São Paulo drought: trends in streamflow and their relationship to climate and human-induced change in Cantareira watershed, Southeast Brazil

The 2013–2015 drought in the Metropolitan Region of São Paulo exposed the lack of resilience of the regional water supply system, highly dependent on the Cantareira reservoirs. In this paper, inflows to each of the four main Cantareira reservoirs are tested for systematic change. Persistent trends in streamflow, rainfall, temperature and evapotranspiration are first evaluated. Streamflow was also tested for step change. Double-mass curves were employed to assess modification in the precipitation–runoff relationship. Subsequently, we used the climate elasticity method and the ABCD model to quantify the relative contribution of climate and human activities into the detected trends. Only Cachoeira and Atibainha sub-basins showed a significant downward trend in streamflow. The results for step change were also significant, and the year of occurrence coincided with breakpoints in precipitation–runoff relationship. For both Cachoeira and Atibainha, human activities had a more significant impact on streamflow reduction than climate variability. Land use and cover maps suggest that the reduction of pasture/abandoned land parallel to an increase in reforestation/ silviculture is behind streamflow reduction. The results highlight the importance of coordinating land-use patterns and water management, as an important contributor beyond any considerations of a changing climate. Implications for better managing regional water resources are discussed.


INTRODUCTION
Ensuring the reliability of water supply services in the Metropolitan Region of São Paulo (MRSP) is increasingly a challenge due to social, climate and environmental dynamics. In 2016, the MRSP population was estimated to be 20.6 million (SEADE Foundation ), representing about 10.3% of Brazil's total population, while covering less than 1% of Brazil's land area (IBGE ). Although Brazil accounts for nearly 20% of the world's water reserves (Ceratti ), most of them (74%) are in the Amazon, rather distant from the main consumption centers. Annual water availability in the Alto Tietê River Basin, where the MRSP is located and which supplies water for part of its population, is merely 201 m 3 per capita (World Bank ).
Consequently, the MRSP becomes highly dependent on water diversion from nearby river basins. The main source is the Cantareira system, a network of reservoirs which transfers water from the Piracicaba River Basin, being responsible for serving nearly half of the MRSP population.
The Metropolitan Region of Campinas (MRC), located downstream of the Cantareira reservoir system, also depends on the Piracicaba River Basin to meet its own needs. The MRC is a highly industrialized and urbanized area, with high population and economic growth rates.
These circumstances lead to conflicts in water demand, especially with an expected growth in demand for both the MRC and the MSRP (ANA & DAEE ).
The fragility of MRSP water supply has been exposed by recent extreme climate events. During the 2013-2015 drought in Southeastern Brazil, the anomalously low inflows to the Cantareira system caused reservoir storage to be reduced beyond minimum operational level (Sabesp ).
Emergency structural and non-structural measures were undertaken to prevent the water supply system from collapsing (Braga & Kelman ). The restrictions imposed on the water supply were accompanied by an increase in water tariffs, prompting social protests and riots in different parts of the MRSP. The impact of the drought on crops caused food prices to rise. More than 60,000 industries in the state of São Paulo, which represented 60% of the industrial GDP, were affected by the water crisis, as well as commercial establishments, hospitals and schools (Marengo et al. ). The years between 1999 and 2004 had already been particularly dry, causing the storage levels of Jaguari-Jacareí, the largest reservoir from Cantareira system, to go to À7% (below minimum operational level) in 2004 (Whately & Cunha ). According to Tomer & Schilling (), there are several factors influencing the watershed hydrology, namely, vegetation type, soil properties, geology, terrain, climate, landuse practices, besides spatial patterns of interactions Several studies have been carried out to quantify the relative importance of climate variability and humaninduced change in runoff. Two different approaches have been widely used for that purpose: process-based and statistical methods. In the first approach, physically based hydrological models are calibrated for the period preceding the detected changes in streamflow. To separate the impacts of climate variability and human interference, streamflow is then simulated using meteorological data for the subsequent period, but with fixed model parameters, which embed original LULC conditions (Xu et al. ). Among the statistical methods, climate elasticity is a commonly used measure. It estimates the proportional response of streamflow to changes in precipitation and potential evapotranspiration (Chang et al. ). Often both approaches are compared. There is a perception that climatic changes may be primarily responsible for the recent water supply crisis, and hence, identifying and attributing the source of streamflow reduction is important (Otto et al. ). Given this context, the aims of this study are (1) to investigate whether a systematic change in streamflow is occurring in Cantareira watershed; and, in the affirmative case, (2) to characterize the nature of the changes in reservoir inflows, using climate elasticity and the ABCD hydrological model to separate the relative influence of climate variability and human interference.

STUDY AREA AND DATA
The Cantareira watershed (46 43 Table 1 and marked in Figure 3. Missing values in monthly precipitation data were filled using the Normal Ratio Method (Chow et al. ). The missing precipitation value P x is calculated as: where P i denotes the recorded precipitation for the ith neighboring station, N x is the normal precipitation for the station 'x' with the missing value, N i is the normal precipitation for the ith neighboring station and n is the total number of neighboring stations being considered, n ! 3.
The distance from station 'x' to each of the neighboring stations should be preferably less than 100 km. Prior to filling missing values, the correlation between station 'x' and each of the neighboring stations was tested.
The consistency of precipitation records was checked by plotting the double-mass curve of cumulative monthly precipitation of a target station against the cumulative monthly precipitation in the region. Inconsistent precipitation records     Weron ) was estimated for each AS series to further support the detection of long-range dependence.
In order to detect possible breakpoints in hydro-meteor-  reference period and disturbed period can be expressed as: where ΔQ C refers to variation in streamflow due to climate variability and ΔQ H to the variation due to human-induced changes. The partitioning was estimated through the climate elasticity method and hydrological modeling using the ABCD model.

Climate elasticity method
Schaake () pioneered the concept of climate elasticity of streamflow. It denotes the proportional response of runoff to changes in a climatic factor X (potential evapotranspiration or precipitation, in this case). It is expressed as: The Budyko hypothesis (Budyko ) assumes water balance is governed by the amount of available energy, or the atmospheric demand (represented by potential evapotranspiration -ETo), and water availability (represented by precipitation -P). From the Budyko hypothesis and the long-term water balance equation (Q ¼ P -E), it derives that the ratio of annual evapotranspiration to precipitation (E/P) is a function of φ ¼ ETo/P, dubbed the aridity index, and that precipitation (ε P ) and potential evapotranspiration (ε ETo ) elasticity of streamflow can be assessed through (4) (Arora ): There are different expressions available in the literature for the estimation of F(φ). They are summarized in Table 2 (Zhan et al. ).
The impact of climate variability on streamflow variation is then calculated as: where ΔP (ΔETo) expresses the difference between the average AP (ETo) in the disturbed period and the reference period and P (ET 0 ) is the average AP (ETo) for the whole period analyzed.

ABCD model
Thomas () first introduced the ABCD hydrological model, which takes as input precipitation, potential evapotranspiration and air temperature to simulate streamflow. The model is governed by four parametersa, b, c and deach holding a particular physical interpretation. Therefore, model parameters embed the hydrological response to LULC. For details on the ABCD model, refer to Thomas ().
The model was first calibrated and validated for the reference period using monthly data. The model was calibrated using Excel Solver ® , by maximizing the Nash-Sutcliffe Efficiency (NSE) (Nash & Sutcliffe ), given by where Q obs denotes the observed inflows and Q sim , the simulated inflows. The water balance error (WBE), expressed as was constrained to be within the range (À5%, 5%).
With fixed model parameters, which stand for fixed LULC, streamflow was simulated for the subsequent period. The difference between simulated inflows for the reference and disturbed periods was attributed to climate variability. The difference between observed inflows for the reference and disturbed periods minus the parcel of streamflow variation credited to climate variability was attributed to human-induced change.

Trend and breakpoint analysis
The AS and AP for each of Cantareira sub-basins are plotted against time in Figure 7.
From the graphs in Figure 7, it is noticeable that AS exhibits a more pronounced downward trend than AP in Cachoeira and  Figure 8.
A perceptible break in slope is observed for Cachoeira and Atibainha basins, and the year of the break agrees with the results of step change tests (1989 and 1998, respectively).     1976-1989and 1990-2009for Cachoeira and 1976-1998and 1999-2009 for Atibainha, respectively. Table 6 contains the average AS, AP and ETo for each sub-basin for the reference and disturbed periods, as defined above.

Climate elasticity
The relative contribution of climate variability to the reduction of streamflow is first assessed through the climate elasticity method. Table 7 shows the values for ε P and ε ETo

DISCUSSION Changes in streamflow
It is observed that, since 1998, the interannual variation for both AP and AS in all Cantareira sub-basins was reduced.
This coincides with the onset of a shorter negative phase In agreement with previous results, little oscillation is observed in LULC in Jaguari-Jacareí throughout the years in terms of proportional area. From 1989 to 2010, most pasture/abandoned land/exposed soil conversion to other LULC types was due to reforestation/silviculture and secondary forest cover in medium or initial stage of regeneration (5.33% of Jaguari-Jacareí basin area). On the other hand, pasture/abandoned land/exposed soil advanced primarily over areas of primary/secondary forest cover in the advanced stage of regeneration and reforestation/silviculture (2.99%). Overall, no substantial transition is noticed (see Appendix B -Supplementary information).
Diversely, the area occupied by pasture/abandoned land/exposed soil has been gradually decreasing in Cachoeira basin since 1989, parallel to a steady increase in reforestation/silviculture. It can be inferred that reforestation/silviculture, primarily, and secondary forest cover in medium or initial stage of regeneration is replacing pasture/abandoned land/exposed soil. This is confirmed by LULC transition rates: between 1989 and 2010, the conversion of pasture/abandoned land/exposed soil to reforestation/silviculture and secondary forest cover in medium or initial stage of regeneration sums up to 14.41% of Cachoeira basin area, while the opposite pattern only accounts for 1.84%. Considering the impact of LUCC to be accumulative, the previous analysis offers a plausible explanation for the observed decreasing trend in Cachoeira reservoir inflows since 1989.
In Atibainha basin, a relatively significant increase in pasture/abandoned land/exposed soil area is observed between 1989 and 1999, whereas little change happens in reforestation/silviculture and native forest cover areas.
This modification in pasture/abandoned land/exposed soil area must be regarded with caution, since in 1989, around 25 km 2 of land (7.8% of Atibainha basin area) could not be properly categorized due to the presence of clouds and shadow, a situation that was not repeated for the subsequent years. This fact may translate into the overestimation of the increase in pasture/abandoned land/exposed soil between 1989 and 1999. This supposition is supported by the fact that no clear response of streamflow to a presumable deforestation process is observed for the 1989-1999 interval.
From 1999 to 2003, analogously to Cachoeira, pasture/abandoned land/exposed soil area is gradually shrinking, while a slow expansion in reforestation/silviculture is perceived, and native forest cover oscillates little. A more expressive alteration is observed between 2003 and 2010, also suggesting the replacement of pasture/abandoned land/ exposed soil for reforestation/silviculture, mostly, and secondary forest cover in the medium or initial stage of regeneration. Indeed, the conversion of pasture/abandoned land/exposed soil to reforestation/silviculture and secondary forest cover in the medium or initial stage of regeneration corresponds to 13.54% of Atibainha basin area, whereas the reverse process was 3.36%.  year, it is plausible to argue that the available time window is not sufficiently long to allow a proper appreciation of the events discussed.

SUMMARY AND CONCLUSIONS
In this study, we analyzed AS in each of the sub-basins of Cantareira watershed, namely Jaguari-Jacareí, Cachoeira, Atibainha and Paiva Castro, to identify trends and the timing of structural changes in streamflow and their relationship with climate variability and LUCC.
Only Cachoeira and Atibainha basins showed a significant decreasing trend in AS (p < 0.05) in the period between 1976 to 2009, whereas AP time series did not follow the same pattern in these two sub-basins, nor did ETo. Meanwhile, MAT showed a significant increasing trend (p < 0.10) for two of the three tests performed.
Step change detection was performed for AS time series.
Neither Jaguari-Jacareí nor Paiva Castro basins showed sig- land and pasture areas. Since regional water withdrawals from both surface and groundwater are not significant, we conclude that the parcel of streamflow reduction attributed to human-induced change is chiefly explained by LUCC.
The role of afforestation in determining the water balance of a region has engendered debate over the last few decades, with results indicating increases or decreases in the streamflow. In this setting, the finding that the afforestation activity associated with increasing Eucalyptus plantations over a wide area have led to streamflow reduction is significant. Given that this effect has evolved over a two-decade period of increasing afforestation and is manifest as a change in the rainfall-runoff relationship, this land-use strategy needs to be revisited through a formal hydro-economic analysis to provide better coordination between LULC patterns and water management in Cantareira. The detected trends regarding climate change, LUCC and demand for water reinforce the urgency to reduce the MRSP water supply system on Cantareira and address and integrated water management plan.