Hydrologic Response to Land Use Change in a Large Basin in Eastern Amazon

: Accelerated land use changes in the Brazilian Amazonian region over the last four decades have raised questions about potential consequences for local hydrology. Under the hypothesis of a lack of frontier governance, projections of future changes in the Amazon basin suggest that 20–30% or more of this basin could be deforested in the next 40 years. This could trigger a cascade of negative impacts on water resources. In this study, we examined how a future conversion of the forest into pasture would inﬂuence streamﬂow and water balance components by using a conceptual and semi-distributed hydrological model in a large (142,000 km 2 ) forested basin: speciﬁcally, the Iriri River basin in the Brazilian Amazon. The results showed that the land use change could substantially alter the water balance components of the originally forested basin. For example, an increase of over 57% in pasture areas increased a simulated annual streamﬂow by ~6.5% and had a signiﬁcant impact on evapotranspiration, surface runoff, and percolation. Our ﬁndings emphasize the importance of protected areas for conservation strategies in the Brazilian Amazonian region.


Introduction
Tropical forests have been the focus of scientific and political discussions on surface energy fluxes, hydrological and carbon cycles, vegetation dynamics, land use, and human alteration of the biosphere through agriculture, mining, and urbanization. Land use change (LUC) has aroused significant interest, especially in the largest tropical forest in the world. The Amazon has undergone an accelerated transformation of its forests in the last four decades which has had environmental and socioeconomic impacts [1][2][3][4][5][6][7][8][9]. The effects of current and future LUC on the hydrology depend on soil-climate-topography-vegetation interactions which vary among basins, and for this reason, these effects remain difficult to predict [1,10,11].
Several studies have analysed different aspects of land use changes in the Brazilian Amazon, including governmental policies to encourage the implementation of agricultural projects, paving roads, population pressure, the commodity market, land speculation, timber exploitation, land tenure conflicts, and lack of public regulation [1,2,4,[12][13][14]. Projections made in 2006 on the future of deforestation in Amazon basin in the absence of frontier governance suggested that 20-30% or more of the basin could be deforested in the next 40 years [8]. However, in practice, since 2006, the deforestation rate has been reduced as a result of the enforcement of laws, commitments by the soy and beef supply chains, restrictions on access to credit, and expansion of protected areas (strictly protected indigenous land and sustainable land use). Between 2005 and 2014, the deforestation rate fell by almost 80%

Study Area
The Iriri River Basin is located at the eastern region of the Amazon Basin (Figure 1), between the Brazilian states of Mato Grosso (headstreams) and Pará, and is one of the main tributaries of the left bank of the Xingu River. The total length of the Iriri River is approximately 1,300 km and the total area of the Iriri River basin is 142,000 km 2 . About 80% of the basin, which comprises eleven indigenous territories and seven protected areas, is legally protected from deforestation.
The Iriri River basin is situated in a climatic transition region between the tropical monsoon climate (Am) in the north and tropical savanna climate (Aw) in the south, according to the Köppen classification. The main atmospheric systems responsible for climatic dynamics in this region are the Intertropical Convergence Zone and the South Atlantic convergence zone. Due to a south-north gradient, the annual average precipitation rate is 1800 mm per year, ranging from 1500 to 2500 mm per year. Seasonality is well-defined, with a rainy season lasting from December to June and a dry season from July to November. Relative air humidity is above 80% almost every month of the year. The average temperature ranges between 24 • C and 28 • C. Minimum temperatures rarely fall below 16 • C and mean maximum temperatures range between 33 • C and 36 • C, with the highest values generally occurring between August and March ( Figure 1B) [28][29][30].
The area is formed by Precambrian crystalline complexes and meta-volcano-sedimentary rocks as well as Pleistocene sediments. In general, this type of lithology is dominated by fractured aquifers covered by weathered mantles (soils and altered rocks) of variable thickness. The predominant soil in the basin is yellow Acrisol which occupies approximately 86% of the total area [28].
The vegetation in the Iriri River basin is characterized by continuous forest (approximately 90% of the total area [31]) formed by open ombrophilous forests (moist tropical rainforest) and climatic gradients with more than 60 dry days as well as dense ombrophilous forests with dry periods lasting for less than 60 days. Enclaves of savanna (cerrado) are associated with ombrophilous forest [32]. In 2014, pastures accounted for 3.2% of the basin [31].

Description of the SWAT Model
The SWAT model was designed to predict the impact of land use and management on water, sediment, and agricultural chemical yields at catchment scale at daily, monthly, and annual time increments. In SWAT, the basin is divided into sub-basins comprising a river segment and hydrological response units (HRUs), which represent areas of homogeneous land use, topography, and soil characteristics [25][26][27].
By modeling the hydrological cycle, SWAT computes the amount of water in the streams and in the soil as well as the fluxes of surface runoff (SurfQ), evapotranspiration (ET), and the amount of water stored in shallow and deep aquifers. The model also determines sediment, nutrient, and pesticide yields via routing processes to describe the movement of water, sediments, and other components through the channel network of the basin to the outlet [27,33]. In this study, we focused on the hydrological processes to predict streamflow (Q), SurfQ, ET, and groundwater. The simulations of these components are based on the water balance given in Equation (1) [25]: where SW t is the final soil water content (mm H 2 O); SW 0 is the initial soil water content (mm H 2 O); t is time (days); R dayi is precipitation on day i (mm H 2 O); Q surfi is surface runoff on day i (mm H 2 O); E ai is evapotranspiration on day i (mm H 2 O); w seepi is the amount of water from the soil profile inflowing to the vadose zone on day i (mm H 2 O); and Q gwi is the base flow on day i (mm H 2 O).

Model Input Data
The main input data used to characterize the basin, soil types, land cover, topography (Figure 2), and climate (temperature and precipitation) are spatialized over the basin. To fulfill this requirement, the main input data in the model in this study include remote sensing and reanalysis products ( Table 1).  The European Reanalysis (ERA) Interim temperature estimate dataset was cross-referenced with in situ data from the Altamira weather station ( Figure 1A). When compared with in situ data, it was found that, on average, the ERA Interim minimum temperatures dataset was systematically overestimated by 1.53 • C, whereas the maximum temperature dataset was underestimated by 6.74 • C. Therefore, a linear bias correction was applied to the ERA Interim dataset by summing −1.53 • C and +6.74 • C to the minimum and maximum ERA Interim temperatures, respectively. In this study, TRMM 3B42 V.7 Daily product precipitation estimates was the forcing variable of the model, and the data series used (daily precipitation rates from 1998 to 2015) were validated for the Xingu Basin region by cross-referencing the dataset with in situ P gauge (Altamira weather station). This validation resulted in an R 2 of 0.76 and Relative Root Mean Square Error-% (RMSE) of 31%, indicating that TRMM 3B42 V.7 can be used to represent the P rates in the Iriri River basin for modeling purposes.

•
Land Use/Vegetation Cover Parameters regarding the physical and physiological characteristics of the vegetation were required to model the interactions between vegetation and water, especially the ET process. SWAT has a database with recommended vegetation parameter values for several vegetation types and crops. These include three types of forests: evergreen, deciduous, and mixed forests. However, the'SWAT vegetation database was developed for temperate zones and was not suitable for tropical climates [39][40][41].
Tropical evergreen forest, which is the predominant vegetation in the Iriri River basin, shows a high spatial variability in canopy properties that affects ET across the region. Furthermore, forest activity (phenology and productivity) in the Amazonian region is mainly controlled by solar radiation (leaf flushing usually coincides with periods of increased radiation) [42,43] and soil water availability (the forest continues to grow even during the dry season, since its deep root system is able to reach water stored in deep soil layers, and there is more radiation available in the dry season because of less cloud cover) [42,[44][45][46][47]. Therefore, we adapted the SWAT database (evergreen forest, pasture, and savanna) to tropical conditions using a combination of parameter values found in the literature [48][49][50].
• Soil Some soil characteristics, including porosity and available water capacity, are crucial for computing soil water content and hence the water balance. The SWAT soil database was developed for soils in temperate climate areas and its parameter values are consequently not representative of tropical soils. Therefore, the soil data and parameters were adapted for this study.

Model Setup
SWAT2012 version [63] was set up for the Iriri River basin using the input data listed in Table 1. The basin was divided into 18 sub-basins, delineated by stream confluences defined using the Digital Elevation Model (DEM). The 18 sub-basins were then divided into 407 HRUs, representing homogenous land use, soil, and slope classes.
The study area lacked several elements of meteorological data. Therefore, the Hargreaves method was used to estimate potential ET [64] because, among the methods available in SWAT, this requires fewer meteorological data. Thus, this choice was also an attempt to avoid the introduction of more uncertainties into the model, which potentially may have arisen from estimating more meteorological data. The surface runoff (SurfQ) was calculated using the Soil Conservation Service's Curve Number method [65] and the Muskingum method was used for channel routing [66].
The We used the Land Use Update Tool [67] to update land use distribution by updating the HRU fractions during the model run.

Calibration and Validation
Model calibration was based on the optimization of the parameter values by adjusting the simulated streamflow (Q sim ) at the outlet of the basin with observed streamflow (Q obs ). The monthly Q obs data from the gauge Pedra do Ó (station 18700000, with a basin area of 123,827 km 2 ) ( Figure 1A) was available for the period 1999-2012 (ANA-Brazilian National Water Agency). In this study, the parameters related to ET, groundwater, and Q were initially adjusted by manual calibration (trial and error). The Parameter Solution (ParaSol) [68] implemented in the SWAT-CUP software package [69] was used for sensitivity analysis, model calibration, and uncertainty analysis. The ParaSol method aggregates objective function (OF), which is calculated based on the model outputs and on the observed data, into a global optimization criterion (GOC). Then, the algorithm minimizes these OF's or a GOC using the Shuffle Complex Evolution algorithm (SCE-UA) and performs an uncertainty analysis with a choice between two statistical concepts. This algorithm uses the sum of squares of the residuals (SSQ) as an objective function and the Nash-Sutcliffe Efficiency (NSE [70]) as a goal function to determine the best parameter values [68,71].
Validation was performed using Q obs dataset that was not used for calibration, and model performance was assessed using the periods of calibration (2001-2009), validation (2010-2012), and the whole series (2001-2015) using the coefficient of determination (R 2 ), the NSE, the Percent Bias (PBIAS) [72], and the Root Mean Squared Error Observations Standard Deviation Ratio (RSR) [73].

Land Use Scenario
We used the 'business-as-usual' (BAU) scenario developed by Soares-Filho et al. [8] through an application of the SimAmazonia model. This model compares the potential influence of protected areas and other conservation approaches on future trends in the Amazon basin to vegetation types, using an empirically-based, policy-sensitive model of Amazon deforestation. The SimAmazonia model simulates monthly deforestation in the Amazon Basin from 2002 to 2050 and includes eight scenarios that encompass a plausible range of future deforestation trajectories.
BAU is an extreme scenario that predicts deforestation using historical images (PRODES data [16]) and their variations between 1997 and 2002 as well as adding the effect of paving a set of major roads. The BAU scenario assumes the deforestation trends that occurred across the Amazon basin up to 2006 will continue in the future, i.e., that highways scheduled for paving at the beginning of the 2000s will be paved, compliance with legislation requiring forest reserves on private land will remain low (like at the beginning of 2000s), and no new protected areas will be created. The BAU scenario also assumes that 40% of the forests inside protected areas and 85% outside protected areas will be subject to deforestation [8,74].
The BAU scenario presents three patterns of land use: "deforested", i.e., the removal of native forest; "forest", i.e., the native forest; and "non-forest", i.e., other types of native vegetation (e.g., savanna). In SWAT, deforested areas were set as pasture. Indeed, pastures for livestock farming occupy approximately 60% of the total deforested areas throughout the Brazilian Amazon [31]. "Forest" was set as "evergreen forest" and "non-forest" areas corresponded to "savanna" areas in the Iriri River basin. Savanna areas represent 2.51% of total area in the basin and this percentage has remained unchanged over the years. Figure 3 gives an overview of the future LUC scenario in the Iriri River basin for the years 2020, 2030, 2040, and 2050.

Model Calibration
Model calibration was divided into two stages: manual and automatic calibration. Parameters were manually calibrated to adjust the evapotranspiration (ET) rates and leaf area index (LAI) curve. The LAI curve for evergreen forest, pasture, and savanna were adjusted thereby changing the SWAT vegetation parameters that determine the shape of the LAI curve to match observations [40,43,75,76]. This reported LAIs with annual averages of 4.02 m 2 /m 2 for evergreen forest, 1.09 m 2 /m 2 for pasture, and 1.25 m 2 /m 2 for savanna. The ET was determined by changing SWAT vegetation, soil, and groundwater parameters and led to an annual average ET rate of 56% of the annual average precipitation, corresponding to the rates reported for the eastern Amazon [44,47,77,78] and to data from the Altamira weather station ( Figure 1A).
Global sensitivity analysis was performed using 14 parameters related to channel, groundwater, and runoff, resulting in eight parameters ranked as significantly sensitive (p-value < 0.05). These parameters were then included in the automatic calibration procedure. The sensitivity rank of the parameters and the results of the automatic calibration are shown in Table 2. Calibration reduced CN2 values for tropical evergreen forest when compared with pasture, confirming the expected high infiltration in forested areas [79][80][81]. The CH_N2 calibrated value represents natural streams with heavy timber and brush [82], which is expected for reaches in forested areas like the Iriri River basin. For parameter CH_K2, the calibrated value (63 mm h −1 ) indicates that the river bed material is mainly a mixture of sand and gravel with low silt-clay content [83]. This is reasonable and agrees with the classification of streams in the eastern Amazon Basin as clear-water rivers with little suspended sediment [84]. The values calibrated for the parameters related to groundwater (GW_DELAY and GWQMN) are coherent because the Iriri River basin is part of the Brazilian shield, which usually implies deep thin aquifers with low yields [85]. The difference between evergreen forest and pasture groundwater parameters can be explained by faster lateral runoff from pasture than that from forest.

Efficiency
As shown in Figure 4A, the resulting simulated monthly streamflow (Q sim ), using the best set of calibrated parameter values, reproduced the observed streamflow (Q), as confirmed by efficiency criteria (Table 3) and by the percentage prediction uncertainty (95PPU is calculated at the 2.5% and 97.5% levels of the cumulative distribution of an output variable obtained through Latin hypercube sampling [69]).  The resulting 95PPU was narrow and did not cover most of the extreme high and low flow periods, corresponding to a P-factor (percentage of observations covered by the 95PPU) and R-factor (average thickness of the 95PPU band divided by the standard deviation of the measured) of 0.64 and 0.58 respectively for the entire series. Figure 4B shows the relative error between Q obs and Q sim (positive values indicate overestimated Q sim and negative values indicate underestimated Q sim ). For some episodes in the calibration and validation periods, Q sim was slightly underestimated during peak flow, compared with observed streamflow (Q obs ). Inversely, the model overestimated minimum values of Q obs . In reality, the recession was faster and the baseflow during the dry season was lower. This error could not be corrected by an increase of baseflow recession constant (ALPHA_BF parameter) because it would lead to overestimating streamflow during the beginning of the rainy season (October to January). The higher relative errors for 2010 and 2011 dry seasons are explained by the very low values of streamflow during these years of strong dryness, with a monthly mean of 73 and 111 m 3 ·s −1 , respectively for September 2010 and September 2011, which is largely beneath the mean of 201 m 3 ·s −1 for September months of the period.

Water Balance Components at the Scale of the Hydrological Response Units
To better understand the influence of land use in water balance components (WBC), we analyzed average annual ET, surface runoff (SurfQ), and Percolation (Perc) simulated at the HRU scale for forest and pasture in Acrisol (which represents~86% of the total area of the Iriri River basin). The results showed that ET and Perc are higher in forest HRU (>3% per year and >4% per year, respectively) than in pasture HRU, on the other hand, SurfQ is higher in HRU represented by pasture (>7% per year) ( Figure 5). This is in agreement with other studies showing that ET and Perc are lower and SurfQ is higher in pasturelands [78,[87][88][89][90].

Scenario Application
The LUC scenarios were run using identical weather data (the 17-year data series) so that variations in Q and other WBC were solely attributable to LUC. The LUC increased the annual average Q from 0.3% to 6.5% between 2020 and 2050, compared with the baseline Q (Q was simulated using land use in 2014 based on data from the TerraClass Project [31]) (Table 4). The monthly average Q of the baseline and of the future scenarios are presented in Figure 6A. During the first part of the recession phase, from the peak flow to the end of the rainy season (March to June), the simulated Q was slightly lower than baseline values, particularly in 2040 and 2050 when the area under pasture has been extended. Inversely, in the remaining year period, the expansion of pasture across the scenarios caused an increase in Q, particularly at the beginning of the rainy season, from October to January. The percent differences ( Figure 6B) showed that even a slight change in land cover (2020: pasture expands from 3.2% to 8.9% of the total basin area) is sufficient to increase Q by up to 8% during the rising water phase in December. The consequence of deforestation for Q was even more dramatic in 2040 and 2050, with a maximum increase of 99% in December and a decrease during low Q with a reduction of up to 8% in April. LUC also had a marked impact on other WBC at basin scale. At yearly scale, ET decreased from −0.1 to −10.6% between 2020 and 2050 (Table 4). ET decreased sharply during the dry season with a 10 to 20% reduction from June to October (Figure 7), caused by the reduction in pasture ET compared with forest ET as a result of higher stomatal leaf resistance and shallower grass roots. The variation was greater for 2040 and 2050 with the major expansion of pasture. The increase of SurfQ in pasture, compared with forest, was caused by lower vegetation cover and higher soil compaction by grazing cattle, which was represented by the Soil Conservation Service's Curve Number (CN) in the SWAT model. At yearly scale, the SurfQ increased drastically from 6.6% in 2020 to 183.1%, in 2050 (Table 4). The monthly SurfQ dynamics depended on the precipitation (Figure 8). SurfQ remained negligible from June to October (<5 mm/month). The increase in SurfQ caused by deforestation was higher during the rainy season with a maximum in March (+26 mm, equivalent to a 158% an increase compared to the baseline). Calculated on an annual average basis, Perc decreased from −0.1% to −21.9% between 2020 and 2050 ( Table 4). The monthly Perc dynamics depended on precipitation (Figure 9). Perc remained negligible in the dry season, from July to October (<5 mm/month). From November to December, Perc increased slightly with deforestation compared to the baseline. Inversely, from January to June, Perc was strongly reduced. Perc is controlled by ET and infiltration capacity. Considering the soils and the climate of the Iriri River basin, the actual ET was significantly reduced from November to December (Figure 7) and as the soil infiltration capacity was not saturated, Perc increased slightly. From January to June, actual ET was weakly reduced and the soil infiltration capacity in pasture was saturated by higher precipitation.

Discussion
Based on hydrological modelling, our study shows that changes in land cover significantly affect streamflow (Q) and water balance components (WBC) in a tropical basin. Calculated on an annual average basis, these results are in accordance with those of other studies showing that deforestation increases Q [10,24,78,87,88,[90][91][92]. In the Xingu River Basin, Panday et al. [78] found that deforestation of 15% of the basin over a period of 30 years caused a 6% increase in annual Q. Also in the Xingu River Basin, Cristina et al. [87] highlighted a greater impact of deforestation; at the scale of completely deforested and cropped small catchments, the measured mean Q in soy catchments was three times greater than that of forest catchments.
Concerning the impact of land use change (LUC) on WBC, previous studies have reported that a reduction of forested areas reduces evapotranspiration (ET) from both canopy interception and plant transpiration and simultaneously increases surface runoff (SurfQ) and Q, particularly during the rainy season [78,87]. The deep roots of the trees enable them to tap the water stored in deep soil layers and to maintain a high ET rate during the dry season [9,42,[45][46][47]93]. In contrast, even in wet equatorial areas, pastures converted to forest are subject to a decline in photosynthetic activity in the dry season as a consequence of increasing water stress [46,47]. The greater sensitivity of pasture to drought than primary forest can be explained by grass phenology and shallower roots; these roots, which have less access to deep soil water, reduces the ET. Consequently, the water stored in the deep soil is less depleted in pastureland and Q begins to rise earlier.
When forest is transformed into pasture, soil compaction due to trampling cattle as well as reduced vegetation cover and organic matter, reduce infiltration capacity and increase the Horton overland flow in pastures compared with forest [88,89,[94][95][96]. In small catchments of the Upper Xingu River Basin, on Oxisols, Scheffler et al. [97] measured a reduction in infiltration from 1258 mm/h to 100 mm/h after conversion of forest into pasture. Consequently, a larger fraction of precipitation is transformed into SurfQ thereby reducing the vertical flow path caused by infiltration, as shown in our results on Percolation (Perc) (Figure 9). The groundwater reserves are thus insufficiently replenished in the rainy season and there is a major decline in Q at the beginning of the dry season, as measured by Bruijnzeel [98], which is in agreement with the lower Q simulated between April and June in the Iriri River basin ( Figure 6). These results are in accordance with those of Bruijnzeel [99], who found an increase in baseflow and reduced ET after deforestation when the surface infiltration characteristics were maintained as well as a reduced baseflow when infiltration was strongly impaired. Inversely, Bruijnzeel's [99] results showed an increase in Q from low water to peak flow (July to February), a lower ET, and a higher SurfQ. Modelling the replacement of forest by pasture in large savanna basins in the Amazon, Lamparter et al. [11] also found an increase in monthly Q at the beginning of the rainy season. These authors explained the increase during this period by a reduced ET and an increased SurfQ. However, in the present study, modelling the impacts of LUC did not lead to higher peak Q in contrast to the results reported by other studies [10,98,100,101].
Our model did not include climate feedback such as reduced precipitation caused by deforestation nor global climate change. Previous studies [10,23,24] analyzed the climate-vegetation balance in the Amazon region; extensive deforestation affects atmospheric dynamics, leading to a reduction in precipitation and consequently an increase in the duration and intensity of the dry season in the Amazon basin with major changes in WBC. The increased water yield simulated by our model may be partially or fully offset by reduced precipitation with complex non-linear relationships between deforestation and Q. After modeling recent decades in the Xingu River basin, Panday et al. [78] suggested that climate variations led to a decrease in the annual average Q as a result of a decrease in precipitation and an increase in ET; this was partially offset by the increase caused by deforestation. The simulations they performed, from the 1970s to the 2000s, suggested that climate variations alone accounted for a −14% decrease in annual discharge while deforestation alone caused a +6% increase in annual discharge. Thus, climate feedback may compensate land use change effects on the Iriri basin.
The results shown here depend on the design and assumptions of the SWAT model and represent one example of an application of this model in the Amazon region. Although the SWAT model was originally developed for temperate zones and that some input data and the method require certain improvements before it can be applied in tropical regions [39][40][41], there has been an increase in its application in tropical zones in recent years. Many of these studies reported that the model performed well for simulated Q and/or water quality (model calibration and validation is usually based on these parameters), but these results do not say which hydrological processes, such as ET, SurfQ, and Perc, were simulated correctly [40]. Our calibration results represented a good performance (calibration based on monthly Q data) and the simulated ET (calibration based on yearly data) matched ET rates measured or estimated for the region [44,47,77,78] as well as showed a decrease during the dry season correlated with water stress. However, this result contradicts the observational syntheses of Shuttleworth [102], who showed that during the first months of the dry season, ET was not lower than the wet season in the Amazon rain forest. More recent studies have drawn the same conclusions [44,46,47]: water stored in the soil that is available for root uptake is sufficient to maintain the ET rate. But independent of monthly ET results, our findings strongly support the likelihood that deforestation will have serious consequences for WBC in a large basin in the Amazonian region.

Conclusions
In this study, we used the SWAT model to investigate the hydrologic response to land use change (LUC) in a large river basin in the eastern Amazon. The LUC involved progressive deforestation (between 2020 and 2050), from >8 to <57%, in the Iriri River basin. The results of this study point to the serious potential consequences of deforestation on monthly streamflow and water balance components in an Amazonian basin. A~6% decrease in the forested area (2020) would lead to a negligible increase in streamflow (0.3%); however, a~57% change in land cover from forest to pasture (2050) would cause ã 6.5% increase in streamflow as well as higher surface runoff, less percolation, and evapotranspiration compared to the baseline scenario.
Today, about 80% of the Iriri River basin is legally protected area. Our findings emphasize the importance of protected areas for conservation strategies in the Brazilian Amazonian region; these include maintaining the climate-vegetation balance and hydrological regimes, preventing forest fires, as well as sustaining traditional livelihoods and "green barriers" to deforestation [6,8]. The insights acquired in this study provide further evidence of and useful references for the importance of regional land use and water resources planning in the Amazon basin.