Prominence of the tropics in the recent rise of global nitrogen pollution

Nitrogen (N) pollution is shaped by multiple processes, the combined effects of which remain uncertain, particularly in the tropics. We use a global land biosphere model to analyze historical terrestrial-freshwater N budgets, considering the effects of anthropogenic N inputs, atmospheric CO2, land use, and climate. We estimate that globally, land currently sequesters 11 (10–13)% of annual N inputs. Some river basins, however, sequester >50% of their N inputs, buffering coastal waters against eutrophication and society against greenhouse gas-induced warming. Other basins, releasing >25% more than they receive, are mostly located in the tropics, where recent deforestation, agricultural intensification, and/or exports of land N storage can create large N pollution sources. The tropics produce 56 ± 6% of global land N pollution despite covering only 34% of global land area and receiving far lower amounts of fertilizers than the extratropics. Tropical land use should thus be thoroughly considered in managing global N pollution.

results, which were simulated by using a LULUC scenario without shifting cultivation 7 . a, b, Land Nr inputs include atmospheric deposition (light orange), synthetic fertilizers (sky-blue), biological N fixation (BNF) in non-agricultural (plum) and agricultural (purple) lands. c, d, Land N outputs include river dissolved organic N (DON) exports (green), river dissolved inorganic N (DIN) exports (blue), soil and freshwater denitrification (light green), fire emissions (orange), net harvest in agricultural (red) and non-agricultural (brown) lands. e, f, N fluxes to land storage. All plots show 30-year moving averages from 1750 to 2005. Figure 8. Land N fluxes in the tropics and extratropics. This figure shows results, which were simulated by using fertilizer inputs from Lu and Tian 6 . a, b, Land Nr inputs include atmospheric deposition (light orange), synthetic fertilizers (sky-blue), biological N fixation (BNF) in nonagricultural (plum) and agricultural (purple) lands. c, d, Land N outputs include river dissolved organic N (DON) exports (green), river dissolved inorganic N (DIN) exports (blue), soil and freshwater denitrification (light green), fire emissions (orange), net harvest in agricultural (red) and non-agricultural (brown) lands. e, f, N fluxes to land storage. All plots show 30-year moving averages from 1750 to 2005. Figure 9. Historical human land-use changes. Thirty-year moving averages of historical (1750-2005) land-use fraction changes for the Mackenzie (blue), Colorado (blue-green), Amazon (green), Parana (yellow), and Mississippi (red) River Basins. a, Agricultural land use (i.e., cropland and pasture); b, All land use disturbed by human activities (i.e., agricultural land use and secondary land useabandoned agricultural land or regrowing forest after logging). Figure 10. Global distributions of historical human land use. Contemporary (1976-2005 mean) basin-wide mean fraction of human land use for 159 globally-distributed major river basins. a, Agricultural land use (cropland and pasture); b, All land use disturbed by human activities (i.e., agricultural land use and secondary land use (abandoned agricultural land or regrowing forest after logging)). Figure 11. A long-term increase in the aboveground biomass density in Amazonian intact forests. Our simulation shows positive net biomass changes in Amazonian intact forests during 1983-2005 (blue solid line), which are consistent with corresponding uncertainty bounds (shaded orange area) reported in Brienen and colleagues 8 .  Table 2 in Galloway and colleagues 4 ) plus N2O emissions from soils and rivers (See Table 3 in Galloway and colleagues 4 ). G NOx and NH3 emissions from agriculture, biomass and biofuel burning, and soils under natural vegetation plus N2O emissions from agriculture, biomass and biofuel burning, human excreta, and soils under natural vegetation (See Table 6  For fertilizers, manure, and urban wastewater, reported input fractions for the Chesapeake Bay Program's Watershed Model were adopted 10 . We did not conduct sensitivity tests for urban wastewaters, because their amount is very small compared to the other N sources. Sensitivity tests show that different fractions (i.e., -25% and +25% fractions of the dominant N species for each N source) have almost no influence on land N storage and fluxes (Supplementary Figure 12).

Supplementary Note 1. Comparison of simulated global land nitrogen (N) budgets with published estimates.
We compare simulated global land N storage and fluxes with published estimates in 16 different studies [4][5][6][11][12][13][14][15][16][17][18][19][20][21][22][23] (Supplementary Table 1). We focus on agreement during the contemporary period, which is of primary interest to the results of this paper and for which more data is available. While published N storage and flux estimates are invaluable in identifying and understanding the dominant global N cycling processes, the authors of these studies also acknowledge considerable uncertainties due to sparse measurements and subsequent assumptions required to estimate the global magnitude of different N cycling processes. We thus emphasize consistency in the magnitude and direction of N storage and fluxes.
Simulated contemporary BNF in our baseline simulation (128 TgN yr -1 ) is between 112 TgN yr -1 estimated by Green and colleagues 5 and 139 TgN yr -1 estimated by Galloway and colleagues 4 (Supplementary Table 1a). Simulated preindustrial BNF (76 TgN yr -1 ) is well within the latest estimate of 58 (40-100) TgN yr -1 by Vitousek and colleagues 12 . Simulated BNF in agricultural lands (cropland and pasture; 69 TgN yr -1 ) is within the latest BNF estimate in agricultural systems (50-70 TgN yr -1 ) by Herridge and colleagues 11 . Our estimate lies on the high end of Herridge and colleagues 11 's range, in part, because the agricultural land area simulated by using the LULUC scenarios of Hurtt and colleagues 7 (49 10 6 km 2 ) is larger than the area considered to calculate the Herridge and colleagues 11 's estimate (27 10 6 km 2 ). See Methods for a land-use description.
Simulated BNF in non-agricultural (primary and secondary) lands (59 TgN yr -1 ) is smaller than published BNF in natural systems by Galloway and colleagues 4 (107 TgN yr -1 ) and by Cleveland and colleagues 14 (128 TgN yr -1 ). In contrast to agricultural BNF, this is partly because the non-agricultural land area derived from Hurtt and colleagues 7 's LULUC scenarios (82 10 6 km 2 ) is smaller than the area considered to calculate the Cleveland and colleagues 14 's estimate (104 10 6 km 2 ). Furthermore, we note the suggestion by Vitousek and colleagues 12 that contemporary natural BNF is likely lower than their preindustrial estimate 58 (40-100) TgN yr -1 due to land conversion for cultivation and perhaps to downregulation of BNF under increasing anthropogenic Nr inputs. This implies that the estimates of Galloway and colleagues 4 and Cleveland and colleagues 14 may be high.
The range of published BNF estimates highlights the substantial uncertainty in this important element of the N budgets. As noted in the main text, we thus conducted uncertainty tests of low and high contemporary BNF (116 and 145 TgN yr -1 , see Supplementary Table 3). The results suggest that the different BNF settings have little effect on the patterns of global, extratropical, and tropical land N fluxes and pollution (Figures 3a and 4, Table 1, Supplementary Figures 3, 5, and 6).
Moving on to atmospheric N deposition, Galloway and colleagues 4 estimated 59 TgN yr -1 as the sum of atmospheric NOy, (25 TgN yr -1 ) and NHx (34 TgN yr -1 ) deposition to land, while Green and colleagues 5 estimated 32 TgN yr -1 (Supplementary Table 1b). We used the Green and colleagues 4 's estimate of atmospheric deposition.
For fertilizer, we applied estimates from Bouwman and colleagues 15 (114 TgN yr -1 ) that are between the Haber-Bosch estimate by Galloway and colleagues 4 (100 TgN yr -1 for 1990s) and that by Galloway and colleagues 16 (120 TgN yr -1 for 2000s) (Supplementary Table 1c). We also conducted an uncertainty test of different fertilizer by Lu and Tian 6 . Applications of the both fertilizers show higher land N sequestration in the extratropics than in the tropics (Figure 3a) and create similar global, extratropical, and tropical land N pollution and fluxes (Figures 3a and 4, Table 1, Supplementary Figures 3 and 8).  Table 1d). We further evaluate N fluxes to the ocean by comparing simulated regional river dissolved inorganic and organic N loads and concentrations with measurement-based estimates from 47 major rivers, which are distributed broadly over the globe and influenced by various climates, biomes, and human activities 1-3 ( Supplementary Figures 1 and 2, Supplementary Table 2). The correlations between the simulated and reported estimates of river discharge, N loads and concentrations all fall between 0.74 and 0.91.
Atmospheric N emissions in LM3-TAN arise from simulated denitrification (on land and within rivers and lakes), fire emissions, and harvested material, which is presumed to be primarily a precursor to atmospheric N pollution via various pathways including wood, biofuel, and waste burning, livestock respiration, emissions from food, human, and livestock waste 4,15,20 . As described in the main text and Methods, the portion of the harvest associated with atmospheric emissions is calculated as the harvest remaining after subtracting manure applications, wastewater discharges to rivers, and the fraction of the harvest used for durable goods (e.g., home building). Manure applications and wastewater discharges are specified model forcings (See Methods, Bouwman and colleagues 15 Table 1g).
Simulated N fluxes to the land storage is 32 TgN yr -1 (Supplementary Table 1h), which is similar in magnitude though less than 60 TgN yr -1 estimated by Galloway and colleagues 4 and consistent with estimates by other terrestrial ecosystem models (e.g., 27 TgN yr -1 ; Zaehle 22 ). Galloway and colleagues 4 and Gruber and Galloway 31 acknowledged significant uncertainty in their global estimate and the gap between our estimate and theirs could be easily closed by higher total land Nr inputs of Galloway and colleagues 4 (+22 Tg N yr -1 ). Simulated global soils/litter N storage (86124 TgN) is within reported estimates (70000-820000 TgN) by Post and colleagues 23 and references in Post and colleagues 23 (Supplementary Table 1i).

Supplementary Note 2. Estimation of total land N pollution.
Directly simulated N compartments and fluxes in LM3-TAN provide closed land N budgets and estimates of the total N fluxes from land to the atmosphere and ocean. Not all N fluxes, however, are harmful. N2 is generally considered benign, as is N sequestered into durable goods (e.g., home building). It has furthermore been suggested that organic N exports from rivers to the coastal ocean are less harmful pollutants than inorganic N exports due to their relative long-lived nature 32 . This is, however, likely to only apply to acute local impacts at river mouths and does not preclude broader impacts on continental shelf and ocean scales. Total land N pollution for our study was thus estimated as total land N outputs, minus the sum of N2 emissions and human appropriation of the net harvest into durable goods.
To estimate total land N pollution, we did three additional partitioning of the fluxes directly simulated by LM3-TAN: 1) the partitioning of the soil and freshwater denitrification into N2O and N2 emissions, 2) the partitioning of the net harvest into N2 emissions, and 3) the partitioning of the net harvest transformed into durable goods. To test the robustness of our results to uncertainty in these partitions, an interval for each partition was assigned based on the scientific literature and 1000 Monte Carlo style calculations were conducted with random draws from a uniform distribution across the uncertainty interval. This was done for the baseline simulation, and for the 4 sensitivity simulations with different BNF, fertilizer inputs, and LULUC. Lastly, we created additional 1000 different total land N pollution estimates by excluding river organic N exports. These provided a total of 6000 permutations.
A N2O fraction of denitrification emissions can vary significantly in different climate, land use, and time (See Supplementary Materials in Bai and colleagues 33 ). Global budgets 4,[19][20]22 , however, suggest large scale characteristic values ~0.08-0.11. For a N2O fraction of soil and freshwater denitrification, we assigned different intervals: (0, 0.2) for global land and (0, 0.3) for tropical land. The higher upper bound for tropical land was because natural tropical systems have been recognized as a major hotspot of N2O emissions [33][34] .
A primary source of N2 emissions associated with our net harvest are emissions from manure storage systems and wastewater treatment plants. Bouwman and colleagues 19 estimated denitrification (N2O+N2) emissions from manure storage systems and wastewater treatment plants as 8 TgN yr -1 , which accounts for a 0.09 fraction of our net harvest. For fractions of the net harvest into N2 emissions, we assigned an interval (0, 0.2) for both global and tropical lands. We note that N2O emissions from these sectors were assumed to be minor relative to N2 emissions, based on the literature 35 (Figure 3a), suggests that tropical land as a whole (including all kind of land use and land cover, such as agricultural lands, intact and disturbed forests) is nearly N neutral. This result aligns with filtered inverse models against an additional observational constraint, suggesting nearly neutral net C fluxes from tropical land for the same period (1992)(1993)(1994)(1995)(1996) 36 . A recent study of plot measurements in Amazonian intact forests demonstrated a long-term increase in the aboveground biomass density since 1983 (~0-3 Mg ha -1 yr -1 ) 8 , and a similar pattern during 1983-2005 was captured in our simulation (Supplementary Figure 11). For the same period , however, our simulation suggests that tropical forests as a whole (including both intact and disturbed forests) are a net C source of 271 TgC yr -1 , based on changes in aboveground C storage. This result appears to be consistent with a recent satellite-data-based study 37 , demonstrating a net C source of 425 TgC yr -1 from tropical forests during 2003-2014. However, we note that this comparison cannot be done more formally, because of unconsidered terms in Baccini and colleagues 37 ' approach, such as herbaceous and nonwoody vegetation, and because our simulations do not span the entire Baccini and colleagues 37 ' period. Analyses of our simulations were limited up to 2005, because the used CMIP5 dataset for land-use changes stops in 2005 7 .
Supplementary Note 4. Lake N cycle. Lakes receive dissolved organic N, ammonium, and nitrate plus nitrite from river inflows, and lose those by outflows to rivers and denitrification. Microbial processes in lakes are simulated by first-order loss function with respect to lake N content and with an adjustment for the influence of lake water temperature. Reported nonlinear regression function based on Lotic Intersite Nitrogen experiment reach-scale measurements 28,38-39 was adopted to estimate reaction rate constants of lake denitrification. The lowest measured value (0.034 day −1 ) was set as a minimum reaction rate constant of lake denitrification (Supplementary Table 3). A maximum reaction rate constant of lake denitrification (0.05 day −1 ) was calibrated to match reported and simulated river N exports. Previously used reaction rate constants of river mineralization and nitrification and river temperature reduction function 24 were also used for those of lakes. )}} Supplementary Eq. (5) where is , 4 + , and 3 − ; is lake N content (kg m -2 ); and are river inflows to lakes and river outflows from lakes (kg m -2 s -1 ); ′′ is lake temperature reduction function; ′′ is a parameter; ′′ is a reference lake temperature (℃); ′′ is lake water temperature (℃); ′′ , ′′ , and ′′ are lake mineralization, nitrification, and denitrification (s -1 ); , ′′ and , ′′ are minimum and maximum reaction rate constants of lake denitrification (s -1 ); 3 − is lake nitrate-N concentration (μmol L −1 ); is lake depth (m); 0 , 1 , and 2 are constants; is a log re-transform bias correction factor; , is a unit-conversion constant (day s −1 ).