A Worrying Future for River Flows in the Brazilian Cerrado Provoked by Land Use and Climate Changes

Abstract

In this study, we assessed the impacts of land use and climate changes on the river flows of 81 watersheds within the Cerrado biome, Brazil, based on a comprehensive analysis of field and secondary data acquired between 1985 and 2018. Complementarily, we projected a future deforestation and climate scenario up to 2050 and predicted their impacts on surface water in the study area. We observed that direct impacts by large-scale deforestation oriented to the production of irrigated agricultural commodities have more significantly impacted river flows than climate changes. We estimated an average decrease of 8.7% and 6.7% in the streamflow due to deforestation and climate changes, respectively. Most of the observed changes (56.7%) were due to land use and land cover changes and occurred in recent decades. Climate and land use and land cover changes combined were responsible for a total surface water reduction of −19,718 m³/s within the Cerrado watersheds. By assuming the current deforestation rates, we predicted a total water reduction of 23,653 m³/s by 2050, equivalent to a decrease of 33.9% of the river flows in the study region. It will cause severe streamflow discontinuity in many rivers and strongly affect agricultural, electric power production, biodiversity, and water supply, especially during dry seasons in that region.

1. Introduction

Areas of extensive agricultural production have been increasingly more vulnerable to the amount of surface water availability due to large-scale deforestation [1], and the effects of climate change [2] are issues demanding a closer look at how much each of those drivers impacts water supplies in tropical regions subjected to high rates of deforestation and climate changes [3].

Different countries have emerged as major exporters of agricultural commodities in recent decades, especially those expanding croplands over large and relatively flat areas with well-defined periods of rain, high suitability to intensive crop production, and agricultural mechanization, such as the case of the Cerrado biome in the central part of Brazil [4,5]. Currently, Brazil is responsible for more than 42% of the soybeans [6] and 20% of the beef consumed in the world [6]. The Cerrado biome plays an important role in the growth of agricultural production in Brazil [7], where 44% (87 M ha) of its territory is occupied by field crops and pastures [8], responsible for 44% of the country’s beef production [9] and for 48% of its soybean exportation [10].

A number of factors are at play: the increasing global demand for agricultural products [11], the rise in internal commodity prices due to the weakening of the Brazilian currency [12], the lack of environmental command and control policies [13], and agricultural occupation policies, such as the one implemented in the MATOPIBA region (acronym for the states of Maranhão, Tocantins, Piauí, and Bahia) have all steadily stimulated the expansion of the agricultural areas in new deforestation frontiers in the Brazilian Amazon and Cerrado regions [14]. The Cerrado itself is the most biodiverse savannah in the world [15]. Unlike the Amazon, prolonged dry seasons and more frequent drought episodes are observed in that region [16,17]. This fluctuation in water availability induces farmers to invest heavily in irrigation systems, leading to the concentration of 80% of the country’s central irrigation systems in that region [18].

Different studies reported changes in river flows (both superficial and groundwater) in the Cerrado biome [19,20,21,22], reporting different variables causing streamflow reduction. The increase in deforestation, irrigated agricultural production [18,21,23], climatic factors such as the decrease in rainfall, and the increase in evapotranspiration [24] are some of the factors discussed. However, detailed hydrological surveys have been restricted to small study sites and time periods, raising questions as to the role of drivers and their implications for future scenarios in surface water changes.

The decrease in river flows in the Cerrado biome has been the center of conflicts over access to water for different uses [25,26], such as water for human consumption, natural ecosystem maintenance, agricultural use, and energy production, especially during years of severe drought episodes. Water scarcity is a growing problem in Brazil due to most of the water reservoirs and hydropower dams within the Brazilian territory being spatially located in the Cerrado biome [27,28]. In addition, this biome is affected by leakages of deforestation effects from neighboring biomes, such as the Brazilian Amazon [29,30] and the Pantanal biomes [31].

River flows are generated by a complex mechanism involving water–energy interactions between the land surface and the atmosphere. Those flows depend on the maintenance of the hydrological cycle, including ecosystem services such as the vegetation function of reducing soil erosion and runoff during rainy seasons, water infiltration in deep and porous soils regulated by the long roots of the Cerrado vegetation, groundwater supply of the base flows during dry seasons [24], and the moisture availability in the atmosphere provided by the processes of evapotranspiration [19].

The hydrological cycles include atmospheric and terrestrial components, which makes it difficult to distinguish the consequences of river flow changes that could be attributed to climatic phenomena, large-scale deforestation [3], and forest degradation. For example, the deforestation effects on rainfall depend on the actual local evapotranspiration [19]. However, direct effects on river flows can be estimated along with their direct causes and drivers [3]. In this analysis, we estimated the effects of changes in potential evapotranspiration, precipitation, and local land uses on streamflow based on a time series of field data for each individual watershed in the study area.

From the climatic point of view, changes in rainfall or potential evapotranspiration can directly affect the river flows [32]. Likewise, the increase in deforestation tends to reduce soil water infiltration [20], evapotranspiration [19], and, consequently, reduce river flow regulation based on groundwater [33] during dry periods. These are crucial elements in seasonal environments such as in the Cerrado biome. Additionally, deforestation can increase water consumption by agricultural irrigation systems, because it demands higher amounts of water during the period of lowest surface water availability and causes even more water scarcity [34]. However, assessing the sustainability of the coupled human–nature system requires the distinction of the potential drivers of surface water changes, which can enable the improvement in water resource utilization [35]. In this analysis, we assessed the impacts of climate and land use changes in the flow of rivers within the Cerrado biome. We analyzed field datasets of 81 watersheds acquired in the last 34 years and developed future scenarios of climate and deforestation changes up to the year 2050. We analyzed their effects, separately and jointly, on the streamflow of the entire study region. Our study results call attention to potential water flow discontinuity soon, which may directly affect agricultural production, electric power production, and water supply in the Cerrado region.

2. Materials and Methods

Our methodological approach consisted of three types of analyses: (I) Time Series Analyses, (II) an Estimate of Contributing Effects, and (III) Future Scenario Analyses of the river flows of 81 watersheds within the Cerrado biome, as shown in Figure 1.

2.1. Time Series Analyses

In the Time Series Analysis, we assessed the contribution from land use changes (Lc), which refer mostly to deforestation to create pastures and agricultural lands, and climate changes (Cc) from historical data series. In addition, we projected estimates of deforestation effects on river flows based on future scenarios of deforestation between now and the year 2050.

Fluviometric and climatic data were collected from stations showing data gap rate of <12% and rainfall data gap rate of less than 7%. In the case of overlapping drainage areas, we selected the smallest one to better detect the direct effects. This process resulted in 81 selected watersheds, covering 30,042,423 hectares, representing 15.1% of the Cerrado biome.

Kalman Filter [36] estimates were used to fill fluviometric gaps, a non-linear procedure widely applied to streamflows [37], as is the case for streamflow. The average precipitation and potential evapotranspiration for each watershed were estimated using the inverse distance weighting (IDW) [38] technique and the data collected by the rainfall stations from the Brazilian national water agency (ANA). The estimates for evapotranspiration, calculated by Penman–Monteith equation, were retrieved from the database of Xavier, A.C., King, C.W., and Scanlon, B.R. (1980–2013) [39].

Climatic variables (rainfall and estimated evapotranspiration) were isolated from land use and cover changes to distinguish from the contribution and effects of Lc and Cc using lasso regression [40]. We then applied the Pettitt test [41] to identify the break pattern in these variables. This allowed us to distinguish two periods: pre-disruption (P1) and post-disruption (P2).

The historical land use data [8] were resampled to 500 m × 500 m grids, clipped by individual watershed boundaries, and grouped into the following classes: forest, savanna, countryside, agriculture, pasture, agricultural mosaic, forestry, urban areas, mining, and water.

Once corrected, the historical data series were analyzed by season and year to check for series stability [42,43], trends [44], break patterns [41,45], and the runoff coefficient (RC), represented by the rate between mean streamflow and precipitation [46].

2.2. Estimate of Contribution and Effects

By assigning an observed break point for each watershed, we assessed the contribution and effects of each driver on each river’s streamflow. The Estimates of Contributor Effects analyses looked at the contribution of land use changes (Lc), climate changes (Cc), and future land change scenarios. Above, we mentioned the climatic variables contributing effects were separated into pre-disruption (P1) and post-pre-disruption (P2) periods. However, assuming the land use and land cover data were kept unchanged, the water flow disruptions in the P1 scenario would be evidenced. The P’ scenario reflects the isolated effect of the climatic variables (Figure 2).

Based on the assigned break points for P1, P2, and P2′, we estimated the contribution and effects of climate variables (Cc), land use, and land cover (Lc) on changes in streamflows in the study area. We estimated the contribution impact (%) by calculating the change a driver had on the observed pre-disturbance river flows (P1). Thus, Q is the effect in m³/s a driver may impact the river flows. The relative change ($\mathsf{\Delta }$Q) is the average difference of river flow pre- and post-disturbance, estimated using the following equations:

$$\mathsf{\Delta } total=\overline{x}\left(p2\right)-\overline{x}\left(p1\right),$$

(1)

$$\mathsf{\Delta }QLc=\overline{x}\left(p2\right)-\overline{x}\left(p{2}^{\prime }\right),$$

(2)

$$\mathsf{\Delta }QCc=\overline{x}\left(p{2}^{\prime }\right)-\overline{x}\left( p1\right),$$

(3)

$$CLc=\left(\mathsf{\Delta }Lc/\left(\mathsf{\Delta }Lc+\mathsf{\Delta }QCc\right)\right)\times 100,$$

(4)

$$CCc\%=\left(\mathsf{\Delta }QCc/\left(\mathsf{\Delta }QLc+\mathsf{\Delta }QCc\right)\right)\times 100 \mathrm{or} 1-CLc\%,$$

(5)

Total relative change (1); effect of changes in land use (Lc) on river flow (m³/s) (2); effects of climate changes (Cc) on river flow (m³/s) (3); contributions of land use changes (Lc) to streamflow (%) (4); contributions of land use changes (Lc) to streamflow (%) (5), where ΔQ is the average variation of streamflow.

We applied a model approach to estimate the river flows under the contributions of Lc and Cc to the P2 post-disruption period and P2′ scenarios. Initially, we applied an empirical hydrological model (type rain–river flow), which explicitly represented the water balance components, mathematically modeled by applying Lasso’s regression [40]. This regression model was chosen based on its simplicity, flexibility, and suitability to integrate data and models. Lasso´s regression model fitting was carried out by decreasing variance, increasing bias, and adding penalties on the absolute values of coefficients, thus reducing the model’s complexity through the elimination of some variables. It simplifies and adjusts the model to those variables that can statistically explain at a significant level the dependent variable, as is the case of the streamflow.

The penalties are controlled by a hyperparameter ($\lambda$), selected by K-Fold Cross-Validation. This parameter speeds up the individualization of the model, prevents excessively weighing the variables, and avoids overfitting. The self and individualized parametrization, based on individual watersheds, enabled us to apply the flexible regression models, optimizing the group analysis. After normalizing the data, the coefficient of each dependent variable retains some intelligibility, suggesting that each variable contributes to the solution that describes the behavior of the dependent variable (streamflow). The equation that describes Lasso´s regression is presented as follows [47].

$${{\displaystyle \sum }}_{i=1}^n{\left(y_i-{\beta }_0-{{\displaystyle \sum }}_{j=1}^p{\beta }_j{x}_{ij}\right)}^2+\lambda {{\displaystyle \sum }}_{j=1}^p\left|{\beta }_j\right|,$$

(6)

Lasso´s regression equation (least absolute shrinkage and selection operator) (6) and the retraction operator and absolute minimum selection. It shows the sum of squares of the errors added to the sum of the absolute values of the dependent variable parameters, multiplied by the lick hyperparameter.

To generate a function capable of predicting the river flows, this model was based on the definition of weights for each of its coefficients: the climatic variables (rainfall and monthly potential evapotranspiration), months of the year, and land use classes (forest, savanna, grassland, pastures, agriculture, silviculture, agricultural mosaic, urban areas, mining, and bodies of water). In our methodological approach, we were able to integrate the historical time-series dataset, the future climate model, and the spatially explicit model of LULC change.

This approach was applied to each watershed, aiming to determine the best lambda from the K-Fold Cross-Validation with the cv.glmnet function, nfolds = 10. We randomly partitioned the sample (P1) into 10 parts, using 9 parts for training and 1 for validation for all possible combinations, which allowed us to select the most accurate lambda value. Subsequently, we generated 100 models, randomly sampling 25 years at a time, to generate the predictive intervals for each basin. We adopted the mean regression of those models and validated them by applying objective functions [48,49] to compare the observed river flow for period P2 with the river flow modeled for that period.

$$F_1=\frac{{{\displaystyle \sum }}_{i=1}^n{\left[ln\left(Q_{sim,i}\right)-ln\left(Q_{Obs,i}\right)\right]}^2}{{{\displaystyle \sum }}_{i=1}^n{\left[ln\left(Q_{Obs,i}\right)-ln\left(\underset{¯}{Q_{Obs}}\right)\right]}^2},$$

(7)

$$F_2=\frac{{{\displaystyle \sum }}_{i=1}^n{\left[Q_{sim,i}-Q_{Obs,i}\right]}^2}{{{\displaystyle \sum }}_{i=1}^n{\left[Q_{Obs,i}-\underset{¯}{Q_{Obs}}\right]}^2},$$

(8)

$$F_3=1-\frac{{{\displaystyle \sum }}_{i=1}^n\left(Q_{sim,i}-\underset{¯}{Q_{simi}}\right)\left(Q_{obs,i}-\underset{¯}{Q_{Obs}}\right)}{\sqrt{{{\displaystyle \sum }}_{i=1}^n{\left(Q_{sim,,i}-\underset{¯}{Q_{sim}}\right)}^2{{\displaystyle \sum }}_{i=1}^n{\left(Q_{obs,,i}-\underset{¯}{Q_{obs}}\right)}^2}}$$

(9)

$$F_4=\left|ln \left(\frac{{{\displaystyle \sum }}_{i=1}^n{Q}_{sim,i}{}^{}}{{{\displaystyle \sum }}_{i=1}^n{Q}_{obs,i}}\right)\right|,$$

(10)

F1 (7) is a logarithmic function and the error weights at low flows rather than the errors at high flows; F2 (8) emphasizes errors at high flows; F3 (9) focuses on time changes between flow rates; F4 (10) evaluates the overall difference between the simulated and observed flow during the calibration period.

2.3. Future Land Use and Climate Change Scenarios

Future estimates of LULC change were carried out using artificial neural network (ANN), available on the Land Change Modeler [50]. The estimates were applied to the land use and land cover dataset from 2015 to 2017 [8] and combined with land use spatial variables. These variables were selected by applying V’Cramer technique for each LC transition when greater than 0.16 or when directly related to the change in question [51]. The neural network uses spatial variables as criteria to search for the most significant land use class transition relationships by learning and replicating the identified major trends (Figure 3).

A large set of land use-related variables was tested with different transformations (e.g., natural log, square root of the distance, and evidence likelihood). The most adherent ones, showing highest explanatory power (V’Cramer), were selected. We also included constraints (weights) to different levels of land use change, such as Permanent preservation areas (APP) [52], legal reserves (RL) [53], integral protection conservation units (UCPI), sustainable use conservation units (UCUS) [54], quilombo areas and indigenous lands. A sample size sufficiently large with respect to the number of pixels in the transition from classes t1 to t2 was necessary so the ANN analysis could identify the main land use change patterns. A large enough sample size could not be found for the minority of classes such as forestry, mining, water, urban, agricultural mosaic, and others since it represented a total area smaller than 2.5% of the entire Cerrado biome. Those class areas were therefore held static, and were not evaluated by ANN.

A set of artificial neurons (neurons) interconnected by a multi-layer perceptron (MLPNN) acquired information about a possible transition of land use pattern, which was then incorporated/processed and later made available to the rest of the neuron network [55]. As the input data fed the nodes with information (signal), the information advanced through the network, accumulating and propagating information from other nodes through the layers until they reached a certain activation state change at the output node layer. A final solution was found by the ANN when the network passed through the following phases: training, incorporation of learning, correction of errors and weights (delta rule), and the calculation of the mean square error (root mean square—RMS).

Initially, we used pixel training samples representing land use change. This procedure helped us identify the relationship between the selected variables and the activation levels of the network’s output nodes. At the end of the process, the pixels with the highest level of activation are represented as those where land changes should be assigned.

Since there is an accumulation of learning during the training phase, the network uses the technique of Reverse Propagation (back propagation), which consists of a forward and a backward step. Changes are made to the state of the nodes in each step, according to the weights assigned by the system. Each sample is fed by the input layer and the weighted signals of all nodes to the receiving node and then connected to the previous layer [52].

Through training, the neural network is able to generate several values (link weights) for the relationships between the variables, which are incorporated into the model (learning), helping it to simulate the process of land use changes [51].

The information learned from the training data was transmitted through the network before propagating errors and weights. The error is then propagated backward, using weights for connections, and corrected based on a relationship known as delta rule.

Repeated back-and-forth iterations were applied until the network had learned the characteristics related to the correct solution for all classes. At the end of the training phase, the network obtains the appropriate weights for the connections among the nodes in the multiple layers, bringing greater accuracy to the output data.

The output data generated represented the change in state according to the nodes with the highest activation level. It is not expected for the output data to be the same as the observed data. It is very likely there were errors in the network causing the difference between the output and the observed data [51,56]. The acceptable error rate is calculated according to the error associated with the network learning, calculated by the errors’ root mean square (RMS).

The land use changes predicted by the ANN (artificial neural network) were distributed using the current trend of deforestation rate. Based on it, we estimated that an average of 6.3 thousand km²/year will be deforested by 2030, an average of 3.4 thousand km²/year between 2030 and 2040, and finally, around 2.5 thousand km²/year between 2040 and 2050.

The land use projection model validation process [57] consisted of verifying the model’s agreement with observed land use for 2018. Subsequently, we observed the agreement between the model prediction and the observed data for that year and from the previous period (2015), which we called the null model [51]. This verification used several indices, including the random agreement—N(n), quantity agreement—N(m), standard agreement—M(m), and adjusted location agreement—P(m), in addition to the standard Kappa and the Kappa for no information (Kno).

The indicators N(m), M(m), and P(m) provide information about the correctness between the reference map (observed data) and the modeled map regarding the location of pixels of each category and the corresponding values by category [57,58].

N(m) refers to the agreement between observed and modeled maps, where pixels were randomly located in the reference map. M(m) indicates the correct portion between the reference and modeled maps, without any change, neither in quantity nor in pixel location. P(m) is the agreement between the reference map and a modified comparison map, where the modification is to rearrange the cell locations within the map as closely as possible to maximize the agreement between the modified modeled map and the reference map [51].

The Kappa index (K) is a measure of agreement between two datasets and measures the degree of agreement beyond what would be expected by chance alone. The K index measure varies from 1 to 0, where 1 represents a total agreement between the datasets, 0 (zero) indicates an expected agreement by chance (random), and below 0 indicates an agreement lesser than expected by chance.

Transition models were generated and validated among the land use and land cover classes, capable of supporting a construction scenario for land use changes during the study period and using above-mentioned validation metrics [57]. This allowed us to project future scenarios of land use changes by applying current deforestation rates up to 2030 and systematically decreasing them between 2030 and 2050 (final year of the projection). The projection fitting was based on two trend scenarios: “Conservation” and “Business as Usual”.

We used the dataset provided by the statistically downscaled earth system models CCSM4 from National Aeronautics and Space Administration’s (NASA) Earth Exchange Global Daily Downscaled Projections (NEX-GDDP) [59] to project the rainfall and potential evapotranspiration. These climate projections showed the best adherence to regional climate patterns of the RCP 4.5 [60,61] scenario, which is considered a medium emission intensity scenario, assuming global stability of carbon emissions. The river flow estimates for the Cerrado biome were modeled and compared for the future P2 post-disruption period and for the P2′ scenarios by using a data series of land use and land cover, potential evapotranspiration, and complete rainfall up to year 2050.

3. Results

The temporal analysis of the river flows in 81 watersheds studied indicated significant (p value < 0.5) rainfall reduction in two watersheds (2.4%); another 46 watersheds showed a non-significant decrease (56.7%); and 28 watersheds (35%) showed a significant increase in rainfall.

During the dry period, 12 watersheds (14.8%) showed significant rainfall reduction. The potential evapotranspiration showed a significant ascending trend for 77 watersheds (95%) and over 75 watersheds (92.6%) during dry periods. We observed a significant average reduction in water flow over 77 watersheds (95%) during the study period and a significant descending trend of water flow over 74 watersheds (91.3%) during dry periods. Only three watersheds (3.7%) showed an ascending trend of water flow. Concerning the average runoff coefficient for watersheds between 1985 and 2018, 57 watersheds (70%) showed a reduction ranging from 3.3 to 2.7 (Figure 4).

The water flow estimate models indicated average accuracies of 99% for the global accuracy (F1), 70% for the dry season (F2), and 70% for the rainy season (F3) and the estimated average adherence to the delay was 84% (Figure 5).

The explanatory power analysis of Cerrado’s model estimates for future land use highlighted the distance from previous deforestation, the travel cost matrix, slope, and soil type, among other variables related to the main land use transitions (

Table 1. Variables related to land use changes and their respective explanatory powers (V’Cramer), showing values close to 1, highly related to land use changes, and values below 0.16, a relationship close to null.

Variables	General	Forest	Pasture	Savanna	Agriculture	Grassland	Silviculture	Water	Others	Urban	Mosaic	Mining
Deforestation distance (natural log)	0.38	0.89	0.72	0.53	0.48	0.47	0.17	0.1	0.06	0.08	0.07	0.007
Friction (travel cost)	0.26	0.63	0.3	0.37	0.35	0.34	0.14	0.36	0.06	0.01	0.03	0.001
Slope	0.28	0.87	0.49	0.41	0.37	0.33	0.12	0.1	0.06	0.05	0.04	0.006
Altimetry	0.28	0.87	0.49	0.42	0.36	0.33	0.14	0.07	0.06	0.05	0.04	0.006
Rainfall of the rainiest month	0.28	0.87	0.5	0.43	0.31	0.34	0.14	0.07	0.06	0.05	0.06	0.007
Rainfall of the driest month	0.24	0.67	0.51	0.25	0.3	0.18	0.16	0.05	0.04	0.05	0.11	0
Soils	0.27	0.87	0.48	0.48	0.3	0.33	0.11	0.06	0.06	0.05	0.04	0.005
Soils (evidence likelihood)	0.27	0.77	0.46	0.42	0.34	0.35	0.23	0.09	0.08	0.06	0.05	0.007
Protected areas (natural log)	0.29	0.87	0.51	0.41	0.31	0.4	0.11	0.07	0.08	0.06	0.06	0.005
Legal reserves (natural Log)	0.26	0.8	0.48	0.33	0.31	0.29	0.11	0.08	0.06	0.07	0.04	0.006
Permanent preservation area (square root distance)	0.28	0.85	0.49	0.4	0.31	0.34	0.11	0.06	0.06	0.05	0.07	0.006
Valley (square root distance)	0.11	0.59	0.32	0.27	0.32	0.19	0.1	0.1	0.04	0.04	0.02	0.003
Hydrography (distance)	0.05	0.01	0.003	0.0004	0.006	0.0019	0.001	0.07	0.14	0.02	0.0004
Limite constrange	29	0.87	0.51	0.41	0.37	0.32	0.12	0.07	0.06	0.05	0.001	0.07

). The accuracy of the land transitions was satisfactory, ranging from 61.01% to 98.04% (

Table 2. On the left: the accuracy of the land use transitions. On the right: the model accuracy as a whole, namely, random agreement—N(n), quantity agreement—N(m), standard agreement—M(m), and adjusted location agreement—P (m), in addition to standard Kappa and Kappa without information (Kno).

Accuracy of Land Use Transitions				Model Accuracy
Transition	Accuracy	Transition	Accuracy	Overall Model Accuracy Metrics	Model	Null Model
Forest to Pasture	71.82%	Grassland to Mining	77.78%
Forest to agriculture	71.82%	Others to savanna	67.51%	P(N)	0.30	0.30
Savanna to pasture	75.20%	Others to grassland	68.00%	P(M)	0.99	0.99
Savanna to silviculture	83.25%	Others to pasture	65.60%	P(P)	1.00	1.00
Savanna to mosaic	88.98%	Others to agriculture	76.38%	M(n)	0.30	0.28
Savanna to urban	89.62%	Others to mosaic	100%	M(m)	0.99	0.96
Savanna to agriculture	73.65%	Pasture to grassland	63.50%	M(p)	0.99	0.96
Grassland to pasture	78.50%	Pasture to agriculture	57.54%	N(n)	0.09	0.09
Grassland to agriculture	80.04%	Pasture to mosaic	77.00%	N(m)	0.63	0.63
Grassland to silviculture	86.69%	Forest to silviculture	98.04%	N(p)	0.63	0.63
Grassland to mosaic	88.89%	Forest to mosaic	98.44%	Kno	0.99	0.96
Grassland to urban	92.45%	Savanna to forest	61.01%	Klocation	0.99	0.90

), and the general accuracy indicated that our model was appropriate to carry out the projections for future scenarios (Table 2).

Our land use model allowed us to estimate deforestation rates of approximately 6400 km²/year by 2030, 3400 km²/year by 2040, and 2500 km²/year by 2050. The suggested decreasing trend of projected deforestation rates suggests new croplands could become scarcer in the study area. In this scenario, between 2018 and 2050, the projected native Cerrado vegetation could be reduced from 54.3% to 47.5%, replacing pastures for agricultural lands (Figure 6, Figure 7 and Figure 8 and

Table 3. Land use and land cover changes observed by 2015, 2018, and 2020 and predicted by 2030, 2040, and 2050. Land use, land cover classes, and total projected deforestation and remnants of native vegetation in scenario Business as Usual.

Land Use	2015	2018	2020	2030	2040	2050
Forest	18.40%	18.70%	18.50%	17.70%	17.20%	16.90%
Savanna	22.00%	21.50%	21.30%	20.10%	19.40%	18.90%
Grassland	14.40%	14.10%	13.90%	12.80%	12.20%	11.70%
Others	0.50%	0.50%	0.50%	0.40%	0.40%	0.30%
Pastures	30.00%	30.00%	29.80%	29.40%	29.20%	29.10%
Agriculture	11.70%	12.20%	13.00%	16.60%	18.60%	20.00%
Silviculture	1.70%	1.70%	1.70%	1.70%	1.70%	1.70%
Agricultural mosaic	0.30%	0.30%	0.30%	0.30%	0.30%	0.30%
Urban areas	0.30%	0.40%	0.40%	0.40%	0.40%	0.40%
Mining	0.00%	0.00%	0.00%	0.00%	0.00%	0.00%
Water areas	0.70%	0.60%	0.60%	0.60%	0.60%	0.60%
Deforestation(km2)	10.635	11.747	63.952	34.108	25.581
Deforestation(km2/year)	3.545	5.874	6.395	3.411	2.558
Native remaining	54.80%	54.30%	53.70%	50.50%	48.80%	47.50%

).

Our assessment of the contribution and effects of climate drivers (Cc) and Land changes (Lc) between 1985 and 2022 indicates that 71 (87.7%) of the total analyzed watersheds showed a decreasing trend of water flows due to land use and land cover changes, and 69 (85.2%) watersheds showed a decreasing trend of streamflow due to climate effect (Figure 9). On average, the climate change contribution (mainly as potential evapotranspiration) affecting streamflow was −43.3%, while the effects of land use change were −56.7%. Comparing the river flows observed during the pre-disturbance period (P1), the average estimated total effect of land use and land cover changes on river flow was −8.7% and −6.7%, which were attributed to climate changes. The total average reduction in the estimated basin water flows was −19,718 m³/s, of which −11,190 m m³/s was related to the effects of land use and land cover changes and −8528 m³/s was related to the climate change effects.

Based on our predicted future scenarios, we estimated that 75 (92.6%) watersheds would decrease their water flows compared to the pre-disturbance period, where the water flow of 69 (85.1%) watersheds would be decreased due to the effects of the expansion of agricultural lands and the water flow of 71 (87.6%) watersheds would be decreased due to the effects of climate changes (Figure 10).

We estimated an average water flow change of −65.8% attributed to the effects of land use changes and −34.2% attributed to the effects of climate changes. Overall, those drivers tend to affect future water flows of the watersheds in the Cerrado biome generating an average decrease of 33.9%, where 23.4% would be attributed to land use expansion, especially agriculture, and 12.8% would be attributed to climate changes (Figure 11). We predicted a decrease of 23,653 m³/s for analyzed watersheds between 2022 and 2050 compared to the water flows observed during the pre-disturbance period.

Our predicted future scenario indicates decreasing rates of streamflow in most of the watersheds in the study area resulting from the combined effects of climate and land use changes (Figure 9). In other words, based on our future estimates of streamflow, we can predict that if agricultural expansion stagnated in the pre-disturbance period, none of the watersheds would show significant streamflow changes in the future after 2016. Based on currently observed trends of land use, in the future scenario of land use changes (P2), more than 90% of the watersheds showed a significant decrease in streamflow (Figure 10). We observed a few watersheds that showed significant increases in streamflow in the study region. Most of these watersheds were spatially related to fewer land use changes as predicted by the future scenarios.

By comparing observed streamflow observed during pre-disturbance (P1) and post-disturbance (P2) periods under these two different scenarios, (a) zero deforestation in future scenario (P2′) and post-disturbance (P2) and (b) a future scenario of continuing deforestation (P2), we predicted a worse outcome for scenario (b), with the greatest water scarcity due to keeping the current pace of deforestation rates (Figure 12).

We had some emblematic cases illustrating specific contexts. The case where the difference between the P2 and P2′ scenarios is smaller is due to a lesser intense occupation trend (example, Quebra Anzol River). Or the case of areas with combined monoculture and deforestation rates can be attributed to the western state of Bahia, which has been intensely occupied by large-scale monoculture agriculture, mostly soybeans. This area aggregates the municipalities with the highest deforestation rates in the Cerrado biome, and a significant streamflow loss (−24.2%) was attributed to the 54.6% expansion of agricultural lands.

An example of streamflow loss is the Arrojado river, showing intermediate streamflow capacity, which has lost 18.2% of its average streamflow, equivalent to 328.5 m³/s. About 56% of this loss was due to agricultural occupation and 44% due to climate changes. Future trends indicate an accumulated loss of 36.2% of its flow (651.5 m³/s) by 2050. Another example is the Ondas River, which has lost 25% of its average streamflow, 391.5 m³/s, mostly due to the expansion of agricultural lands. For this river, we predicted an accumulated streamflow loss of 56% (864.32 m³/s) by 2050. Lastly, in the Corda River, spatially located in the Northern e Cerrado biome in the state of Maranhão, we observed an average streamflow of 778 m³/s during the initial period of our analysis (P1), decreasing to 570 m³/s in P2 (−27% of the streamflow). The streamflow could have increased to 716 m³/s if no deforestation had occurred in that watershed. Future estimates indicate if the average flow reaches 266 m³/s (−66% of the streamflow), the river will become intermittent, drying out during the dry season by 2047 (Figure 13).

4. Discussion

Our study results of 81 watersheds indicate macro trends of decreasing streamflow in the Cerrado biome due to agricultural expansion and climate changes. From the climatic point of view, the historical series observed in the Cerrado region showed a clear pattern of increase in the potential evapotranspiration associated with a rise of temperature and radiation in the region presented in the results above and in other studies [61,62]. The climate variables studied have become more intensive during dry seasons when the rainfall substantially decreases. On the other hand, annual rainfall showed variable behavior, mostly showing no significant negative trends, such as the studies conducted by [23]. Therefore, we cannot attribute changes in annual precipitation as the leading factor in decreasing streamflow in the Cerrado biome, as previously mentioned in our analysis of driver contributions and effects.

Current scientific knowledge related to the effects of replacing native vegetation with other land use types (e.g., agriculture and pasture) has been controversial [63,64,65,66,67,68]. Nevertheless, most of the current scientific literature indicates that deforestation increases streamflow [65], mainly maximum and mean flows [63,66]. Those studies’ results were affected by several factors such as climate, causes of surface runoff, physiographic characteristics of watersheds, vegetation characteristics (i.e., age, species, and regrowth), landscape characteristics, and their environmental interactions, soil conservation practices, and the type of change in soil conditions [65]. However, the Cerrado biome shows particular physical and biological characteristics that make any assessment of the impact of eventual land use changes very complex and difficult to compare with previous results reported by other scientific studies and analyses carried out in other regions of the planet.

We developed a promising methodological approach to assess the effects of land use, land cover, and climate change by innovating and adjusting available technical approaches. Based on [3,69], we estimated and distinguished the effects of climate change and land use change in several river flows within the Cerrado biome by adjusting those available approaches to the Lasso regression and by combining future climate, land use, and land cover modeling. Our approach substantially contributed to integrating previously developed modeling approaches into more integrated models that enabled us to better understand natural systems. We observed that the period analysis may cause model limitations to capture macro climatic cycles, masking out eventual intra-cycle trends, such as in the following cases: (1) by considering that parts of the Cerrado have been occupied prior to 1985, the streamflow during the pre-disturbance period would have been affected by previous expansion of agricultural or pasture lands expansion, implying an underestimate in land use changes’ effect; (2) the coefficient values of the variables in the lasso modeling may lose part of their intelligibility due to the Cerrado’s climatic seasonality; (3) there are different delays between the interaction of climate variables and land use changes, which may or may not be properly considered for each analyzed basin, reflecting in a loss of the regression accuracy; (4) land use changes observed in other regions may have an impact on our studied watersheds, such as deforestation in the Amazon, by reducing the humidity that reaches the Cerrado region through the flying river phenomena; (5) our model of land use changes assumes that the main arable lands have already been occupied and, therefore, deforestation rates should be lower over the next decades; and (6) our methodological approach has enough elements to indicate the role of each driver in reducing the streamflow of Cerrado rivers, although we have observed model constraints distinguishing some dynamics between climate and land use changes.

Our methodological approach considered the systemic complexity by assuming that each assessed watershed shows some physical and biological characteristics that make it unique. We carried out individual analyses, taking advantage of the available long time-series data available for each watershed, and applied a methodological framework that allowed the integration of the individual results of the watersheds. We applied the necessary adjustments and inferences from the regression model of each of the 81 hydrographically analyzed watersheds. As a result, the fit quality (Figure 5 and Figure 13) of the parameters representing the variability of streamflow indicates a strong adequacy, especially when using the time-series data showing those more pronounced trends. Based on it, we discussed the contribution or “effects” of individual variables used in our analysis.

The statistical regression approach allowed us to distinguish the effects of land use and climate change on the streamflow in the study area. We applied the regression model analyses to assess the relationships among variables using the long time-series data acquired in the field. These models produced values from a sum of contributing variables, where the predicted streamflow was the net sum of all variable contributions. The results of the regression parameters indicated that although the climate is an important factor that affects streamflow, it is no longer the most important in explaining the recently observed decreasing trends of streamflow in the study region. For example, practically all regression models (>95% of the 81 analyzed watersheds) showed at least one significant negative parameter associated with variables related to land use. A total of 70% of the analyzed models showed that the agricultural use parameter was negative, and in more than 50% of those cases, it was a dominant predictor of that decreasing streamflow trend. In addition, pasture and forestry showed a relevant percentage (>50%) of regression models showing a negative parameter.

Although our results are not in agreement with some previously reported results by (19, 20, 22, 33, 97) for the Cerrado region, we argue that our analysis differs from previous work due to the applied space and time scales, the uncertainties associated with data collection, and the quantification of the vegetation changes, which are associated with our applied methodological approach that was able to distinguish effects of climate and LULC change. We also observed that the time-lapse data of the streamflow varied in previous studies. A short streamflow record may fail and only partly capture long-term cumulative hydrological responses following changes in native vegetation cover [70]. For example, in [71], the study area encompassed only a small portion of the Cerrado, and the time-series data of streamflow were analyzed up to 2012, which was a relevant period impact of climate and LULC change in this biome. In the most recent research conducted by [72], which included the entire country of Brazil, their results agree with our findings in the Cerrado biome. [72] used the same empirical statistical context by applying regression model analyses and observed trends toward drier climate conditions and water scarcity in the Cerrado, mainly driven by changes in the climate component. In several regions of Brazil, they observed that the expansion of intensive agricultural lands was the most significant component of streamflow changes, especially in the Cerrado biome.

We call attention to land use changes caused by the expansion of agricultural lands and pastures, which is the main cause of greenhouse gas emissions in Brazil (46%). Agricultural activity alone directly emits a total of 27% of CO₂ emissions for the whole country [73]. Therefore, some effects of climate change on surface water availability are derived or originated from agricultural activities. Although there are some relationships between climate and land use changes that cannot be isolated, it did not prevent us from estimating the impacts of deforestation on streamflow.

Ecosystems such as the Cerrado biome, which show a well-defined [16] and high water baseflow contribution to the streamflow during the dry season, deforestation, and land cover conversion from native vegetation to soybean plantations or other types of monoculture, affect a wide chain of hydrological components at medium and long-term time lapses. Those hydrological effects include the reduction in real evapotranspiration [19], reduction in water infiltration capacity and aquifer recharge [28], the increase in surface flow during the rainy period, and the decrease in stream flow during dry periods in the medium and long term. Together they may decrease the ability of aquifer recharge during the rainy season and the ability to sustain high water consumption for crop irrigation during dry seasons [74]. Therefore, it is very likely that the observed increased streamflow in the short-term just after deforestation and at the beginning of agricultural production [20,22] can be considered responses to the reduction in the observed evapotranspiration. In those cases, the streamflow will tend to decrease at medium and long-term time lapses. It is important to note that the trend of increasing potential evapotranspiration in the Cerrado biome reported by [39,62] is related to the climatic parameters only. As for, the studies by [19,20] observed a decrease in real evapotranspiration in environments occupied by agricultural systems.

Consequently, the trends are of equal concern for biodiversity conservation not only for the Cerrado aquatic biota [75], which relies on both tree survival [76] but also on the availability of small streams and lentic habitats in that landscape for survival. The Cerrado borders several other South American biomes, contributing to large continental water basins. The water shortage forecasted by our analysis will likely affect the resilience of those biomes to fire and climate changes. The Pantanal biome, for example, largely depends on the water flow from the Cerrado rivers for its hydroperiod dynamics, and has already experienced unexpected fires that deeply threaten its biota [77].

The extensive and continued expansion of agricultural lands in the Cerrado leads to the conclusion that the production of agricultural commodities and the scarcity of water are intrinsically linked. The increasing export of intensive water-consuming commodities has changed water governance, shifting the water control from local, regional, and national actors to those who dominate global agricultural production chains [78]. Some studies in Latin America have observed that political alliances between the private sector and the state have driven environmental policy reforms that have enabled the concentration of water rights to agro-industrial companies, heavily affecting the social equity and sustainability of water use [78,79,80]. Approximately 80% of central pivots in Brazil are spatially located in the Matopiba region, western Bahia state, where the environmental policy reform after 2012 allowed industrial agriculture to access water supplies in a region showing increasing water scarcity and social conflicts related to water access [81]. Some studies have enhanced the concept of “water justice” combined with demands for an equitable distribution of water access rights and political water decision-making rights [82].

The role that producers, exporters, and consumers of agricultural commodities have is of great concern, as well as the role of definition and enforcement of public policies of command and control in the mitigation of and adaptation to ongoing and future climate changes. In this context, Brazil plays a major role in global agribusiness exportations, reaching more than 120 billion dollars a year exporting agricultural commodities to China, the European Union, and the USA [83].

However, the expansion of agricultural lands has critically affected surface water availability in the Cerrado biome, which may heavily impact the international economy and food security, especially under the recent scenario of weakening environmental policies in Brazil [52,84,85]. The importation policies of several countries that depend on Brazilian agricultural commodities should be appropriately reviewed to prevent the importation of any agricultural commodity related to new deforestation in the Cerrado biome. In this case, environmental impacts, such as increasing water scarcity, carbon emissions, biodiversity lost, etc., should be commercially shared to those importers and informed to consumers.

The future trend scenario includes a worsening of surface water availability in almost the entire Cerrado biome. It will require rigorous evaluation and monitoring of farming production, which is responsible for 78.3% of surface water consumption in Brazil [86] and 96.6% of deforestation in the Cerrado biome [8]. In tune with other studies we have shown, the Cerrado Biome acts as an exporter of virtual water [78,87,88] attached to the exported agricultural commodities from Brazil. The high water consumption required to sustain agricultural commodity exports is mainly responsible for increasing surface water scarcity to supply human livelihood and biological processes in the study region [89]. There is a perception that the agricultural producers are “shooting themselves on the foot” by destroying important ecosystem services, which ultimately will affect both environmental and economic sustainability and the water supply, which has been increasingly critical to maintain local communities in the Cerrado biome [89].

Therefore, any expansion of agricultural lands in the Cerrado biome should be properly and rigorously evaluated based on the water requirements of each user. It would also require strengthening participation in the water management process (e.g., watershed committees), aiming to prevent unequal or overloaded water consumption in the Cerrado biome. We foresee that legal reserves on private properties and other protected areas can substantially contribute to the conservation of water resources and limit the expansion of agricultural lands by preserving a native vegetation [90] quota per watershed capable of maintaining water flows at satisfactory levels.

5. Conclusions

Our results suggest that land use and climate changes have directly affected surface water availability in the Cerrado biome in Brazil, and this is likely to be aggravated in the coming years, assuming the future land use and climate changes used in this study. The same results suggest the trends in our scenarios will threaten agricultural production, public and private water supplies, and energy supply, as was experienced in Central Brazil in 2021.

Future scenarios indicate that more than 90% of watersheds within the Cerrado biome tend to decrease their water flow due to the expansion of agricultural lands. This may lead to recurrent critical levels of water scarcity during dry periods in the study region. The watersheds spatially located in the west of Bahia state, for example, showed clear decreasing trends and significant reductions in river flows, which will potentially worsen surface water availability, especially during dry periods, as found in some study cases [18].

Our analysis indicates we are embracing an uncertain future of surface water availability in the Cerrado biome. Most of the studied watersheds are subjected to high rates of deforestation linked to the equally high expansion of agricultural lands. For the medium and long term, we predict future critical levels of surface water supply and intensification in the current water conflicts [26] in the medium and long term. We conclude that Cerrado’s water capacity has been substantially weakened by agricultural activities and climate changes. Yet, the Cerrado biome has been considered “Brazil’s water reservoir”, although it is now heavily “leaking” and showing future scenarios of brief water scarcity.

Stakeholders overusing water resources, or those with insufficient access to water in the study region, will be directly affected by a decreasing surface water availability, and will be challenged to find alternatives to sustainable use of water and land. It is critical to heed nature’s warnings and reorient processes of land occupation and production in the Cerrado biome. Improving the production chains of non-timber forest products and production systems more adapted to local conditions could substantially contribute to protecting the native vegetation in the Cerrado biome.

We concluded that land use changes and the additional effects of climate change are impacting water availability in the Cerrado biome. Climate and land use changes are closely linked to the international consumption of certain agricultural commodities, which ultimately impacts local communities and increases water conflicts at local and regional levels. Finally, our study results can support the definition of political strategies and the legal measures needed to mitigate the negative effects of water overuse, and avoid an overwhelming expansion of agricultural lands into Cerrado.