; Fig. 2b). Pluvial basins, mostly located towards the south of the study region (35–41∘ S), have mean annual precipitation values ranging from 428 to 3376 mm (mean of 1862 mm), with a significant water accumulation (mostly rainfall) outside the autumn–winter season of maximum precipitation. Streamflow in pluvial basins follows the seasonality of precipitation closely (Fig. 2c).

To visualize the importance of groundwater processes within the study catchments, and their relationship with snow processes, in Fig. 2d we relate the GWI computed from HBV simulations with the SF derived from ERA5-L for each basin (Table 1). These variables come from independent datasets and show a significant correlation (r2=0.49), which indicates the interdependence of snow and groundwater processes. The HBV model calibration resulted in overall acceptable simulation performance in snow-dominated and pluvial basins, with NPE values of 0.72, 0.88, and 0.95 for the 10th, 50th, and 90th percentile, respectively. For comparison, the corresponding Kling–Gupta efficiency values (Gupta et al., 2009) were, respectively, 0.57, 0.8, and 0.92.

In addition to the climatic characteristics (e.g. precipitation, snow fraction, and aridity), GW contribution to runoff depends on physical factors (e.g. geology, topography, soil properties, and soil drainage density), which may explain the large scatter in Fig. 2d. In fact, the geologic characteristics vary across basins, as can be seen in Fig. S1 in the Supplement. The most common geologic classes in snow-dominated basins are acid volcanic rocks (main class in 59 % of basins), followed by acid plutonic rocks (main class in 18 % of basins) and pyroclastics (main class in 18 % of basins). In pluvial basins, there is greater heterogeneity in geologic classes, with 22 % of basins dominated by pyroclastics and 19 % of basins by acid plutonic.

Additional characteristics of snow-dominated and pluvial catchments, including PA–S ratio, QA–S ratio, SF, elevation, area, baseflow index, main land cover class, and granted water used rights are presented in Fig. S1. By construction, there is a clear distinction in PA–S ratio, QA–S ratio, and SF from both clusters, which is translated into differences in mean catchment elevations (mean elevation of snow-dominated catchments is 2667 m a.s.l., while mean elevation of pluvial basins is 599 m a.s.l.). Both clusters present similar areas, and pluvial basins present lower baseflow indices compared to snow-dominated basins. Higher baseflow indices in snow-dominated basins are consistent with the larger GWI simulated in these basins (Fig. 2d).

Regarding the land cover, 68 % of the snow-dominated catchments are mainly covered by barren soil and snow, while the rest is mainly covered by shrubland. None of these land cover classes is directly associated with anthropic activities. 94 % of the snow-dominated basins have less than 5 % of their areas covered by croplands.

The region where pluvial basins are located features a higher heterogeneity of land cover classes, compared to the Andean region of central Chile (where snow-dominated basins are located) (Alvarez-Garreton et al., 2018). The dominant land cover classes in pluvial basins are native forest (main class in 61 % of pluvial basins) and shrubland (main class in 11 % of pluvial basins). Anthropic-related land cover classes dominate the rest of the pluvial catchments: forest plantation in 11 % of basins, grassland in 9 % of basins, and cropland in 7 % of basins.

The correlation between autumn–winter precipitation (PA–S) and the observed streamflow and simulated GW over the following seasons are presented in Fig. 3. PA–S was used instead of annual precipitation since it is more directly related to snow dynamics. Precipitation in autumn and winter represents between 50 % and 100 % of the total precipitation volume (Garreaud et al., 2017).

Figure 3a indicates that PA–S explains more than half of the variance in spring and summer streamflow (QOND and QJFM, respectively) in snow-dominated catchments, which is consistent with the seasonal-lag hydrological memory produced by snow accumulation and melting processes. In general, the PA–S control on streamflow in these basins is stronger during spring and decreases towards summer (larger blue bars compared to cyan bars in Fig. 3a). This memory effect drastically decreases during the subsequent autumn season (QAMJ, represented by the green bars in Fig. 3a) for most snow-dominated basins, which indicates that streamflow during autumn is more influenced by the current seasonal precipitation than by the precipitation during the preceding year. These correlations, representing memory times of 3 to 9 months (Sect. 3.1.3), are consistent with the streamflow and snowmelt seasonality presented in Fig. 2b.

Interestingly, the correlation between PA–S and the spring and summer GW generally increases towards the summer months (Fig. 3b). This indicates a longer response time of GW to the solid autumn–winter precipitation, compared to the streamflow time response. Consistent with this longer GW memory, the correlation to PA–S does not decrease so drastically after summer as streamflow does (larger green bars in Fig. 3b compared to Fig. 3a). Since snowmelt is mostly finished by January, the high r2 values (up to 0.75 in some catchments) indicate that autumn–winter precipitation contributes to the following year's runoff after being stored as groundwater (Fig. 2b).

The control of PA–S on subsequent seasonal streamflow is consistently weaker in pluvial catchments compared to snow-dominated ones, which is expected given the more direct runoff generation from rain. Also, in pluvial basins, the spring precipitation is larger than in northern basins and explains most of the streamflow generated during the same season, hence overshadowing the streamflow dependency to the previous season precipitation. The spring GW memory in most pluvial catchments is stronger than the streamflow memory (larger blue bars in Fig. 3b compared to Fig. 3a), which indicates a 3-month GW memory after winter precipitation. The GW memory during spring decreases towards summer (cyan bars in Fig. 3b) and mostly disappears towards the autumn of the following year (green bars in Fig. 3b close to zero).

To remove the effect of the precipitation of the current year, Fig. 3c presents the dependency of the P–R regression residuals from Sect. 3.2.2 (i.e. annual streamflow not explained by the current precipitation) to the precipitation from the previous year. The r2 values in Fig. 3c (all corresponding to positive correlation coefficients; not shown here) represent the control that the precipitation from the previous year exerts on the annual streamflow and thus indicate hydrological memory beyond 1 year. Interestingly, although the residual r2 values are larger in snow-dominated catchments, there are also significant values in the pluvial basins, suggesting that residence times associated with GW processes in pluvial catchments also contribute to more extended catchment memory.

The relationship between hydrological memory beyond 1 year and catchment storage dynamics driven by GW and snow processes is further explored in Fig. 4. The GWI derived from HBV simulations (Sect. 3.1.2) and SF (Table 1) are used to summarize GW and snow processes within a catchment in Fig. 4. These plots show that both snow and GW mechanisms contribute to the hydrological memory, which is consistent with the time lags these processes introduce to the pathways of precipitation within a basin. The adopted approach to compute the memory of a catchment, based on seasonal streamflow and GW flows at the catchment outlet, represents the composite response times of all catchment mechanisms. Therefore, there might be other factors contributing to the overall response time of a catchment, including topography, soil properties, geology, drainage area, and water table levels (Robinson and Ward, 2017), which may explain the large scatters in Fig. 4.

Given the marked north to south precipitation gradient (Fig. 1), droughts characteristics in this section are analysed at the catchment scale but following a latitudinal order. In the following sections, we analyse drought propagation per hydrologic regime. Heat maps in Fig. 2 illustrate the precipitation and streamflow annual anomalies (Fig. 5a and b) and z scores (Fig. 5c and d). The unprecedented dry conditions during the last decade are evident. The temporal and spatial extent of precipitation deficits are shown by general negative anomalies for the 106 catchments between 30 and 41∘ S (Fig. 5a). Over the previous 3 decades, there are few analogues showing such a spatial pattern, such as the 1988–1990 3-year drought and the single-year droughts in 1996 and 1998 (one of the driest years in the last millennium; Garreaud et al., 2017) and 2007. The decade-long persistence of the MD spatial pattern, however, has no analogue in the studied period, nor in the last century (Garreaud et al., 2017).

The relative precipitation anomalies are consistently higher in catchments north of 35∘ S (Fig. 5a), which is partly due to the very low annual precipitation values in the northern region compared to the southern region. The probability of having high anomalies is larger in this northern region (z scores closer to zero in Fig. 5c), where few meteorological events contribute to most of the annual accumulation. Thus, the interannual variability is higher than in the southern region (south of 35∘ S, where most pluvial basins are located, Fig. 2), where several precipitation events occur during the year. During the MD, annual precipitation in basins south of 35∘ S decreased up to 1.5 times below the historical standard deviation of annual precipitation, which represents an exceedance probability of 94.3%. The z scores during the MD have usually been lower in the southern region compared to central Chile (30–35∘ S).

The annual streamflow relative anomalies (Fig. 5b) present larger values and larger variations in space and time than precipitation, which is due to the lower absolute values compared to precipitation and the dependency of streamflow on local terrestrial characteristics. If we compare the spatial patterns of anomalies (Fig. 5b) and z scores (Fig. 5d) during the MD, we see that consistently with precipitation, larger anomalies have been observed in catchments north of 35∘ S (extreme dry years up to 90 % of streamflow deficits with a median of 57 %). Still, the moderate deficits experienced in the southern region (median deficit of 25 %) have a lower probability of occurring (lower z score values, indicating higher exceedance probabilities).

To further characterize the hydro-climatic anomalies during the MD, Fig. 6 presents the frequency distribution of 10-year mean precipitation, observed streamflow, and simulated ET for each basin, with the mean values during the MD plotted by red dots. These plots indicate that the MD has been extremely unusual in terms of precipitation and streamflow. The average precipitation during the MD is within the first decile for 91 % of the study catchments (96 out of 106). The average runoff during the MD has been more extreme than precipitation deficits, with 96 % of catchments (102 out of 106) presenting 10-year mean runoff values within the first decile. These values represent the minimum value over the last 4 decades for some basins located north to 32∘ S and correspond to an outlier of the historical distribution for the rest of the catchments. The average 10-year ET during the MD has been less unusual than precipitation and streamflow anomalies, with only 20 % of catchments (21 out of 106) within the first decile and 73 % of basins below the median (quantile 0.5) historical 10-year ET values. This indicates that in general, the estimated ET within the study basins during the MD has not been as different from normal conditions as precipitation and streamflow. This also shows that, despite the higher mean temperatures due to the warming trend experienced in the region (Boisier et al., 2018), ET is closely related to precipitation, notably in the northern (water-limited) basins.

Given the relatively small scale of catchments in Chile, runoff in most of them has a strong dependency on the interannual precipitation variability, explaining typically ∼75 % of the streamflow variances. Yet, the annual P–R relationships observed during the MD changed significantly for some of the catchments within the study region (examples in Fig. 7b, c, e, and f), while other catchments showed no significant change (Fig. 7a and d). Table 2 summarizes the number of catchments that experienced a significant shift in P–R relationship during the MD and their associated shift magnitudes. This table also shows the results for those catchments with no change. From the 106 studied catchments, 61.3 % had a significant change in the P–R relationship during the MD, and 38.7 % had not changed (Table 2).

For the 65 catchments showing a change, the historical P–R regressions consistently underestimate the runoff deficits during the MD. 95 % (62 out of 65) of the catchments with a significant change in P–R relationship during the MD, had a negative shift, that is, an intensification in drought propagation. For similar precipitation deficits to other dry years, observed streamflow during the MD in snow-dominated catchments was up to 57 % lower than predicted by the historical P–R relationship. In pluvial catchments, these shifts reached up to 37 % (Table 1).

For those catchments with a significant change, higher GWI values are associated with larger shifts in P–R relationships (r2=0.40, not shown here) and thus with the intensification of drought propagation during the MD, with respect to historical annual responses to droughts. The relationship between snow fraction and P–R shifts is weaker (r2=0.29, not shown here), probably because snow processes control the hydrologic response of a subset of catchments experiencing a significant change in P–R relationships, whereas GW is inherent to all basins.

While these results provide insights about changes in P–R relationships during a multiyear period, they do not indicate if the changes are progressive, i.e. if the basins progressively generate less water for a given precipitation amount, compared with their historical behaviour. We address this in the following section, where annual drought propagation is analysed in detail.

Figure 8 presents the time series of annual precipitation and streamflow anomalies averaged across snow-dominated catchments (Fig. 8a) and pluvial catchments (Fig. 8b) over the last 4 decades. Drought propagation is represented by the difference between average runoff and precipitation anomalies (secondary y axis in Fig. 8). To focus on drought propagation and not on the propagation of positive anomalies, the difference in runoff to precipitation anomalies was computed only for those years with negative streamflow anomalies (red line in Fig. 8).

In snow-dominated basins, the difference between runoff and precipitation anomalies consistently increases (becomes more negative) in the second year of precipitation deficits, showing that in catchments with longer hydrological memory, consecutive years with precipitation deficits are associated with intensified drought propagation. This plot also provides insights about hydrological recovery, understood as the hydrological condition after a meteorological drought has ceased (Yang et al., 2017). While 2016 had near-average precipitation in snow-dominated basins, there was probably not enough water entering the system over enough time to recharge groundwater systems up to levels such as those before the MD (similarly to the conceptual drought propagation illustrated in Fig. 3 from Van Loon, 2015). This is reflected by the larger streamflow deficits in 2016 compared to 2008, even when the above-mean precipitation values in 2008 following the deficits in 2007 are comparable to those in 2016 and 2015, respectively. This can be related to the catchments' memory (Sect. 4.2) and the 7-year (2009 to 2015) precipitation deficits prior to 2016, which probably prevented a full hydrological recovery after a single year of above-average precipitation. These results are consistent with large recovery times reported for semi-arid Australian catchments following extreme droughts (Fowler et al., 2020; Yang et al., 2017). In this way, hydrological memory would be an explanatory factor for both the intensification in drought propagation and a delayed hydrological recovery.

For pluvial catchments (Fig. 8b), prior to 2010, the propagation of meteorological to hydrological drought has ranged from around 0 % to 15 % (i.e. streamflow anomalies have been, on average, up to 15 % lower than precipitation anomalies), independently of the precipitation deficits of previous years. Such propagation was observed even in the driest 2 years of the historical record, 1996 and 1998. After 2010, there are 2 years (2012 and 2016) where drought propagation has been intensified up to 25 %. These larger streamflow deficits during the MD may be due to different factors, including the large ET in 2012 and 2016 (positive anomalies and z scores above 1.5, as can be seen in Fig. S2) combined with large precipitation deficits.

Regarding the overall response during the MD, since P–R relationships do not explicitly account for variations in ET, the positive ET anomalies during specific years of the MD (2012, 2016 and 2018, Fig. S2) may partly explain the P–R shifts identified in some pluvial catchments (Fig. 7). Towards the south of the study region, basins move from being water-limited to energy-limited (Alvarez-Garreton et al., 2018). Therefore, ET in pluvial catchments is modulated by both the available water and the available energy, in contrast to snow-dominated catchments, where ET is primarily driven by precipitation (these are water-limited basins). This suggests that ET may be a factor influencing the intensification of drought propagation during the MD in pluvial catchments. However, despite higher ET in 3 years during the MD, the average ET during the 10-year MD period has been lower than other 10-year windows (Fig. 6c). Therefore, the hydrological memory beyond 1 hydrological year in pluvial basins (Fig. 3) is likely another factor contributing to the P–R shifts in pluvial catchments.

Figure 9 presents the observed and simulated drought propagation during single-year severe and extreme droughts and during persistent mild and moderate droughts, for snow-dominated (Fig. 9a and b) and pluvial catchments (Fig. 9c and d). Consistent with the hydrological memory in snow-dominated catchments, we observe that drought propagation in these basins is highly dependent on the meteorological conditions from the previous year (Fig. 9a), which define the initial condition of soil water storages. If a severe or extreme drought happens after a wet year, such as 1981, 1985, 1988, 1998, 2007, and 2009 (Fig. 9a), drought propagates without amplification; i.e. streamflow deficits are lower than precipitation deficits. By contrast, if the extreme drought happens after a dry year, such as 1995, 1996, 2013, and 2019, meteorological droughts are amplified by nearly 20 % (difference between median streamflow and median precipitation deficits in Fig. 9a). This is also observed in persistent but moderate droughts, when under similar precipitation deficits, the surface water supply decreases after 1 year with below-average conditions.

If we look at streamflow predictions, these plots indicate that the HBV model represents catchment response for extreme and persistent droughts well, consistently outperforming the prediction from P–R regressions. It should be noted though that HBV allows for memory effects only to a certain extent given the groundwater storage capacity defined by calibration.

Pluvial basins in the study region have shorter hydrological memory compared to snow-dominated catchments, which leads to a more similar behaviour under extreme meteorological droughts occurring after a wet and a dry year (difference in streamflow to precipitation anomalies between 10 %–20 %; Fig. 9c). However, even when these basins are largely controlled by precipitation during the same year, there is some memory that is over 1 year (Fig. 3c), which may be influencing the observed decrease in streamflow generation after 2 years of consecutive precipitation deficits (Fig. 9d). This effect is well captured by the HBV model, while the annual P–R relationship tends to overestimate observed runoff. This indicates that a good representation of fluxes such as ET, soil moisture, and groundwater dynamics, allows catchment response to persistent droughts and to extreme droughts to be foreseen.