The fertilizer P application maps are developed by allocating country-level inventory data into grid cells according to land use maps, crop rotation maps, and crop-specific fertilizer application data (Table 1). Annual national P fertilizer inputs on agricultural land during 1961–2020 were obtained from FAOSTAT (2024). We first separated total P fertilizer use into P application on cropland and pasture by assuming fractions of P used for cropland and pasture are the same as those of N fertilizer, following the method used in Zou et al. (2022). The annual fractions of fertilizer use allocated to cropland and pasture for major countries from 1961 to 2019 were adopted from Einarsson et al. (2021) and Lassaletta et al. (2014). Then, the country-level P fertilizer data were spatially distributed with cropland and pasture maps following the workflow in Fig. 1.

To determine the P fertilizer application on cropland, we considered 17 dominant crop types (wheat, maize, rice, barley, millet, sorghum, soybean, sunflower, potato, cassava, sugarcane, sugar beet, oil palm, rapeseed, groundnut, cotton, and rye) and crop-specific harvested areas to generate crop-area-weighted average of P fertilizer application in each grid during 1961–2020, which is calculated as follows: 1Fcrop,y,g‾=∑iFcrop,i,2000×AHi,y∑iAHi,y where Fcrop,y,g‾ is the crop-area-weighted average of P fertilizer application rate in a cropland grid g (g P m⁻² cropland per year) in year y, Fcrop,i,y and AH_i,y are P fertilizer application rate (g P m⁻²) and harvested area (m²), respectively, for crop type i in year y. Fcrop,i,2000 used a fixed value derived from the global crop-specific P fertilizer application data circa 2000 (at a resolution of 5 arcmin) developed by Mueller et al. (2012). This dataset was used as the spatial baseline for distributing fertilizer inputs across cropland. The crop distribution maps for individual years were derived by allocating FAO statistics for crop-specific harvested area into grids using the Dissever algorithm (Malone et al., 2012), which is a general spatial downscaling method using fine-resolution covariate data. Specifically, the covariate data included climate variables (temperature and precipitation) (Fick and Hijmans, 2017), soil information (FAO and IIASA, 2023), and topography. In each year, crop-specific distribution from SPAM2010 (Yu et al., 2020) was used as initial inputs for starting an iterative process to obtain an optimal machine learning model within a region in each year. Though this climate-based disaggregation strategy for statistical analysis inevitably brings uncertainties, it has usually been adopted in generating crop distribution maps (e.g., Li et al., 2025; Yu et al., 2020). In practice, to minimize these uncertainties, the machine learning-based distribution maps are adjusted to meet multiple constraints. For example, the sum of areas of all crops in a grid should not exceed the actual arable land area (see our cropland data generation using HYDE3.3 and LUHv2 in text below), and the area of a specific farming system should not exceed the suitable area within the grid (adopted from SPAM2010).

Annual national P fertilizer consumption from FAOSTAT was subsequently used to scale the baseline application rates, allowing fertilizer inputs to vary over time: 2Rcrop,y,j=FAOcrop,y,j∑g=1g=nincountryjFcrop,y,g‾×ACy,j,g3Pfercrop,y,j,g=Fcrop,y,g‾×Rcrop,y,j where Rcrop,y,j represents the regulation ratio (unitless) for the year y and country j, FAO_crop,y,j is the national total P fertilizer usage (g P yr⁻¹) on cropland, AC_y,j,g is the cropland area (m²) for the year y and grid g of the country. The historical land use data (cropland, pasture, rangeland) covers 1860–2020 and has an annual time-step and a 5 arcmin spatial resolution, developed by reconciling the HYDE 3.3 (https://landuse.sites.uu.nl/datasets/, last access: 25 October 2024) and LUHv2 datasets (Hurtt et al., 2020; Klein Goldewijk et al., 2017; Tian et al., 2022). For the period prior to 1949, HYDE 3.3 provides land use maps only at decadal intervals (i.e., one map every ten years). To generate continuous annual data, we applied a temporal downscaling approach: the LUHv2 data for each year within a given decade were used to proportionally interpolate the HYDE 3.3 map from the beginning of that decade. This method preserves the decadal baseline provided by HYDE 3.3 while incorporating the interannual variability captured by LUHv2. The P fertilizer in cropland grid g of the country j in year y, Pfer_crop (g P m⁻² yr⁻¹), was then calculated as the product of Fcrop,y,g and Rcrop,y,j.

The earliest year of P fertilizer statistics that the FAO database provides is 1961. Prior to the widespread adoption of phosphate rock as a fertilizer in the 1940s, guano – accumulated bird droppings over millennia – and human excreta were utilized as fertilizers for food crops. The historical P fertilizer data during 1860–1960 were generated based on data reported by Cordell et al. (2009), which provide global total P fertilizer usage estimates back to 1800. Annual relative change rates of global P fertilizer were first calculated based on the estimates from Cordell et al. (2009). These change rates were then applied to the spatially explicit P fertilizer distribution in the reference year of 1961 to reconstruct annual global P fertilizer maps for the period 1860–1960. This approach assumes that temporal dynamics of P fertilizer data before 1961 were primarily driven by global-scale trends.

The grid-level manure P inputs, including manure application on cropland, manure application on pasture, manure deposition on pasture, and manure deposition on rangeland, were generated by multiplying the corresponding manure N inputs from HaNi (Tian et al., 2022) and manure P : N ratios (Fig. 2): 4Pmany,g=Nmany,g×RPNy,g where Pman_y,g and Nman_y,g indicate the P and N manure application or deposition on grid g in year y, respectively. The HaNi dataset offers global manure input data on cropland, pasture, and rangeland during 1860–2020 at a resolution of 5 arcmin. RPN_y,g represents the manure P : N ratio in grid g. 5RPNy,g=∑tAPNy,t×Anumy,g,t∑tAnumy,g,t6Anumy,g,t=AGLWnumg,tGLMnumg,t×LFAOy,t,j∑g=1g=nincountryjGLMnumg,t The annual grid-level manure P : N ratios RPN_y,g were estimated as the animal-number weighted sum of these animal-specific P : N ratios within grids. We obtained manure P : N ratio APN_y,t for each animal species t from Lun et al. (2018). Livestock distribution data Anum_y,g,t were derived by fusing the annual gridded livestock of the world (AGLW) dataset (Du et al., 2025) and a newly downscaled product. The AGLW dataset covers 1961–2021 at a 5 km resolution and was developed based on annual FAOSTAT census information using a machine learning method. But in some areas (particularly the former Soviet Union region), livestock information is missing. To fill the gaps, we used the national FAO livestock statistics and the Global Livestock of the World 3 (GLW3) database (Gilbert et al., 2018) to generate a long time-series gridded livestock distribution dataset. Specifically, GLW3 provides global livestock populations GLMnum_g,t at a 5 arcmin resolution for cattle, buffaloes, horses, sheep, goats, pigs, chickens, and ducks, but the data are available for a single reference year. Therefore, the annual country-level livestock population data LFAO_y,t,j from FAOSTAT were employed to extend the GLW3 dataset into a time series of livestock distribution maps spanning from 1961 to 2020 following the method used in Eqs. (2) and (3). Based on the fused product of livestock distribution, a spatially explicit manure P : N ratio dataset covering the period 1961–2020 was constructed and subsequently used to estimate manure P inputs from corresponding N inputs. The final nutrient ratio dataset is highly consistent both spatially and temporally (Fig. S1 in the Supplement). Notably, although temporal changes in livestock composition are considered when estimating manure P : N ratios, potential changes in species-specific manure nutrient ratios associated with evolving feed composition and livestock management practices are not explicitly represented, which may introduce biases in manure P estimates.

Atmospheric P deposition originates from both natural and anthropogenic sources and was estimated by combining a P emission inventory, an atmospheric chemical model, and an algorithm (Fig. 3 and Table 1) (You et al., 2026). Natural P emissions primarily comprise dust, sea salt, Primary Biological Aerosol Particles (PBAPs), volcanoes, and wildfires. In this study, emissions from dust, sea salt, and volcanoes were derived from the GEOS-Chem HEMCO data (Meng et al., 2021; Weng et al., 2020; Carn, 2019; Carn et al., 2015; Ge et al., 2016). PBAPs emissions were estimated by utilizing specific humidity data and LAI data according to the method used in Myriokefalitakis et al. (2016). Wildfire emission data were sourced from the GFED4 dataset (Randerson et al., 2015). Meanwhile, anthropogenic emissions were estimated using monthly sector-specific data from the EDGAR database (Crippa et al., 2023). By integrating all the emission sources, we developed a global atmospheric P emission inventory, which was subsequently used as input for the GEOS-Chem model (Feng et al., 2021) to simulate global atmospheric P deposition from 2001 to 2019. GEOS-Chem is a global three-dimensional atmospheric chemical model driven by meteorological inputs provided by the Goddard Earth Observing System (GEOS) of the NASA Global Modeling and Assimilation Office. It has been widely applied to address various atmospheric composition issues. The modeled P deposition by GEOS-Chem was at a resolution of 2.5°×2°.

To obtain the atmospheric P deposition at a high resolution, the dissever algorithm (Malone et al., 2012; Roudier et al., 2017), which is a mass-conserving spatial downscaling method, was applied to the GEOS-Chem outputs. First, the coarse-resolution P deposition data were resampled to 0.1°×0.1°; Second, the dissever algorithm was used to iteratively train an XGBoost model for eight different regions, including Africa, Asia excluding China and India, China, Europe, India, North America, Oceania, and South America, to retrieve resampled P deposition and fine-resolution environmental variables. These environmental variables included ERA5 climate factors (wind components, temperature, surface net solar radiation, surface pressure, and monthly total precipitation) (Muñoz-Sabater et al., 2021), LAI (Cao et al., 2023), Aerosol Optical Thickness (AOT, Randles et al., 2017), and P emission data. The iterative model training process was terminated when the improvement in performance between successive generations fell below a predefined threshold. Model performance was assessed using the Mean Absolute Error (MAE; unit: g ha⁻¹), calculated for each model's target region across all terrestrial grid cells using the test dataset. The convergence threshold for MAE was initially set to 0.001 g ha⁻¹. Using this approach, we generated global atmospheric P deposition estimates at a spatial resolution of 0.1° for the period 2000–2019. To ensure consistency with the HaPi dataset, the 0.1° data were subsequently resampled to a 5 arcmin resolution and temporally extended back to 1900 using annual P deposition change rates derived from Ringeval et al. (2024). P deposition for the period 1860–1899 was assumed to remain constant and equal to the early 20th century levels.

Global anthropogenic P inputs have undergone dramatic changes since the Industrial Revolution, increasing from 3.8 Tg P yr⁻¹ in the 1860s to 41.0 Tg P yr⁻¹ in the 2010s, a nearly 11-fold increase (Fig. 4). The most rapid acceleration occurred between 1945 and 1989, coinciding with post-war agricultural intensification. The amplified usage of mineral fertilizer and livestock manure P contributed nearly equally, over the study period, i.e., 52 % and 46 % of the increase in TP inputs to the terrestrial biosphere, respectively. The composition of P inputs shifted significantly over this period (Table 2). Prior to the 1980s, livestock manure accounted for over half of total P (TP) inputs. However, by the 2010s, mineral fertilizer P in TP inputs surpassed manure P for the first time. Atmospheric P deposition, while relatively stable, declined from 50 % to only 7 % of TP inputs from the 1860s to the 2010s, due to the marked increase in anthropogenic inputs. The intensified application of P fertilizer on cropland was the primary driver behind the increased TP inputs after 1945.

Regionally, the TP inputs initially increased in Europe and eastern USA between the 1860s and the 1910s, primarily due to enhanced manure inputs (Figs. 5 and 6). Subsequently, P input hotspots emerged in Europe, driven by the rising use of P fertilizer. Europe (5.2 Tg P yr⁻¹) and the USA (2.8 Tg P yr⁻¹) were the top two regions with the highest P inputs in the 1960s. After the 1980s, TP inputs in Europe declined quickly, while leveling off in the USA. By the 2010s, new hotspots emerged in eastern China, northern India, southern Brazil, and eastern Africa, sourcing from the widespread application of mineral P fertilizer and the expansion of livestock production in these regions. Notably, TP inputs in China began to decrease around 2010, whereas inputs in South Asia and Brazil maintained growth trends. Livestock manure remains the dominant P source in most regions of the South Hemisphere (Africa, South America, Oceania), while P fertilizer plays a more important role in industrialized regions of the North Hemisphere (Asia, Europe, North America). In addition to Europe, TP inputs in Russia, Korea, and Japan (KAJ), and Central Asia (CAS) also declined substantially, largely due to the reduction in P fertilizer usage. Manure P inputs continued to increase in Africa, South/Southeast Asia, and Central America (CAM), but have shifted to a decreasing trend in Oceania, Europe, and Russia in recent decades.

The global total P fertilizer inputs surged from 6.6 Tg P yr⁻¹ in the 1960s to 19.4 Tg P yr⁻¹ in the 2010s, with cropland receiving over 90 % of these inputs (Fig. 7). Annual P fertilizer applied to croplands increased rapidly at a rate of 0.4 Tg yr⁻² during 1961–1989, then declined sharply until 1995 before resuming an upward trajectory. Despite this global growth, significant regional disparities emerged (Figs. 6 and 8). Europe and the USA were the dominant regions for early P fertilizer consumption, accounting for 38 % and 24 % of global usage, respectively, in the 1960s. By the 2010s, China (30 %) and South Asia (20 %) had become the predominant consumers of P fertilizers, driven by intensive agricultural expansion. Regional shifts in P fertilizer inputs were pronounced. In the USA, P fertilizer application to cropland decreased from 2.0 Tg P yr⁻¹ in the 1970s to 1.5 Tg P yr⁻¹ in the 2000s. In Europe, P fertilizer usage peaked in the 1980s but decreased by 66 % in the 2010s. The increase in P fertilizer usage in China and South Asia contributed to 44 % and 29 % of the global total increase in P fertilizer application on cropland from the 1960s to the 2010s, respectively. However, China's usage plateaued in the 2010s, while South Asia's continued to increase rapidly. In the 2020s, hotspot regions with high P fertilizer application rates (>3 g P m⁻² yr⁻¹) were mainly distributed in eastern China, northern India, and southern Brazil. The P fertilizer application rates on cropland were consistently low across most areas in Africa over the whole study period, accounting for only 4 % of global consumption in the recent decade.

Global P fertilizer application on pasture increased from 0.3 to 1.6 Tg P yr⁻¹ during 1961–2020. Before the 1980s, European countries were the predominant consumers of P fertilizer on pasture (Fig. S2). Thereafter, the USA and India significantly increased their usage, accounting for 23 % and 22 % of total P fertilizer application on pasture, respectively, in the 2010s. P fertilizer application rate per pasture land was high in India, Japan, and southern Canada, but remained low in most other countries in recent decades.

Livestock manure served as the primary P source for agricultural soils historically, with global manure P inputs on cropland, pasture, and rangeland increasing from 1.7 to 18.9 Tg P yr⁻¹ during the period from the 1860s to the 2010s (Fig. 9). Pasture received half of total manure P inputs, while cropland and rangeland each shared a similar proportion of the remaining half. Over 1961–2020, manure P deposition on pasture kept increasing at a rate of 0.07 Tg P yr⁻², while manure P application on pasture was relatively stable. Meanwhile, annual manure inputs on cropland and rangeland both increased, but at relatively slow rates than those on pasture. In the context of rising global manure P inputs, developing countries have demonstrated stronger growth rates than developed countries in recent decades. China has surpassed Europe as the region receiving the largest manure P inputs since the 2000s, as manure P inputs in Europe have declined since the 1980s. Over the past four decades, manure usage in North Africa (NAF) and Equatorial Africa (EQAF) has grown at the fastest rate, increasing by 96 % and 138 %, respectively. In the 2010s, China, South Asia, Brazil, and North Africa, as the top four regions, received 13 %, 11 %, 11 %, and 10 % of the global total manure P inputs, respectively.

Application rates of manure P per unit area of cropland increased significantly in Asia, Europe, and North and South America since 1860 (Fig. 10). Manure P application on cropland initially increased in western Europe in the 1910s, with subsequent intensified application occurring in eastern Asia and northern South America by the 2010s. Manure application and deposition rates on pasture were extremely high in South and Southeast Asia over the last century (Figs, 11, S3, and S4). The proportion of manure deposition in total manure inputs to pasture gradually increased from 79 % to 92 % from the 1860s to the 2010s. Prominent manure P deposition on rangeland was observed in South and Southeast Asia, with new hotspots developing in Central Africa, eastern South America, northern China, and Europe in the 2010s (Fig. 12).

Atmospheric P deposition is the major P source for natural ecosystems, such as forests and shrubs, but plays a less important role in cropland. Although anthropogenic activities have led to the increase of atmospheric P deposition from 2.0 Tg P yr⁻¹ in the 1960s to 2.7 Tg P yr⁻¹ in the 2010s, the proportion of atmospheric P deposition in total P inputs decreased dramatically to only 7 % by the 2010s. China was the region with the largest atmospheric P deposition (21 % of the global total) in the 2010s, followed by North Africa (15 %), the USA (11 %), and Europe (9 %) (Fig. 13). Atmospheric P deposition in northern Africa was primarily derived from natural sources and remained relatively stable during the study period. Influenced by anthropogenic activities, atmospheric P deposition in China has continually increased since the 1960s, while that in Europe has shifted from an increase to a decrease in the 1980s. Atmospheric P deposition was relatively low across most regions, with the exception of several notable hotspots in northern Africa, eastern China, central Europe, and eastern USA.

The HaPi dataset provides spatially explicit, annual estimates of anthropogenic P inputs to the terrestrial biosphere. This gridded P input dataset reveals subnational heterogeneity, enabling the detection of localized hotspots of high P input that national averages may obscure. When aggregated to the global scale, HaPi aligns well with existing studies, particularly for cropland fertilizer P inputs, which have been extensively evaluated (Table 3). Nearly all these studies rely on the country-level fertilizer inventory data provided by FAO or IFA, which give consistent global totals. Additional validation against independent national statistics from China and the US shows that HaPi captures comparable magnitudes and similar spatial-temporal patterns of P fertilizer use (Fig. S5), although with slightly lower estimates relative to both datasets. These results provide further support for the robustness of national-scale temporal dynamics in HaPi.

In contrast, the HaPi estimates of manure P applied to cropland are consistent with those reported by Zou et al. (2022) but are lower than other estimates, while manure P inputs in grasslands (pasture and rangeland) are higher than previous estimates. Despite these differences, total manure P inputs in HaPi remain consistent with other studies. The primary source of discrepancy arises from differences in the allocation of manure inputs between cropland and grassland and is partly driven by inconsistencies in cropland definitions, especially regarding the inclusion of temporary pastures. In previous studies, manure P inputs were calculated by multiplying manure N inputs provided by FAO and P : N ratio in livestock manure products. According to the latest FAOSTAT dataset, around 75 % of total manure N inputs are deposited on grassland, and the remaining 25 % is applied to cropland (around 21 %) and grassland (4 %) soils. The majority of manure is directly deposited on grasslands through grazing animal excretion, whereas a smaller proportion is collected and subsequently applied to pastures and croplands as a managed nutrient input. HaPi estimates 23 % of manure P was applied on cropland, closely matching FAOSTAT's current methodology. In contrast, Lun et al. (2018) and Sattari et al. (2016) estimated 36 % and 38 % of total manure P inputs to cropland.

Compared to previous studies, the HaPi dataset offers the most comprehensive coverage of anthropogenic P input fluxes, featured by a high spatial resolution and extended historical coverage. These features support the analysis of legacy P accumulation and depletion, and simultaneously provide consistent forcing data for land surface and biogeochemical models. This harmonized P input dataset enhances our ability to assess regional nutrient trends, environmental risks, and management needs at multiple spatial scales.

Globally, mineral P fertilizer has surpassed manure P as the largest P input to cropland, while livestock manure remains the dominant P source for pasture and rangeland. Overall, livestock manure is a more significant source of P for terrestrial ecosystems compared to mineral fertilizer. In recent decades, livestock manure contributed over half of the P sources in most regions, especially in Africa where around 90 % of total P inputs were derived from manure. As P fertilizer application began to decrease in European countries, manure P application on cropland played an increasingly important role in food production. Grassland, including pasture and rangeland, received over 70 % of global manure P inputs and over 40 % of total P inputs; therefore, it is critical to take account of grassland when investigating global P balance and P-related environmental issues. For instance, it is estimated that P loading to the Gulf of Mexico originated primarily from manure on pasture and rangeland (37 %), followed by corn and soybeans (25 %) within the Mississippi River Basin (Alexander et al., 2008). Despite high manure P input, manure excretion is still partly an internal P cycling in the grassland-livestock system since it derives from soil P uptake by grass. Furthermore, global manure P input cannot compensate for the grazing P output in grassland systems because around 23 % of manure was transferred from grassland to cropland systems. Due to the limited application of fertilizer to grasslands, global grasslands are facing the challenge of a negative P budget (Sattari et al., 2016). Given the projected increase in livestock production to meet future demand for meat and milk, effectively managing P flows in the whole crop-livestock production systems is critical for the sustainable human P cycle (Bouwman et al., 2013).

Socio-economic development clearly drives the long-term changes in total anthropogenic P inputs to agricultural lands (Fig. S6). The rapid increase in P fertilizer application on cropland in the latter half of the 20th century stimulated crop production but ultimately resulted in a decrease in P use efficiency. After about 60 years of growth, global total P fertilizer use has leveled off in the 2010s. Europe, previously the largest consumer of P fertilizer, began reducing its usage in the 1980s. China, currently the largest consumer, has also shown a decrease in P fertilizer consumption since 2013. Although P fertilizer application decreases, crop yield and P uptake may not decline correspondingly. The large P surpluses due to previous P application in European countries have built up soil residual P pool which can continually supply crop production (Sattari et al., 2012). Similarly, the reduced usage of P fertilizer in China may also result from the supply of soil legacy P resources accumulated over the last decades. The increased crop yield and decreased P fertilizer usage indicated an enhancement in P use efficiency in these industrialized countries. Meanwhile, many developing countries, such as India and Brazil, continue to experience elevated P fertilizer usage and an amplified P surplus on croplands (Zhang et al., 2017b). Utilizing residual soil P can be a key strategy to reduce reliance on imported mineral P fertilizer and improve the sustainability of agriculture in these countries.

Despite numerous improvements in the HaPi dataset over previous P input datasets, several limitations remain in its development. First, the fertilizer input dataset was developed using the spatial distribution of crop-specific fertilizer application rates around the year 2000 as a baseline map. Although fertilizer application rates vary over time through scaling with national statistics, the fixed spatial pattern of fertilizer use would disregard subnational variability in fertilizer application rates for each crop, and HaPi may not fully capture temporal changes in relative fertilizer intensity among individual crops. Furthermore, the climate-based disaggregation strategy for FAOSTAT crop harvest area needs further improvement. A key challenge is the lack of globally consistent, dynamic maps (e.g., yearly suitable area for specific farming systems) for constraining spatial allocation of statistical data. To improve the crop-specific harvest area estimates, more efforts are required to obtain fine-grained statistics (national- or provincial-level statistical data is used in this study) of varied crops. Another plausible solution is to fuse satellite-retrieved crop distribution products with these statistical disaggregation maps. But the major limits of this pathway are the discontinuous temporal coverage and lack of global availability for many crop types in satellite products.

Second, we used historical cropland, pasture, and rangeland data from HYDE/LUHv2 to spatialize country-level P fertilizer use amounts, but HYDE/LUHv2 data have been shown to exhibit inconsistent spatial and temporal patterns of land use relative to satellite-derived land use at the regional scale. For example, the HYDE dataset overestimates the cropland area in India, which can lead to the underestimation of the P fertilizer application rate on cropland. In this study, we assumed that the ratios of P fertilizer application on pasture relative to cropland were the same as those for N fertilizer, which may not accurately reflect the actual allocation of P fertilizer usage on pasture. Since country-level data for P fertilizer are only available from 1961 onwards, we assumed that the change rates of global fertilizer inputs before 1961 followed the annual global trends reported by Cordell et al. (2009). Although this approach provides a temporally continuous long-term dataset, it cannot explicitly represent historical changes in crop composition, fertilizer accessibility, and agricultural management practices, and it ignores regional variations in the changes of P fertilizer usage across different countries before 1961. Consequently, uncertainties are substantially higher in the pre-1961 period, and the resulting maps should be interpreted primarily as broad-scale reconstructions rather than precise representations of historical subnational fertilizer distributions.

Third, we calculated manure P inputs based on manure N inputs and P : N ratio with spatiotemporal heterogeneity, but the gridded manure P data were not constrained by survey data, as country-level manure P inputs data were not available from the Soil Nutrient Budget database in FAOSTAT. Although temporal changes in livestock composition are considered when estimating manure N : P ratios, potential changes in species-specific manure nutrient ratios associated with evolving feed composition and livestock management practices are not explicitly represented, which may introduce biases in manure P estimates. Aside from mineral fertilizer, livestock manure, and atmospheric deposition, other anthropogenic P sources, including guano, livestock bones, and human excreta, were historically used to enhance soil fertility. These P sources were difficult to quantify at a global level without reliable data sources and were therefore not included or accurately quantified in the HaPi dataset.

The uncertainties in the HaPi dataset mainly arise from the input datasets and methodological assumptions used in developing the global P input estimates. Quantifying the overall uncertainties in the HaPi dataset is challenging due to the heterogeneity of underlying data sources and the scarcity of independent datasets for robust validation. Four major sources of uncertainty were identified as contributing to the overall uncertainty in the HaPi dataset. First, the FAO survey data, which serve as the primary constraint for national total P inputs, represent a critical source of uncertainty in estimating total P inputs to the terrestrial biosphere. Based on expert judgment, a generic uncertainty of approximately 20 % was assigned to the FAO national estimates (Tubiello et al., 2021). Consequently, the spatially explicit P input maps derived from these national data inherently carry at least the same level of uncertainty, with a minimum estimated uncertainty of 20 %. Second, uncertainties arise from the land use maps used to spatialize P inputs. Uncertainties in land cover classification or temporal interpolation in HYDE/LUHv2 data may propagate into the spatial allocation of P fertilizer and manure applications, thereby affecting local input rates and hotspot identification. Third, uncertainties stem from empirical assumptions and static parameters used during dataset construction, including the assumption of partitioning ratios for fertilizer use between cropland and pasture, the use of time-invariant crop-specific fertilizer and manure application patterns, the fixed manure N : P ratios for different livestock species, and the application of globally uniform change rates for reconstructing pre-1961 data. Finally, uncertainties in atmospheric P deposition simulated by the GEOS-Chem model are primarily associated with model structure and parameterization. Although the HaPi dataset primarily focuses on anthropogenic P inputs, it additionally includes both natural and anthropogenic sources. Collectively, these sources contribute to uncertainties in both the magnitude and spatial distribution of anthropogenic P inputs. Given the limitations and uncertainties in the HaPi dataset, it is important to collect or conduct surveys of crop-specific P fertilizer and manure use at subnational scales and update global land use data to reflect more precise regional patterns of global fertilizer and manure P inputs.