Thank you for visiting nature.com. The browser version you are using has limited CSS support. For the best experience, we recommend using a newer browser (or disabling compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we will display the site without styles and JavaScript.
We launched a new geospatial dataset, NPKGRIDS, which for the first time provides data on application rates of the three main plant nutrients—nitrogen (N), phosphorus (P, expressed as phosphorus pentoxide P2O5), and potassium (K, expressed as potassium oxide K2O)—for 173 crops as of 2020 at a geospatial resolution of 0.05° (approximately 5.6 km at the equator). NPKGRIDS was developed using a data fusion approach that combines crop mask information with eight published datasets of fertilizer application rates compiled from georeferenced data or national and provincial statistics. In addition, total N, P, and K application rates were compared with country-level information provided by FAO and the International Fertilizer Association (IFA), and validated against data provided by the National Statistics Office (NSO). NPKGRIDS can be used for global modeling, decision making and policy development to help maximize crop yields while reducing environmental impacts.
The use of chemical and mineral fertilizers has increased almost tenfold over the past 60 years1, and growing global demand for food and feed, driven by an expanding global economy, has contributed decisively to the growth of crop and livestock production over the same period2. At the same time, overuse and misuse of fertilizers through agricultural runoff creates serious environmental problems that can affect ecosystems and human health at all levels, from localized soil and water pollution to regional eutrophication hotspots and marine dead zones at the confluence of major rivers flowing through important agricultural regions3,4,5,6. The dual objective of ensuring food supplies to meet global demand while restoring and reducing environmental damage is a major challenge for people and the planet and is fundamental to the 2030 Agenda for Sustainable Development6,7 and the Global Biodiversity Framework8, particularly as fertilizer efficiency is needed to achieve high yields and sustainable agriculture.
Currently, two global datasets, FAO1,9,10 and IFA11, provide extensive information on nitrogen (N), phosphorus (P) and potassium (K) inputs to agriculture, with country-level statistics for the period 1961–2022, updated annually. In addition, limited data on crop-specific input rates are available12. This information is a recognized global source of information for analyzing agricultural fertilizer use and trends at the national, regional and global levels, as evidenced by dozens of published articles13, international reports14, sustainability indices15 and planetary frontier science data16.
At the same time, studies addressing local or regional issues often require more detailed subnational information to assess interactions between fertilizer use and key covariates (e.g., climate conditions, soil properties and water flows, ecosystem and crop distribution, farm management type, infrastructure, and population data). To address these needs, global spatial maps of fertilizer use have begun to appear in the literature17,18,19,20,21,22,23,24, primarily in the context of informing global biogeochemical studies and earth system science models. These products are a useful step toward refining information from national to subnational and grid-level data, but they have a number of important limitations. First, these new maps are often created by spatially representing existing national-level information without incorporating more detailed published data and subnational information from national statistical offices. Another reason is that creating such maps requires enormous amounts of data and computational resources to develop and validate, so existing products are largely ad hoc efforts that lack the coordination needed to ensure continuous improvement and updating. In fact, the most widely used geospatial dataset today, which contains nitrogen, phosphorus, and potassium application rates by crop type25 (hereafter referred to as MFM, for Mueller et al.’s Fertilizer Maps), is limited to data since 2003. Significant changes in agricultural land use and fertilizer use over the past 20 years9, coupled with dramatic changes in computing power and data storage space, suggest that the time is ripe for a major update to currently available products.
Here we present the results of a major new data fusion initiative to create NPKGRIDS, an updated dataset of global inorganic fertilizer application rates by crop type in 2020 for the major plant nutrients nitrogen (N), phosphorus (P2O5), and potassium (K2O). NPKGRIDS includes fertilizer application rates for 173 crops at a global spatial resolution of 0.05° (approximately 5.6 km at the equator). NPKGRIDS was developed using a data fusion approach combining crop mask information recently provided in CROPGRIDS26,27 with other relevant published data sources, as illustrated below. First, we searched and collected available peer-reviewed and national crop-specific fertilizer use data, selecting eight datasets containing information describing individual crops or aggregated crop groups in either geo-referenced or tabular format. We then selected the most appropriate dataset for each crop and provincial unit using the same data merging optimization process and quality assessment system as in CROPGRIDS. Subsequently, national statistics for general application published by FAO and IFA, as well as national statistical offices, were used for the NPKGRIDS benchmarking.
We examined and collected georeferenced and tabular datasets reporting application rates or amounts of N, P, and K fertilizers applied to individual crops at the national and/or subnational levels. To ensure data reliability, we collected only peer-reviewed and nationally sourced datasets. We then reviewed these datasets in detail following the workflow illustrated in Figure 1, which includes three main steps: step 1) harmonizing input datasets into tabular format at the subnational level; step 2) defining endogenous data quality indicators; and step 3) global spatial distribution of fertilizer application rates.
NPKGRIDS development workflow. Step 1: Harmonization of input datasets in tabular format; Step 2: Definition of endogenous data quality indicators; Step 3: Global spatial distribution of fertilizer application rates; Step 4: Validation. MFM: Mueller et al. Fertilizer Map; MRF: Monfreda et al. dataset; GAUL: Global Administrative Units Level Dataset.
The starting point for NPKGRIDS is CROPGRIDS26,27, a recently developed georeferenced crop map dataset that details where crops are grown and the area where they are harvested. We then reviewed the available peer-reviewed literature and official national statistics to obtain georeferenced and tabulated fertilizer use datasets that specify individual crops and/or aggregated crop groups. We included only datasets that were dated years later than 2003, were the latest time coverage of MFM25 and had crop names consistent with the FAO Indicative Crop Classification (ICC)28. We excluded datasets that were not crop-specific or contained aggregated crops without further specification of the constituent crops. The collected datasets contain the mass and/or application rate of total nitrogen (N), total phosphorus (P or P2O5) and total potassium (K or K2O) from direct and/or compound fertilizers. Of these, we selected eight N datasets and seven P2O5 and K2O datasets (Table 1 and Supplementary Table 1). Among the selected datasets, the Fertilizer Use in Cyclical Crops (FUBC)29 includes crop-specific and aggregated crop group data for 63 countries for the period 2016–2018. We divided it into two datasets: one contains only individual (IDV) crops (FUBC18-IDV) and the other contains only aggregated (AGG) crop groups (FUBC18-AGG). The Historical Fertilizer Use in Crop (HFUBC)12 dataset combines all crop fertilizer use data from IFA and FAO for individual crops and crop groups in 111 countries from 1978 to 2018. Only 12 individual crops from 65 countries in the HFUBC from 2006 to 2018 were used in this study. Note that FUBC18 and HFUBC are nationally resolved datasets. Also included are four National Statistical Office (NSO) datasets from the United States30 (US), Belarus31 (BY), the United Kingdom32 (UK), and Australia33 (AU), which provide crop-specific fertilizer data at the subnational level. These eight datasets were used as input to construct NPKGRIDS (Table 1).
Additional output datasets are also used, in particular to assist in the calculation and spatial placement of NPKGRIDS. In particular, two georeferenced global crop map datasets were used to inform crop locations and harvested areas, namely the CROPGRIDS26,27 dataset containing 0.05° maps of 173 crops in 2020, and the Monfreda et al.34 dataset (hereafter referred to as MRF after the initials of the original authors) containing 0.0833° maps of 175 crops around 2000, in both cases using the FAO crop species nomenclature. When the selected dataset did not contain fertilizer application rates but only total fertilizer application data, we estimated fertilizer application rates in terms of mass per unit of harvested crop area using country-level crop area harvested data publicly available from national statistical offices (e.g. CROP-AU35 and CROP-BY36) or FAOSTAT37. We used the FAO Global Administrative Units Level (GAUL) dataset to define country boundaries and subnational units38 (Table 2).
The eight input data sets (Table 1) were combined into a common tabular format where N, P2O5 and K2O application rates were expressed as input per unit area harvested. The tabular resolution represents the best scale of each data set, i.e. subnational (level 1) for MFM, US, BY, AU, UK and national (level 0) for FUBC18-IDV, FUBC18-AGG and HFUBC.
For the georeferenced MFM dataset, we first constructed a table of application rates using a GAUL level 1 mask at the native resolution of the dataset, i.e. 0.0833° (approximately 10 km at the equator). In subnational units with missing application rate data, we imputed the missing information using the national weighted average FMFM application rate [kg ha−1] for fertilizer n, crop i, and country j, calculated as follows:
where fMFM(n,i,j,r) (unit: [kg ha−1]) is the available fertilizer application rate n for crop i in provincial unit r of country j in the MFM dataset, and AMRF is the corresponding harvested area obtained from the MRF dataset.
For all other tabular datasets, the matching process consisted of converting the variables into application rates expressed as the mass of N, P2O5, and K2O applied per unit area harvested. If a dataset contained only application quality information, we calculated application rates using the crop-specific harvested area for the corresponding year from the corresponding output dataset (Table 2). Specifically, harvested areas from FAOSTAT were used for the HFUBC, FUBC18-IDV, and FUBC18-AGG datasets, while CROP-AU and CROP-BY were used for the AU and BY datasets, respectively. For the US dataset, which contains information on the application rate per unit area and the percentage of harvested area fertilized, we calculated the application rate for the entire harvested area for each crop by multiplying them together. For the BY dataset, the P and K masses were multiplied by 2.29 and 1.20, respectively, to convert them to P2O5 and K2O. In the UK dataset, application rates for some crops varied between seasons. In this case, due to the lack of intra-annual information in our product, we used the average seasonal application rate to determine the average annual application rate for a given crop. For the AU dataset and for some countries where the data spans two calendar years in FUBC18-IDV and FUBC18-AGG, we assigned the first calendar year as the base year. For datasets that contain a list of crops in a crop group, such as FUBC18-AGG and UK, group application rates were assigned to all component crops in the group based on the summary in Supplementary Table S3.
Other georeferenced and tabular outputs and control datasets (Table 2) were also converted to the same data format and administrative unit level as the input datasets.
We developed a multi-criteria ranking scheme to identify values that best reflect crop-specific fertilizer application rates in local units using multiple sources across eight selected input data sets. The ranking is based on three endogenous data quality metrics: Qc, crop specification; Qr, data resolution; and Qy, synchronicity. Each metric is assigned a value ranging from zero (lowest quality) to one (highest quality). For each data set, metric values may vary across crops and local units.
The Qc indicator indicates whether the fertilizer data relate to a single crop or a group of crops:
Datasets that include both specific crops and aggregate crop data (e.g. UK) will have different Qc values between crops, with individual crops ranking higher.
The Qr metric ranks the administrative resolution of datasets: the higher the resolution, the higher the ranking of the dataset, as shown below:
The Qy metric is used to assess the degree of synchronization between the base year Yr of the input dataset and the base year of NPKGRIDS, which is set to the period from 2015 to 2020, hereafter referred to as “around 2020″, and is defined as
Qy increases as \({Y}_{r}\) approaches the 2015–2020 period and can vary for a wide range of base years \({Y}_{r}\) within the same data set, such as HFUBC and the US.
Table 3 summarizes the above endogenous data quality measures for all data sets used to compile NPKGRIDS. For practical purposes, we relate the average endogenous quality for each data set k, crop i, and subnational unit r as
Following the algorithm presented in Figure 2, the georeferenced application rates of N, P2O5, and K2O for each crop were globally summarized. First, we decomposed the national application rates of the three fertilizers in HFUCB, FUBC18-IDV, and FUBC18-AGG into subnational application rates using the national-to-subnational apportionment calculated based on the MFM, if MFM data were available. In this step, the crop-specific application rate fk(n,i,j,r) is calculated for a data set k of fertilizer n for crop i in subnational unit r in country j as follows:
where fMFM is the application rate for the respective crop and local unit in MFM, and α is a scaling factor defined as
where Fk(n,i,j) is the national fertilizer rate n for crop i and country j in dataset k, and ACR is the corresponding crop area in CROPGRIDS. In equation (7), we assume that these proportions α remain constant between 2000 and 2020. This is true only if we assume that this geographical variation is largely due to agrometeorological differences rather than management practices, or that the latter geographical variation remains the same in both time periods.
Synthesis algorithm for global maps of crop-specific fertilizer application rates. The names of the data sets are shown in Tables 1 and 2.
For each fertilizer n applied to crop i in provincial unit r, we checked whether the application rate could be obtained from multiple data sets. If only one data set k is available, the selected application rate is f(n,i,r) = fk(n,i,r) (Figure 2). If multiple data sets are available, we select the best-fitting data set kbest, which has the highest internal quality Qk,i,r defined in equation (5), so f(n,i,r) = \({f}_{{k}_{{best}}}\)(n,i,r). If the Qk,i,r of the two data sets is equal, the data set with the most recent base year is selected as kbest. Alternatively, if the base year of these data sets is the same, then \({f}_{{k}_{{best}}}\left(n,i,r\right)\) is calculated as the average of all data sets with the same Qk,i,r and base year. If a data set was missing, we first checked whether neighboring provincial units had available data to fill the gaps. Specifically, if the fertilizer application rate for crop i is available in w neighboring provincial units, the area-weighted average fertilizer application rate favg(n,i,w) is calculated in w neighboring provincial units as follows:
Where nw is the number of common boundary grids. If no fertilizer was applied to crop i in adjacent provincial units n, we estimated f(n,i,r) based on the application rates for similar crops according to three criteria defined by FAO39: (a) classification (i.e. cereals, legumes, nuts, fruits and berries, spices, perennial oilseeds, annual oilseeds, forages, fibre crops, vegetables and other perennial crops); (b) life span (i.e. temporary or permanent); and (c) stem type (i.e. grass, shrub or tree; see Supplementary Table 4). Crop i and crop c are considered similar if they share at least two of the above three criteria. If a fertilizer application rate n for similar crops c is available in a provincial unit r (Figure 2), we calculate f(n,i,r) = favg(n,c,r), where favg(n,c,r) is the area-weighted average fertilizer application rate for c similar crops in provincial unit r, that is,
If there are no similar crops in national unit r (Figure 2), we calculate f(n,i,r) = favg(n,c,g), where favg(n,c,g) is the area-weighted average application rate c of similar crops in all national units g, that is,
Finally, to generate a global geographic map of nutrient and crop application rates, provincial crop-specific fertilizer application rates were spatially uniformly distributed across the grid cells containing crop i within a given provincial unit using the crop mask in CROPGRIDS. Figure 3 shows example application rate maps for cotton N, P, and K, along with the corresponding overall data quality and data sources used to generate the maps.
An example of a distribution plot of NPKGRIDS data for cotton. From left to right, columns are: N, P2O5, and K2O; from top to bottom, fertilizer application rate, data quality, and data source.
The NPKGRIDS dataset distributes global georeferenced maps of N, P2O5, and K2O fertilizer application rates for 173 crops (see Supplementary Table 4 for a list of crops) at 0.05° resolution (about 5.6 km at the equator) from −180° to 180° longitude and −90° and 90° latitude using the WGS-84 coordinate system. The georeferenced maps are distributed as NetCDF files in which grid cells containing ocean/water are labeled as “-1”. The files included in the dataset are shown in Table 4. The NPKGRIDS dataset is available for download from the figshare40 repository at https://doi.org/10.6084/m9.figshare.24616050. The phosphorus and potassium fertilizer data were categorized by oxide application rate. These can be converted to element-based application rates using the following conversions: 1 kg P2O5 is equivalent to 0.436 kg P, and 1 kg K2O is equivalent to 0.83 kg K.
Due to the lack of data sets other than the crop fertilizer data already used in this paper, we estimated NPKGRIDS data using total fertilizer input data for N, P2O5 and K2O at the national level (Table 1) provided by FAOSTAT41 (160 countries) and IFA11 (110 countries). To do this, we first calculated the total fertilizer input M(n,j) for country j estimated by NPKGRIDS as follows:
where ACR(p,i,j) is the harvested area of crop i in grid p of country j in CROPGRIDS26, and f is the corresponding fertilizer application rate n in NPKGRIDS. Country boundaries are defined based on the GAUL38 dataset (level 0). We then compare M(n,j) with the corresponding fertilizer use reported in FAOSTAT and IFASTAT, MFAO and MIFA, respectively, averaged over the period 2015–2020. These comparisons were characterized using the coefficient of determination R2 (similar to the Nash-Sutcliffe efficiency coefficient), the concordance correlation coefficient (CCC), and the normalized root mean square error (NRMSE), expressed as
where \({O}_{x}\) represents the logarithm of the MFAO or MIFA, and \(E\) represents the logarithm of the national application mass (M) calculated according to NPKGRIDS. \(\bar{{O}_{x}}\) and \(\bar{E}\) are the corresponding means for all countries, \({{\sigma }_{{O}_{x}}}^{2}\) and \({{\sigma }_{E}}^{2}\) are the corresponding variances, and \(\rho \) is the Pearson correlation coefficient between Ox and \(E\). Mx represents the MFAO or MIFA, \({M}_{x,\max }\) and \({M}_{x,\min }\) are the maximum and minimum fertilizer masses corresponding to each country, and \({n}_{p}\) is the number of data points.
According to NPKGRIDS, the total nitrogen applied worldwide is 100 million tonnes, which is about 10% lower than the global FAOSTAT and IFASTAT estimates for 2020 (110 million tonnes and 112 million tonnes, respectively). At the national level (left column of Fig. 4), nitrogen application estimated using NPKGRIDS was in relatively good agreement with FAOSTAT (R2 = 0.76, CCC = 0.89 and NRMSE = 0.01) and fairly good agreement with IFASTAT (R2 = 0.66, CCC = 0.87 and NRMSE = 0.01). Compared to FAOSTAT data, underestimation of nitrogen application was found mainly in Africa, such as the Democratic Republic of Congo, Namibia and Madagascar. Compared with FAOSTAT and IFASTAT data, NPKGRIDS consistently overestimates nitrogen application rates in Iraq, Syria and Jordan.
Comparison of fertilizer application rates in NPKGRIDS with FAOSTAT (top row) and IFASTAT (bottom row) for N (left column), P2O5 (middle column) and K2O (right column). Each marker in the scatterplot represents a country, and the black line indicates a 1:1 ratio.
For phosphorus, NPKGRIDS estimates global application at 46 million tonnes, which is very close to the FAOSTAT and IFASTAT estimates for 2020 (48 million tonnes and 49 million tonnes, respectively). A comparison of total P2O5 use at the country level showed the strongest correlation with FAOSTAT (R2 = 0.82, CCC = 0.91 and NRMSE = 0.02) and IFASTAT (R2 = 0.70, CCC = 0.88 and NRMSE = 0.01, Fig. 4, middle column). In general, as with nitrogen data, the differences in phosphorus application data between NPKGRIDS and FAOSTAT and IFASTAT data are more pronounced in countries in Africa and the Middle East.
For potash, NPKGRIDS reports a global application of 40 million tonnes, which is close to the global FAOSTAT and IFASTAT estimates (39 million tonnes and 41 million tonnes, respectively). However, comparison of total K2O applications at the country level showed low agreement with the FAO/IFA estimates (Figure 4, right column), with low correlations for FAOSTAT (R2 = 0.68, CCC = 0.84, and NRMSE = 0.01) and IFASTAT (R2 = 0.50, CCC = 0.77, and NRMSE = 0.01). NPKGRIDS tends to overestimate potash applications in North Africa and West Asia.
We obtained crop-independent national and provincial total N, P2O5, and K2O application data for 37 countries and 166 provincial units from 2006 to 2020, including 32 European countries42, India43, Pakistan44, China45, Iran46, and Sri Lanka47 (Table 2, Supplementary Table 2). Of the 37 countries, 11 provided provincial-level data and 26 provided only national data. Only five countries (99 provincial units in total) provided K2O data. We calculated the average over 2015–2020 for all NSO data except Iran, for which the latest available data are from 2006. We summarized pixel-level data in NPKGRIDS into national and provincial total N, P2O5, and K2O application rates according to formula: 11, where j is a national (level 0) or subnational (level 1) unit defined within the boundaries of a GAUL38 administrative unit. The quality of the comparison of NPKGRIDS and NSO data was quantitatively assessed using R2 (Equation 12), CCC (Equation 13), and NRMSE (Equation 14).
Comparison of NPKGRIDS and NSO for national and provincial N applications showed relatively good agreement with R2 = 0.80, CCC = 0.90 and NRMSE = 0.03 (Fig. 5), whereas P2O5 and K2O estimates were weaker, with R2 values of 0.74 and 0.75 with NSO data, respectively.
Comparison of the quality of fertilizer applied to all crops at national and provincial level between NPKGRIDS and the National Statistical Office (NSO). The total mass of material applied was (a) N, (b) P2O5, and (c) K2O. The colored markers represent the different national statistical organizations: EU (European Union), LK (Sri Lanka), PK (Pakistan), IR (Iran), IN (India), and CN (China). The black line represents the 1:1 ratio.
A comparison of data from 32 countries with EUROSTAT42 data shows good agreement between national and subnational data, with some exceptions. In particular, total fertiliser use N and P2O5 was significantly underestimated in Iceland and Ireland (N and P2O5) and in Malta (N). This underestimation is likely due to high uncertainty in the harvested area data provided by CROPGRIDS for these countries and, for Ireland, in particular uncertainty in fertiliser use in grassland and pasture. In Iceland, only potatoes are mapped in NPKGRIDS. In contrast, fertiliser application was slightly overestimated in China.
NPKGRIDS takes into account uncertainties and errors in input datasets, such as the original fertilizer dataset and the CROPGRIDS dataset used to spatially distribute fertilizer application rates. Uncertainties can arise from erroneous or missing fertilizer application and crop area data submitted for national and international reporting. For example, MFM faces data limitations in many low- and middle-income countries, and phosphorus and potassium application data show more anomalies than nitrogen application data. CROPGRIDS, on the other hand, is constructed by coordinating multiple data sources (including surveys, remote sensing, and models), each with uncertainties that are passed on to the construction of NPKGRIDS.
The spatial arrangement of national and provincial data within grid cells in NPKGRIDS introduces additional uncertainty. For example, the spatial distribution of national data (e.g., HFUBC, FUBC18-IDV, FUBC18-AGG) assumes that relative rates of fertilizer use within a country follow the same pattern as in the MFM (Equation 7), ignoring potential relative differences in cropping practices that may occur between different provinces within a country. In addition, information obtained directly from the MFM does not take into account changes in fertilizer use that may have occurred in these areas over the past 20 years.
Finally, due to spatial resolution limitations, NPKGRIDS excludes a number of small countries and territories, including the Falkland Islands, Faroe Islands, French Southern and Antarctic Territories (SAT), Hart Island, Isle of Man, Kingman Reef, Kiribati, Matansala, Mayotte, Netherlands Antilles, Palau, Réunion, Saint Pierre, South Georgia, Svalbard, and the Virgin Islands.
To quantify potential uncertainties, we calculated subnational data quality scores based on endogenous quality scores and comparisons with FAOSTAT and IFASTAT data. The overall NPKGRIDS data quality \(Q(n,i,j,r)\) for nutrient n (i.e. N, P2O5 and K2O) of crop i in subnational unit r of country j is calculated as follows:
where Qk is the endogenous quality of the selected dataset calculated using equation (5), and the base quality of QFAO and QIFA with the FAOSTAT41 and IFASTAT11 datasets is defined as
x is FAO or IFA, and Qx ranges from 0 (low quality) to 1 (high quality). For subnational units with gaps in fertilizer application rates, we assign a value of zero to the corresponding Qk. Data quality maps are distributed with the NPKGRIDS dataset. An example of a data quality chart for cotton is shown in Figure 3 (second row).
All georeferenced maps distributed in the NPKGRIDS40 dataset are formatted as standard NetCDF4 files. These files can be read and analyzed using a variety of coding languages (e.g. MATLAB, Python, Julia, R) and software (ArcGIS, QGIS, Panoply). The NPKGRIDS dataset contains the same crops as the CROPGRIDS26 dataset, following the naming system used by FAO28.
Post time: Mar-31-2025