Remote land use impacts on river flows through atmospheric teleconnections

Abstract

The effects of land-use change on river flows have usually been explained by changes within a river basin. However, land–atmosphere feedback such as moisture recycling can link local land-use change to modifications of remote precipitation, with further knock-on effects on distant river flows. Here, we look at river flow changes caused by both land-use change and water use within the basin, as well as modifications of imported and exported atmospheric moisture. We show that in some of the world’s largest basins, precipitation was influenced more strongly by land-use change occurring outside than inside the basin. Moreover, river flows in several non-transboundary basins were considerably regulated by land-use changes in foreign countries. We conclude that regional patterns of land-use change and moisture recycling are important to consider in explaining runoff change, integrating land and water management, and informing water governance.

Introduction

River flows ($Q$) are fundamental for ecosystems, nutrient transport, hydropower, navigation, and human well-being . Land-use change (LUC) has been suggested to be the most important driver of both past and future changes in river flows ($\mathrm{\Delta }Q$). Central to the analysis of $Q$ is the river basin unit, and estimates of $\mathrm{\Delta }Q$ from LUC often assume that impacts occur exclusively within a basin . Water governance is strongly focused on frameworks such as the Integrated River Basin Management (IWRM) and largely assumes that there is no land–atmosphere feedback, even in discussions of spatial misfit between institutions and hydrological realities . In fact, land–atmosphere feedbacks are not incorporated into most recent literature on a wide range of topics of relevance for water management, such as virtual water , the freshwater planetary boundary , water scarcity , the relative role of climate and LUC for water flows , and land acquisition impacts on water .

However, studies on land–atmosphere interactions clearly show that changes in land surface properties can considerably influence precipitation ($P$) and $Q$ through land–atmosphere feedback, sometimes well beyond the local scale . For example, general circulation model simulations suggest that complete deforestation of central Africa may decrease February $P$ by 35 % in the Great Lakes region , and irrigation in India may support up to 40 % of the $P$ in some arid regions in eastern Africa . Under a business-as-usual deforestation scenario, $Q$ in the Xingu River basin in the Amazon was found to increase by 10–12 % without land–atmosphere feedback, and decrease by 30–36 % when such feedback was taken into account . Furthermore, statistical analyses of observed data suggest that irrigation in the US High Plains enhances downwind $Q$ , and coupled regional climate modelling shows that irrigation in the California Central Valley can be linked to an about 30 % increase in Colorado $Q$ . At the global scale, $\mathrm{\Delta }Q$ from future climate and LUC scenarios changed from decrease to increase by considering land–atmosphere feedback and by closing the water balance .

Land–atmosphere interactions can influence $Q$ through thermal layer processes, terrestrial moisture recycling (TMR), and circulation perturbation . First, thermal layer processes refer to the boundary layer and mesoscale circulation perturbation that may lead to a change in total terrestrial evaporation ($E$) and can locally lead to both positive and negative $P$ responses . Local forest clearing has for example been shown to enhance $P$ in downwind areas due to turbulence changes . Second, TMR refers to the process of terrestrial $E$ returning to land as $P$ and is underpinned by the mass conservation of water . TMR is often the dominating land–atmosphere process at the regional to continental scale . About 40 % of global terrestrial $P$ originates from terrestrial $E$ and the average distance travelled in the atmosphere is 500–5000 $\mathrm{km}$ – a distance likely to exceed the size of most river basins. Lastly, large-scale atmospheric circulation perturbation allows extreme LUC (e.g. complete tropical deforestation) to impact $P$ in geographically remote regions and continents in unexpected ways . Monsoon regions are particularly sensitive to circulation perturbation, and irrigation may for example reduce $P$ by weakening the monsoon onset .

The previous studies that illustrated the importance of remote LUC for basin $P$ and $Q$ did not examine the effect of taking moisture recycling into account for estimating LUC effects on $Q$ and attributing them to influence from different nations, nor did they analyse the interplay between LUC within and outside the river basin. These effects are, however, important to disentangle since they can have profound water governance implications for, for example, riparian water rights and transboundary river basin treaties . Thus, there is a missing interdisciplinary bridge between understanding the role of land–atmosphere feedback over large distances and its importance for water governance at the basin scale.

This study aims to (i) investigate the potential impacts of human LUC on $Q$ worldwide accounting for TMR, (ii) disentangle the relative influence on $Q$ from within- and extra-basin LUC, (iii) attribute potential human LUC impacts on $Q$ to nation states, and (iv) discuss the potential implications for water governance. We focus on the TMR effect because it is transparent, closes the water balance, and explicitly links changes in land and water geographically. Given these advantages, similar TMR approaches have in recent years been used to analyse unexplored relations, e.g. LUC impacts of crop yields , self-amplifying forest dieback from TMR changes , and vulnerability to LUC-induced reductions in $P$ . For a comparison of different methods for analysing LUC impacts on $Q$, see Table S1 in the Supplement.

Methods

Modelling

Hydrological modelling

We used the process-based Simple Terrestrial Evaporation to Atmosphere Model (STEAM) hydrological model to simulate water fluxes based on land cover and land use. STEAM partitions evaporation into five fluxes: vegetation interception, floor interception, transpiration, soil moisture evaporation, and open-water evaporation. STEAM uses the Penman–Monteith equation to estimate potential evaporation, the Jarvis–Stewart equation to compute stomatal resistance, and Jolly's growing season index (function of minimum temperature, soil moisture content, and daylight) to describe phenology . STEAM operates at $\mathrm{1.5}{}^{\circ }\times \mathrm{1.5}{}^{\circ }$ and a 3 $\mathrm{h}$ resolution. Based on the long-term water balance, mean annual river flow ($Q$) is assumed to approximately equal the difference between mean annual $P$ and $E$, i.e. $Q=P-E$. STEAM was validated in previous studies and compared well with recent observation-based analyses of evaporation partitioning by land-cover type . Modifications from the original version of STEAM include (1) update of land-use classification, parameterization, and parameterization approach (Table S2 and Fig. S1 in the Supplement), (2) use of a temperature threshold of 0 ${}^{\circ }$C for differentiating snowfall from rainfall, and (3) differences in input data (i.e. root zone storage capacity, land surface map, precipitation data source as described in Data). Evaluation against runoff data is shown in Fig. S2. Simulated land-use change effects on evaporation increase and decrease are compared with literature values in Table S3 and found to be in the conservative range. With the study period being 2000–2013, the years 1995–1999 were used as spin-up for STEAM.

Moisture tracking

Atmospheric moisture is tracked using the Eulerian Water Accounting Model-2 layers (WAM-2layers) moisture tracking scheme . WAM-2layers tracks atmospheric moisture from zero pressure to surface pressure in two layers. Within the layers, atmosphere is assumed to be well mixed. WAM-2layers tracks vapour flows by applying the water balance. For example, the following equation is used to track where evaporation from a given region falls as precipitation (i.e. forward tracking):

$$\begin{array}{cc}{\displaystyle \frac{\partial S_{\mathrm{tracked}}}{\partial t}}& {\displaystyle }={\displaystyle \frac{\partial \left(S_{\mathrm{tracked}}u\right)}{\partial x}}+{\displaystyle \frac{\partial \left(S_{\mathrm{tracked}}v\right)}{\partial y}}\\ & {\displaystyle }& {\displaystyle }+E_{\mathrm{tracked}}-P_{\mathrm{tracked}}\pm F_{\mathrm{vertical},\mathrm{tracked}},\end{array}$$ where $S_{\mathrm{tracked}}$ is the tracked atmospheric storage in an atmospheric column in one layer, $t$ is time, $u$ and $v$ are wind components in the $x$ zonal and $y$ meridional directions, $E_{\mathrm{tracked}}$ is tracked evaporation entering and $P_{\mathrm{tracked}}$ is precipitation exiting an atmospheric column and layer, and $F_{\mathrm{vertical},\mathrm{tracked}}$ is the tracked vertical moisture transport between the two layers. An analogous equation is used for tracking the source of precipitation to a given region (i.e. backward tracking). The spatial resolution of WAM-2layers is 1.5${}^{\circ }$ and input data are linearly interpolated to the 15 $\min$ time step to maintain numerical stability. WAM-2layers has been employed previously for analysing atmospheric moisture transport over terrestrial areas and validated against other types of moisture tracking algorithms . We used the MATLAB version of WAM-2layers, but a Python version is also openly available on Github . With the study period being 2000–2013, the year 1999 is used as spin-up in forward tracking in WAM-2layers, and 2014 is used as spin-up for backward tracking in WAM-2layers.

Coupling of the moisture tracking scheme and the hydrological model

Hydrological flows in the current land-use scenarios are simply represented by current data and simulation. To obtain $E$ and $P$ under potential land cover, STEAM is coupled with WAM-2layers by (1) simulating present-day $E$ in STEAM and forward tracking terrestrial $E$ with WAM-2layers, meaning that the $E_{\mathrm{tracked}}$ is equal to all evaporation from terrestrial surfaces, i.e. not belonging to the oceans, (2) simulating $E$ in STEAM based on present-day $P$ and potential land cover, and forward tracking the fate of terrestrial $E$ with WAM-2layers, (3) calculating the change in $P_{\mathrm{tracked}}$, (4) updating the present-day $P$ with the changes in $P_{\mathrm{tracked}}$, and (5) simulating $E$ in STEAM based on updated $P$ and potential land cover, and forward tracking the fate of terrestrial $E$ with WAM-2layers; see Fig. . Steps 3–5 are iterated until the annual $P$ change is below 1 % and the monthly $P$ change is below 5 $\mathrm{mm}{\mathrm{month}}^{-\mathrm{1}}$ in every grid cell, which in our case ultimately resulted in four iterations in total. This procedure assumes that land-use induced changes in terrestrial $E$ will result in proportional changes in $P$ with terrestrial origin.

Model coupling schematic. Model coupling between STEAM and WAM-2layers based on current land use and potential vegetation scenarios. $P$ stands for current precipitation; $E$ stands for evaporation. Subscript t stands for terrestrial origin, pv denotes simulation with potential vegetation, cur denotes simulation with current land use, and $n$ stands for the number of iterations.

[Figure omitted. See PDF]

Data

Land data

Land-use and land-cover data input to STEAM are based on the Ramankutty potential land-cover and current land-use scenarios for consistency. We further added permanent wetlands, permanent snow or ice, and urban or built-up areas from the Land Cover Type Climate Modeling Grid (CMG) MCD12C1 International Geosphere Biosphere Program (IGBP) land classification created from Terra and Aqua Moderate Resolution Imaging Spectroradiometer (MODIS) data for the year 2005. Monthly irrigated rice and irrigation non-rice crops were obtained from the data set of Monthly Irrigated and Rainfed Crop Areas around the year 2000 (MIRCA2000) V1.1 . The urban and irrigated areas were only added to the current land-cover map. In this merging procedure, MODIS is allowed to override the Ramankutty data sets, and MIRCA2000 is allowed to override the Ramankutty map as long as it does not extend over the cropland areas. The scenarios used are shown in Fig. S3 and the land-use change is illustrated in Fig. .

The root zone storage capacity map is based on a climate-observation-based root zone storage capacity ($S_{\mathrm{R}}$) derived from satellite and energy balance-based evaporation, gauge-based precipitation, and modelled irrigation. The best performing Gumbel normalized root zone storage capacity ($S_{\text{R,CRU-SM,merged}}$) was used. Root zone storage capacity for both current and potential land-cover and land-use scenarios was constructed from the mean of land-cover type and Köppen–Geiger climate class . The mean root zone storage capacity of single land-cover types was used only in places where the combination of land-cover type and climate zone that exists in the potential land-cover scenario did not exist in the current land-use map.

Changes in land use resulting from the replacement of the potential vegetation scenario with the current land-use map. Changes in (a) land use (current land use is shown, with grid cells without major land-use change masked out) and (b) total area of difference between potential vegetation ($y$ axis) and the current land-use map (colour legend).

[Figure omitted. See PDF]

Meteorological forcing and runoff data

Meteorological data used in WAM-2layers and STEAM, except for land precipitation, were taken from the Earth Retrospective Analysis Interim (ERA-I) from the European Centre for Medium-Range Weather Forecasts (ECMWF) . ERA-I meteorological forcings to STEAM are snowmelt, temperature at 2 $\mathrm{m}$ height, dew point temperature at 2 $\mathrm{m}$ height, wind speed (meridional and zonal vectors) at 10 $\mathrm{m}$ height, incoming shortwave radiation, and net longwave radiation. In addition, ERA-I evaporation data were used to downscale calculated daily potential evaporation in STEAM to the 3 $\mathrm{h}$ time step. ERA-I model level forcings used in WAM-2layers are specific humidity and wind speed at 6-hourly resolution, spanning from zero to surface pressure. Moreover, 3-hourly ocean evaporation is taken from ERA-I. The Modern-Era Retrospective analysis for Research and Applications (MERRA) reanalysis has in a previous study been used as input to WAM-2layers for comparison and generated similar persistent moisture recycling patterns, except in South America, where differences arise due to underestimation of precipitation in MERRA (Keys et al., 2014). Precipitation forcing for WAM-2layers and STEAM both come from the state-of-the-art Multi-Source Weighted-Ensemble Precipitation (MSWEP V1) product that was specifically created for hydrological modelling. The use of MSWEP as forcing for STEAM resulted in runoff estimates that compare well to observed runoff data (Fig. S2). All meteorological forcing data cover temporally 1995–2014.

Runoff data used for benchmarking were taken from the composite (observed river discharge consistent with the water balance model) from the Global Runoff Data Centre (GRDC) . The separate GRDC water balance model runoff fields are included in the comparison for reference (Fig. S2).

The spatial coverage of all data used is 57${}^{\circ }$ S–79.5${}^{\circ }$ N latitudes at $\mathrm{1.5}{}^{\circ }\times \mathrm{1.5}{}^{\circ }$ resolution. MSWEP originally at 0.25${}^{\circ }$ and GRDC runoff at 0.5${}^{\circ }$ were aggregated to 1.5${}^{\circ }$ resolution by simple averaging.

Analyses

Changes in hydrological flows

River flow change without TMR ($\mathrm{\Delta }{Q}_{\mathrm{noTMR}}$) is $$\mathrm{\Delta }{Q}_{\mathrm{noTMR}}=(P_{\mathrm{cur}}-E_{\mathrm{cur}})-(P_{\mathrm{cur}}-E_{\mathrm{pv},\mathrm{1}}),$$ where $P_{\mathrm{cur}}$ is current-day precipitation data from MSWEP, $E_{\mathrm{cur}}$ is current-day evaporation based on STEAM simulation, and $E_{\mathrm{pv},\mathrm{1}}$ results from STEAM simulation in the potential vegetation scenario and forced with current-day precipitation (Fig. ). River flow change after accounting for TMR ($\mathrm{\Delta }Q$) is $$\mathrm{\Delta }Q=(P_{\mathrm{cur}}-E_{\mathrm{cur}})-(P_{\mathrm{pv},\mathrm{4}}-E_{\mathrm{pv},\mathrm{5}}),$$ where $P_{\mathrm{pv},\mathrm{4}}$ is the converged precipitation (i.e. meeting the convergence requirement of mean annual precipitation change $<\mathrm{1}$ % yr${}^{-\mathrm{1}}$ and monthly precipitation change $<\mathrm{5}$ $\mathrm{mm}{\mathrm{month}}^{-\mathrm{1}}$ in every grid cell) achieved at the fourth iterative coupling between STEAM and WAM-2layers, and $E_{\mathrm{pv},\mathrm{5}}$ is the evaporation under the potential vegetation scenario simulated in STEAM with precipitation forcing $P_{\mathrm{pv},\mathrm{4}}$.

Change in tracked basin precipitation ($\mathrm{\Delta }{P}_{\mathrm{tracked},\mathrm{basin}}$) occurring outside the river basin boundaries is referred to as $\mathrm{\Delta }{P}_{\mathrm{import}}$, whereas $\mathrm{\Delta }{P}_{\mathrm{tracked},\mathrm{basin}}$ originating from within the basin boundary is referred to as $\mathrm{\Delta }{P}_{\text{basin-recycling}}$. Internally recycled evaporation ($\mathrm{\Delta }{E}_{\text{basin-recycling}}$) corresponds to $\mathrm{\Delta }{P}_{\text{basin-recycling}}$ and all other basin evaporation change is considered exported ($\mathrm{\Delta }{E}_{\mathrm{export}}$).

Country influence on changes in river flows

The influence on river flow change in river basin b from country c without considering TMR ($I_{\mathrm{b},\mathrm{c},\mathrm{noTMR}}$) is $$I_{\mathrm{b},\mathrm{c},\mathrm{noTMR}}=\left|\mathrm{\Delta }{E}_{\mathrm{b},\mathrm{c}}\right|,$$ where $\mathrm{\Delta }{E}_{\mathrm{b},\mathrm{c}}$ is evaporation change in the part of river basin b located in country c. The influence on river flow change in basin b from country c with consideration of TMR ($I_{\mathrm{b},\mathrm{c},\mathrm{TMR}}$) is $$I_{\mathrm{b},\mathrm{c},\mathrm{TMR}}=\left|\mathrm{\Delta }{E}_{\mathrm{b},\mathrm{c},\mathrm{export}}\right|+\left|\mathrm{\Delta }{P}_{\mathrm{b},\mathrm{c},\mathrm{import}}\right|),$$ where $\mathrm{\Delta }{E}_{\mathrm{b},\mathrm{c},\mathrm{export}}$ is the evaporation change exported from the part of basin b located in country c, and $\mathrm{\Delta }{P}_{\mathrm{b},\mathrm{c},\mathrm{import}}$ is the precipitation change imported to basin b from country c.

Influences from countries below 5 % of total influences in a specific basin ($I_{\mathrm{b},\mathrm{c},\mathrm{noTMR}}<\mathrm{0.05}\times \sum I_{\mathrm{b},\mathrm{c},\mathrm{TMR}}$) were lumped into the category “Other”.

Results

LUC impacts on global water flows

Our results show that human LUC (from potential land cover to current land use) (Fig. ) has led to reductions in $E$ and $P$, and to increases in $Q$, in large parts of the world (Fig. b–d). $E$ has decreased primarily in Southwest China, Europe, western Africa, the south of Congo, and south-eastern South America, resulting from substantial pasture and agricultural expansion . Following prevailing wind directions (Fig. c), subsequent $P$ has decreased in all tropical regions, southern central China, the eastern US, and Europe.

Current mean annual hydrological flows 2000–2013. (a) Current evaporation simulated by STEAM, (b) current precipitation (MSWEP data), (c) current continental precipitation recycling ratio (i.e. precipitation with terrestrial origin divided by total precipitation: $P_{\mathrm{tracked}}/P$) where arrows show average winds in the lower atmosphere, and (d) current river flow at outlet based on $P-E$. Values below about 0.5 % of the maximum display value are in grey.

[Figure omitted. See PDF]

Land-use change-induced changes in hydrological flows (current land-use–potential vegetation scenario): (a) absolute change in evaporation, (b) relative change in evaporation, (c) absolute changes in precipitation, (d) relative change in precipitation, (e) absolute change in continental precipitation recycling ratio (i.e. precipitation with terrestrial origin divided by total precipitation $P_{\mathrm{tracked}}/P$ and converted to the unit of percent), (f) relative change in continental precipitation recycling ratio, (g) absolute change in river flows at outlet, and (h) relative change in river flows at outlet.

[Figure omitted. See PDF]

Nevertheless, in some areas, $E$ increased due to incremental irrigation – notably in India, the western US, Northeast China, and the Middle East (Fig. a, b). Due to the combination of heavy irrigation in India and orography, $P$ has increased substantially along the Himalaya mountain ridge (Fig. b, c). Weak increases in $P$ are observed in other downwind regions: the Sahel (i.e. downwind irrigation areas along the Nile) and in the western US. Continental precipitation recycling ratios are modified – with some exceptions – in a similar pattern to $P$ (Fig. e, f). Large $\mathrm{\Delta }Q$ are seen in the La Plata basin in South America, the Zambezi in southern Africa, the Yangtze in China, and the Indus in northern India (Fig. g), and relative changes in $Q$ are large in for example the Colorado basin in the US, the Odra basin in eastern Europe, and the Lake Chad river basin in Africa (Fig. g).

The role of TMR for $\mathrm{\Delta }Q$

In aggregate (Fig. ), when accounting for TMR, LUC changed global terrestrial $E$ by $-\mathrm{1251}$ ${\mathrm{km}}^{\mathrm{3}}{\mathrm{yr}}^{-\mathrm{1}}$ ($-\mathrm{1.8}$ % from 69 211 ${\mathrm{km}}^{\mathrm{3}}{\mathrm{yr}}^{-\mathrm{1}}$), $P$ by $-\mathrm{586}$ ${\mathrm{km}}^{\mathrm{3}}{\mathrm{yr}}^{-\mathrm{1}}$ ($-\mathrm{0.5}$ % from 107 800 ${\mathrm{km}}^{\mathrm{3}}{\mathrm{yr}}^{-\mathrm{1}}$), and $Q$ by 664 ${\mathrm{km}}^{\mathrm{3}}{\mathrm{yr}}^{-\mathrm{1}}$ ($+\mathrm{1.7}$ % from 38 589 ${\mathrm{km}}^{\mathrm{3}}{\mathrm{yr}}^{-\mathrm{1}}$). The estimated changes to $Q$ tend to fall in the conservative end of previous estimates (Fig. ). However, recent research suggests that consumptive water use is severely underestimated in earlier studies e.g.. $\mathrm{\Delta }Q$ with TMR corresponds to the difference between $\mathrm{\Delta }E$ and $\mathrm{\Delta }P$ change including TMR (Fig. , solid bars), whereas $\mathrm{\Delta }Q$ without accounting for TMR simply corresponds to $\mathrm{\Delta }E$ without TMR (Fig. , hollow bars).

Human impact on global hydrological flows. The solid bars show our estimated net change (terrestrial area 131.7 million $\mathrm{km}$${}^{\mathrm{2}}$ and time period 2000–2013) in evaporation, precipitation, and river flows including consideration of TMR. Hollow bars show flow changes without TMR. Circles and stars indicate river flow change estimates from other studies (Table S4), where LUC implicitly accounts for consumptive water use. Note that while consumptive water use alone always reduces river flows, other human impacts have both positive and negative influences that are concealed by the global aggregate.

[Figure omitted. See PDF]

Including TMR nearly halves the global $\mathrm{\Delta }Q$ estimate. This is because $E$ returns as $P$ over land and thus compensates for the initial water “loss” from the basin. This suggests that previous studies without TMR e.g. may have substantially overestimated the net LUC impacts on $Q$. Our estimate of LUC impact on $Q$ is slightly larger than some of the estimates of ${\mathrm{CO}}_{\mathrm{2}}$ fertilization e.g., but substantially smaller than climate change and overall human impact e.g. (Fig. ).

Our river basin analysis shows that accounting for TMR considerably alters estimates of $\mathrm{\Delta }Q$ (Fig. a): in the Congo, Volga, and Ob basins, $\mathrm{\Delta }Q$ are reduced by more than half; in the Amazon, $\mathrm{\Delta }Q$ drops from 1630 to 270 $\mathrm{m}$${}^{\mathrm{3}}$ s${}^{-\mathrm{1}}$; and in the Yenisei, the sign of $\mathrm{\Delta }Q$ is reversed from an increase (150 ${\mathrm{m}}^{\mathrm{3}}{\mathrm{s}}^{-\mathrm{1}}$) to a decrease ($-\mathrm{220}$ ${\mathrm{m}}^{\mathrm{3}}{\mathrm{s}}^{-\mathrm{1}}$).

The effect of accounting for TMR on river flow change estimates, shown (a) as absolute difference between river flow change without and with TMR effect, i.e. $\mathrm{\Delta }{Q}_{\mathrm{noTMR}}-\mathrm{\Delta }Q$, (b) as this difference relative to river flow change with TMR effect, i.e. ($\mathrm{\Delta }{Q}_{\mathrm{noTMR}}-\mathrm{\Delta }Q)/\mathrm{\Delta }Q$, and (c) as this difference relative to current river flows, i.e. ($\mathrm{\Delta }{Q}_{\mathrm{noTMR}}-\mathrm{\Delta }Q)/Q_{\mathrm{cur}}$.

[Figure omitted. See PDF]

At the basin level, the TMR effect on river flow change is estimated to be the largest in large and relatively wet basins such as the Amazon, Congo, and Yangtze River basins in terms of absolute volumes (Fig. a). Not accounting for TMR clearly generates the largest relative deviations in river flow change estimates in the Amazon (i.e. $\mathrm{\Delta }{Q}_{\mathrm{noTMR}}$ is approximately 5 times larger than $\mathrm{\Delta }Q$), and large relative TMR effects are seen in many large basins worldwide, including e.g. the Congo ($\mathrm{\Delta }{Q}_{\mathrm{noTMR}}$ is 150 % higher than $\mathrm{\Delta }Q$), Yenisei ($\mathrm{\Delta }{Q}_{\mathrm{noTMR}}$ is 165 % lower than $\mathrm{\Delta }Q$), and Ob ($\mathrm{\Delta }{Q}_{\mathrm{noTMR}}$ is 140 % higher than $\mathrm{\Delta }Q$) river basins (Fig. b). The TMR effect relative $Q_{\mathrm{cur}}$ (Fig. c) shows that TMR effect can be important also in more arid basins such as the Colorado, Niger, and the Yellow River.

The interplay between internal and external LUC

Furthermore, atmospheric moisture does not respect river basin boundaries (Fig. a, and spatial maps in Figs. , , S4, S5, S6, and S7). In fact, $P$ over the basins has been modified more significantly by external than by internal LUC (change in imported precipitation $\mathrm{\Delta }{P}_{\mathrm{import}}>$ change in internally recycled precipitation $\mathrm{\Delta }{P}_{\text{basin-recycling}}$) in some of the largest basins (Fig. a). Likewise, internally recycled evaporation changes ($E_{\text{basin-recycling}}$) (Fig. b, II) are substantially smaller than $\mathrm{\Delta }E$, affecting $P$ elsewhere ($\mathrm{\Delta }{E}_{\text{basin-recycling}}<$ change in exported evaporation $\mathrm{\Delta }{E}_{\mathrm{export}}$) for all selected river basins (Fig. a).

Changes in hydrological flows at river basin scale. (a) Changes in hydrological flows in 10 of the basins with the largest TMR effect on river flows ($Q$) (8 basins with increased and 2 with decreased $Q$), and 5 basins with large relative changes in river flows (2 basins with positive $\mathrm{\Delta }Q$ and 3 with negative $\mathrm{\Delta }Q$). (b) Conceptual figure of hydrological flow changes in a basin. The ($-$) and ($+$) in (b) may be different for different basins, and the ($-$) and ($+$) as displayed here are for example seen in the Amazon, Congo, and Yangtze; see (a). Note also that the figure has two $y$ axes, ${\mathrm{m}}^{\mathrm{3}}{\mathrm{s}}^{-\mathrm{1}}$ to the left and % to the right

[Figure omitted. See PDF]

Internal moisture recycling (Fig. b, II) does not affect $\mathrm{\Delta }Q$ directly, but only indirectly if $\mathrm{\Delta }{P}_{\text{basin-recycling}}$ affects subsequent $\mathrm{\Delta }{E}_{\mathrm{export}}$ under transient change (Fig. b and Methods). Thus, provided steady state, $\mathrm{\Delta }Q$ simply corresponds to the difference between $\mathrm{\Delta }{E}_{\mathrm{export}}$ and $\mathrm{\Delta }{P}_{\mathrm{import}}$ (Fig. a). For example, $\mathrm{\Delta }Q$ in the Amazon is very small because the reduced $\mathrm{\Delta }{P}_{\mathrm{import}}$ is almost entirely offset by reduced $\mathrm{\Delta }{E}_{\mathrm{export}}$. In Congo, about half of the within-basin LUC-induced $Q$ increase is counteracted by extra-basin LUC (i.e. $\mathrm{\Delta }{P}_{\mathrm{import}}\approx \mathrm{0.5}\mathrm{\Delta }{E}_{\mathrm{export}}$). The effect of TMR on $\mathrm{\Delta }Q$ ($\mathrm{\Delta }{Q}_{\mathrm{noTMR}}$$-$$\mathrm{\Delta }Q$, where subscript ${}_{\mathrm{noTMR}}$ denotes simulation without TMR) corresponds to total $\mathrm{\Delta }P$ (i.e. $\mathrm{\Delta }{P}_{\mathrm{import}}+\mathrm{\Delta }{P}_{\text{basin-recycling}}$) and any indirect $\mathrm{\Delta }E$ (i.e. $\mathrm{\Delta }{E}_{\mathrm{noTMR}}$$-$$\mathrm{\Delta }E$, not shown). In the Yangtze, the $\mathrm{\Delta }Q$ is mitigated mostly by $\mathrm{\Delta }{P}_{\text{basin-recycling}}$. The strong flow reduction in the heavily irrigated Indus, however, is only mildly compensated for by TMR (i.e. $\mathrm{\Delta }{P}_{\mathrm{import}}\ll \mathrm{\Delta }{E}_{\mathrm{export}}$).

The pattern of overlapping precipitationsheds (i.e. $P$ source regions)and evaporationsheds (i.e. $E$ sink regions) illustrated in Fig. b and moderated by wind directions can also be clearly seen in the basin-specific precipitationshed and evaporationshed maps (Fig. ). In the Amazon (Fig. a, b), the moisture arrives from the east, is stopped up by the Andes, and changes direction towards the south-east. The hotspot of precipitation source and sink within the Amazon basin does not overlap, with major moisture providing spots located along the north-eastern border and the major moisture receiving spots located along the Andes in the west. In the Yangtze (Fig. c, d), the moisture comes from a large area in the south, and leaves in the direction of Japan in a relatively narrow band. In the Yenisei (Fig. e, f), the moisture follows the westerlies, coming in straight from the west, and leaving straight towards the east. In the Niger (Fig. g, h), the moisture is mostly supplied from the east from terrestrial areas, and flows towards the west into the Atlantic. For precipitationsheds and evaporationsheds of other basins, see Figs. S4 and S6 respectively.

While changes in precipitationsheds and evaporationsheds are conditioned by the original moisture flows, the resulting pattern ultimately depends on the distribution of LUC-induced hydrological change (compare Figs.  and ). For example, although the Amazon precipitationshed is weak over Africa (Fig. a), the precipitationshed change is in fact relatively strong there due to strong LUC-induced hydrological change (Fig. a). In other cases, aggregated changes in Fig.  hide spatially heterogeneous increases and decreases in moisture flows. For example, agricultural activities and irrigation in India, the Sahel, and regions around the Nile increase moisture supply to the Yangtze, Yensisei, and Niger basins and offset deforestation-induced moisture supply decrease elsewhere (Fig. c, e, g). For changes in precipitationsheds and evaporationsheds of other basins, see Figs. S5 and S7 respectively.

Mean annual precipitation sources and evaporation sinks for selected river basins (boundaries in orange).

[Figure omitted. See PDF]

Impacts of human land-use change on mean annual precipitation source (i.e. $\mathrm{\Delta }{P}_{\mathrm{import}}+\mathrm{\Delta }{P}_{\text{basin-recycling}}$) and evaporation sinks (i.e. $\mathrm{\Delta }{E}_{\mathrm{export}}+\mathrm{\Delta }{E}_{\text{basin-recycling}}$) for selected river basins (boundaries in dark yellow).

[Figure omitted. See PDF]

Attributing influence on $\mathrm{\Delta }Q$ to nations

Typically, TMR attributes LUC influence on $Q$ (methods described in Sect. 2.3.2) to a larger number of nations than when only basin boundaries are considered (Fig. ). In several of the studied basins (such as the Amazon, Congo, Volga, Ob, Yenisei, and Niger basins; see Fig. a, b, e, f, i, l), the share of nations contributing less than 5 % to $\mathrm{\Delta }Q$ grows considerably when TMR is considered. In some cases, nations not considered key influencers of $\mathrm{\Delta }Q$ in fact influence $\mathrm{\Delta }Q$ by more than 5 % when TMR is accounted for: in the Mekong, India is only an important influencer (10 % influence) when TMR is considered (Fig. g); in the Yenisei, Mongolia falls below 5 % influence, while Kazakhstan (11 %) and China (6 %) climb considerably in influence (Fig. i); and in the Niger basin, Sudan/South Sudan (8 %) and Niger (5 %) replace Ivory Coast and Guinea as important influencers (Fig. l). Notably, basins geographically confined within one nation can be influenced by LUC taking place in foreign nations. This is for example the case in the Yangtze, Yellow, and Huai, where irrigation in India increases the basins' $P$ (Fig. d, m, n). The TMR leads to a limited difference in nation influence only in the North American basins (Fig. h, o) and La Plata (Fig. c).

Nation influence on river flow change depending on whether TMR is taken into account in the 15 studied basins. Without considering TMR, river flow change influence originates entirely from evaporation change within the basin. With consideration of TMR, nation influence to river flow change is considered as the sum of absolute changes in precipitation import and the sum of absolute changes in evaporation export (Methods). Single country contribution below 5 % is bundled into category “Other”.

[Figure omitted. See PDF]

Discussion

Interplay between TMR and LUC

At the global scale, $\mathrm{\Delta }Q$ as a response to LUC can be almost halved by taking TMR into account (Fig. ). However, these effects vary widely by regions. While the TMR effects are negligible in some basins, remote LUC can compensate for the majority of the impact on $Q$ from local LUC in other basins (e.g. Amazon, Fig. a) and even propose new transboundary relationships (e.g. Yangtze, Fig. d). From a TMR perspective, the impact on $Q$ from within-basin LUC depends on the $\mathrm{\Delta }E$ exported from the basin as much as the $\mathrm{\Delta }P$ imported to the basin.

Our analysis shows the importance of considering LUC on par with TMR to identify anthropogenic influence on water resources, beyond analyses of pure moisture exchanges . While Africa does not constitute a major moisture source of Amazonian $P$ (7 % of all Amazon $P$, 13 % of Amazon $P$ with continental origin; see also Fig. a), the spatial extent of $\mathrm{\Delta }E$ from LUC was sufficient to elevate the relative importance of African LUC for Amazonian $\mathrm{\Delta }Q$ (28 % of Amazon $\mathrm{\Delta }P$; see also Fig. a). Similarly, India is not identified as a major moisture source of the Yangtze (see Fig. c and ), but has about 10 % influence on Yangtze $\mathrm{\Delta }Q$ (Fig. d).

Potential governance relevance

Our results indicate that both precipitationsheds and evaporationsheds of river basins are relevant governance units. Previous studies of TMR for water management have emphasized the importance of considering the $P$ source region, i.e. the precipitationshed , which was introduced as a concept analogue to watershed for water resource management. This study finds that the evaporationshed , i.e. the $E$ sink region, is just as important when considering changes to $Q$.

LUC impacts $Q$ through TMR in different ways depending on how precipitationshed, river basin, and evaporationshed are aligned. For example, where an evaporationshed has a limited overlap with river basin boundaries, reforesting a river basin may lead to unexpectedly large reductions in $Q$, if considerable deforestation simultaneously occurs in the precipitationshed outside the river basin.

The magnitude of TMR effects from remote LUC on $Q$ can be comparable to managed water flows. For example, the Yangtze River provides 36 % of the country's surface water resources, and is subject to two of the world's most ambitious water engineering projects: the Three Gorges Dam and the South-to-North Water Diversion . The overall TMR effect on mean annual LUC-induced $\mathrm{\Delta }Q$ is here estimated at 980 ${\mathrm{m}}^{\mathrm{3}}{\mathrm{s}}^{-\mathrm{1}}$ in the Yangtze basin, and the mean annual moisture change imported to the basin from foreign countries is estimated at 1110 ${\mathrm{m}}^{\mathrm{3}}{\mathrm{s}}^{-\mathrm{1}}$ (Fig. c). As a comparison, the $Q$ difference between a normal year and a dry year is about 300 ${\mathrm{m}}^{\mathrm{3}}{\mathrm{s}}^{-\mathrm{1}}$ and the total amount of water to be transferred from the Yangtze through the South-to-North Water Diversion is aimed to be 1420 ${\mathrm{m}}^{\mathrm{3}}{\mathrm{s}}^{-\mathrm{1}}$ . Seasonal and interannual flow variability is a major challenge facing the Yangtze, and future research in the seasonal LUC influence and interaction with the monsoon system is needed. Note, however, that our estimates are associated with parameter sensitivity (see Fig. S9) and large uncertainties as discussed in the Limitations.

We note that the relevance of considering TMR governance depends on future LUC. The simulated $\mathrm{\Delta }Q$ in this paper follows from a rather extreme LUC scenario (from potential to current land use). The current LUC in this study is 15 million $\mathrm{km}$${}^{\mathrm{2}}$ cropland and 28 million $\mathrm{km}$${}^{\mathrm{2}}$ pasture conversion . As a comparison, models used in the Intergovernmental Panel on Climate Change (IPCC) Fourth Assessment Report (AR4) estimated cropland changes from $-\mathrm{1.2}$ to $+\mathrm{12}$ million $\mathrm{km}$${}^{\mathrm{2}}$ between 2000 and 2050 . A more recent multi-model comparison ranged cropland conversion until 2050 from $-\mathrm{1}$ to $+\mathrm{8.5}$ million $\mathrm{km}$${}^{\mathrm{2}}$ across different scenarios . In total, the potential land for agricultural conversion has been estimated at 17 million $\mathrm{km}$${}^{\mathrm{2}}$ . Thus, future LUC can be considerable, and potential TMR impacts on $Q$ will be dependent on the type and geographical distribution of LUC, as well as dependent on prevailing winds, hydroclimate, and orography.

Limitations

In interpreting our results, it should be noted that our approach only accounts for the TMR effects. The frequency or intensity of $P$ is assumed to remain unaffected by thermal layer processes or circulation perturbation, which may introduce a bias into the quantitative estimates of hydrological flows under water-limited conditions (i.e. semi-arid regions and temperate regions during summertime) . Furthermore, vegetation response to $\mathrm{\Delta }P$ is not simulated, such as forest dieback from increased fire risk under drying conditions. Human modification of $Q$ through dams and climate change are also not considered in this study. In addition, the land-use change over land may affect the above ocean processes mainly through modification of the energy balance and circulation in monsoon regions, which we do not account for. Changes in freshwater discharge to the oceans might have implications for ocean circulation and climate, as studies of for example river discharge to the Arctic Ocean showed . However, moisture recycling's buffering effect (which mitigates river flow changes) should have a mitigating effect on the ocean's response to freshwater inflow. Otherwise, precipitation over the ocean can influence ocean salinity and precipitation patterns over land can be influenced by sea surface temperature , but we consider this outside the scope of our study and likely to be of minor importance for the research questions that we address. Our TMR analyses should, thus, be seen as an inquiry to better understand the relative importance of local and remote LUC effects on $Q$ from a water balance perspective, rather than an exact prediction. Nevertheless, due to the inevitable recycling of moisture in the global hydrological cycle, uncertainties in the magnitude are unlikely to affect our key conclusions that upwind extra-basin LUC can be essential for $Q$.

The magnitude of our estimated $\mathrm{\Delta }P$ (Fig. ) and $\mathrm{\Delta }Q$ from LUC is conservative in comparison to the literature . For example, a meta-analysis of 96 different general circulation models (GCMs) and regional climate model (RCM) deforestation simulations showed that under 10 % conversion of Amazon forest to pasture or soybean production, the inter-quartile range of rainfall change in the Amazon basin is 0 to $-\mathrm{4}$ % . In comparison, the STEAM-WAM2layers approach with change from potential to current land-use change (i.e. 8.8 % deforestation extent in the Amazon) causes a rainfall reduction of 0.4 % in the Amazon and thus falls in the conservative range. In addition, our analyses concern mean annual $\mathrm{\Delta }Q$, and can also be considered conservative in the sense that seasonal signals are expected to be much stronger.

The limitations of our methods should also been seen in the light of the strengths and limitations of alternative methods for studying hydrological LUC effects; see Table S1. The most complex and coupled modelling approaches account for the highest number of feedback processes. However, the high degree of freedom in GCMs also contributes to the high sensitivity of precipitation to initial conditions and the low signal-to-noise ratios. For example, a scenario replacing natural with present-day land cover only detected a significant response in less than 5 % of all grid cells in a single model analysis and less than 5 % in non-perturbed grid cells across seven different models . Regional deforestation scenarios generate higher ratios of significant results near the source of change, albeit noise remains high in distant regions . The challenges in simulating precipitation due to cloud formation, aerosol representation, and inherent uncertainties in circulation response , and non-closure of water balance in semi-coupled modelling approaches , also contribute to a high model dependence in estimates of river flow change from LUC . Thus, the sign, magnitude, and location of impacts vary widely among models . Observation-based methods relate presence of vegetation or irrigation to precipitation or river flows using statistical methods, often in combination with moisture tracking to determine the geographical origin of rainfall . Limitations of this type of method include variations in data quality, challenges in isolating effects of land use from climate variability, and difficulties establishing causation from correlation . Key elements missing in all approaches including our own are socio-economic dynamics and landscape resilience, which are complex issues currently explored in experimental model settings .

Future research outlook

Research of land-use change dynamics and moisture recycling is becoming increasingly detailed, looking at e.g. the role of transpiration for moisture recycling during dry periods , studying the interplay with groundwater use and recharge , and identifying influential source areas within a basin . Nevertheless, a key challenge for considering TMR effects in water governance is the modelling uncertainties and inherent variabilities associated with land–atmosphere feedback processes. The most complex modelling approaches account for the highest number of feedback processes. However, the sign, magnitude, and location of impacts vary widely even among state-of-the-art climate models . Key future improvements in climate models' ability to simulate $\mathrm{\Delta }P$ from LUC will contribute to the governability of TMR. In-depth examination of differences in model simulation of $P$ (e.g. the ongoing Precipitation Driver Response Model Intercomparison Project ) is one step in this direction. Tracking moisture in coupled climate models could further help identify causes of simulated differences in atmospheric and hydrological outputs. Key elements missing in current research on LUC effects on hydrological flows include socio-economic dynamics and landscape resilience, which are complex issues currently explored in experimental model settings .

In the meantime, “no-regret” policies in river basin management, where TMR objectives align with other aims, can potentially be explored in conjunction with LUC scenarios that include TMR effects.

Conclusions

We analysed the potential impact of human LUC on $Q$ worldwide through TMR, and separately looked at the remote and local LUC effects of relevance to water governance. Despite the river basin being the standard unit in water governance and water resource management, we find that $\mathrm{\Delta }Q$ are ultimately dependent on the modifications in both incoming $P$ and outflowing $E$. At the global scale, $\mathrm{\Delta }Q$ as a response to LUC is almost halved by taking TMR into account. Due to variations in moisture recycling patterns and LUC, the magnitude and spatial sources of the TMR effect on $\mathrm{\Delta }Q$ vary substantially among individual basins. In some basins, the remote LUC effect on $\mathrm{\Delta }Q$ exceeded local within-basin effects (e.g. in the Amazon), while in others, TMR introduced considerable foreign nation influence on $\mathrm{\Delta }Q$ (e.g. in the Yangtze). International governance arrangements of teleconnnected LUC influence could be needed, even for river basins that today are not considered transboundary. We conclude that consideration of TMR is essential for understanding $Q$ modifications and managing water resources in a rapidly changing and tele-coupled world facing increasing pressure on both land and water . Further research in both climate modelling and water governance strategies is needed to internalize land–atmosphere interactions in future water resource considerations.

The Water Accounting Model-2 layers (WAM-2layers) moisture tracking scheme in Python code can be obtained from GitHub (https://github.com/ruudvdent/WAM2layersPython, ). Earth Retrospective Analysis Interim (ERA-I) meteorological data can be obtained from the European Centre for Medium-Range Weather Forecasts (ECMWF) (http://apps.ecmwf.int/datasets/data/interim-full-daily/levtype=sfc/, European Centre for Medium-Range Weather Forecasts, 2018). The Multi-Source Weighted-Ensemble Precipitation (MSWEP) data can be downloaded from http://www.gloh2o.org/ (Beck, 2017). The Ramankutty potential land data come from https://nelson.wisc.edu/sage/data-and-models/global-potential-vegetation/index.php (Ramankutty and Foley, 2018). The current cropland and pasture map can be obtained from EarthStat (http://www.earthstat.org/data-download/, Ramankutty et al., 2018). Land Cover Type Climate Modeling Grid (CMG) MCD12C1 International Geosphere Biosphere Program (IGBP) land classification created from Terra and Aqua Moderate Resolution Imaging Spectroradiometer (MODIS) data can be downloaded at https://modis.gsfc.nasa.gov/data/dataprod/mod12.php (Friedl, 2018). Monthly irrigated rice and irrigation non-rice crops were obtained from the data set of Monthly Irrigated and Rainfed Crop Areas around the year 2000 (MIRCA2000) V1.1 and can be downloaded at http://www.uni-frankfurt.de/45218031 (Portmann et al., 2018).

The supplement related to this article is available online at: https://doi.org/10.5194/hess-22-4311-2018-supplement.

The research was conceived by LWE, RJvdE, PWK, HHGS, and LJG. IF contributed ideas for analyses. LWE carried out the model simulations, analysed the data, and wrote the paper with input from all the authors.

The authors declare that they have no conflict of interest.

Acknowledgements

We thank Victor Galaz, Chandrakant Singh, and two anonymous reviewers for providing feedback on the manuscript. Lan Wang-Erlandsson, Patrick W. Keys, and Line J. Gordon, are funded by the Swedish Research Council Formas (grant no. 1364115). Lan Wang-Erlandsson is also funded by the Japan Society for the Promotion of Science (JSPS). Ruud J. van der Ent received funding from the European Union Seventh Framework Programme (FP7/2007-2013, grant agreement no. 603608). Ingo Fetzer receives financial support from the Stordalen Foundation. Ingo Fetzer and Line J. Gordon are also supported by the Swedish foundation for strategic environmental research (MISTRA). Edited by: Stefan Uhlenbrook Reviewed by: two anonymous referees