CropGlobe: Global Crop-Specific Fertilizer Data

• CropGlobe is a global, crop-specific dataset that reconstructs historical fertilizer application rates for major inorganic nutrients (N, P2O5, and K2O) from 1961 to 2019.
• It combines machine learning predictions with spatial downscaling to generate both country-level estimates and 5-arcmin resolution maps for 13 major crop groups.
• The dataset supports environmental modeling and policy-making by linking fertilizer use trends to socioeconomic and agrological drivers worldwide.

Searching arXiv for "CropGlobe" and closely related papers to ground the article. CropGlobe is a global, crop-specific fertilization dataset built to reconstruct historical application rates of the three major inorganic nutrients—nitrogen (N), phosphorus pentoxide (P$_2$O$_5$), and potassium oxide (K$_2$O)—for major crop groups around the world from 1961 to 2019. Its central purpose is to fill a major data gap in global agriculture: although fertilizer use is one of the most important drivers of yield, environmental impacts, and nutrient cycling, there has been no consistent, spatially explicit, long-term dataset describing how much fertilizer is applied to different crops in different countries over time. CropGlobe provides both country-level estimates and gridded maps, enabling analyses of fertilization trends, spatial hotspots, and the socioeconomic and environmental drivers behind fertilizer use (Coello et al., 2024).

1. Definition and scope

CropGlobe is a reconstructed, machine-learning-based, globally gridded dataset of crop-specific fertilizer application rates for N, P$_2$O$_5$, and K$_2$O from 1961–2019 (Coello et al., 2024). The fertilizer variables are expressed in the standard oxidized forms used in international fertilizer reporting: N, in kg ha$^{-1}$; P$_2$O$_5$, in kg ha$^{-1}$; and K$_5$0O, in kg ha$_5$1. The final product is released as 5-arcmin resolution GeoTIFF maps covering the globe from 180°E to 180°W and 90°S to 90°N, in WGS84 / EPSG:4326. Each file is named by crop, fertilizer, and year.

The dataset spans Time: 1961–2019, Space: global, at 5 arc-min resolution, Units: country-level predictions and gridded cell-level maps, Crop coverage: 13 major crop groups, and Nutrients: N, P$_5$2O$_5$3, and K$_5$4O (Coello et al., 2024). This gives a 60-year historical view of fertilizer application patterns across major crops and countries, later downscaled into spatially explicit maps.

The stated motivation is analytical rather than merely descriptive. As global fertilizer application rates increase, high-quality datasets are presented as paramount for comprehensive analyses to support informed decision-making and policy formulation in crucial areas such as food security or climate change. CropGlobe is therefore framed as an input resource for environmental modeling, causal analysis, fertilizer price predictions, and forecasting (Coello et al., 2024).

2. Data contents and representation

The dataset provides crop-specific fertilizer application rates for 13 crop groups: Wheat, Maize, Rice, Other cereals, Soybean, Palm oil fruit, Other oilseeds, Vegetables, Fruits, Roots and tubers, Sugar crops, Fiber crops, and Other crops (Coello et al., 2024). These groups were aligned with the FAO Indicative Crop Classification and were designed to standardize and merge a wide range of heterogeneous source datasets.

CropGlobe combines country-level estimates with gridded maps. The country-level component originates from machine-learning predictions for crop-, country-, and year-specific fertilizer application rates. The map component is a set of 5-arcmin resolution products derived by downscaling those country estimates using cropland and crop distribution data (Coello et al., 2024).

The representational choice is important because the underlying source material was heterogeneous in crop naming and crop detail. Because the original sources differed in crop naming and crop detail, all crops were standardized into the 13 groups listed above. The resulting harmonized dataset contained 3712 country–crop–year combinations, covering 114 countries from 1973 to 2018 for the observed training data (Coello et al., 2024). Important preprocessing steps included converting seasonal data to the starting year of the season, selecting the most recent source when duplicate values existed, aggregating split crop varieties or management systems using area-weighted averages, transforming percentage-of-fertilized-area records into average application rates over total cropland, excluding implausibly high values above 5000 kg ha$_5$5, and one-hot encoding categorical variables.

This suggests that CropGlobe is not a direct observational archive for all countries and years. Rather, it is a reconstructed panel that combines observed historical reports with estimated values and then imposes national consistency and spatial downscaling.

3. Construction workflow and modeling framework

The workflow had three major stages: 1. Compile and harmonize crop-specific fertilizer data; 2. Train machine learning models to estimate missing country-year-crop values; 3. Downscale country estimates to 5-arcmin maps using cropland and crop distribution data (Coello et al., 2024). The authors assembled 14 global and national datasets on fertilizer use by crop, country, and year. These included historical FAO, IFDC, IFA, and Fertilizer Europe reports, plus the Ludemann et al. database and several national datasets used later for validation.

The models used a rich feature set spanning environmental, agrological, and socioeconomic drivers. These included climate: annual precipitation, temperature, evapotranspiration, aridity index; soils: N, organic carbon stock, sand, silt, clay, pH, cation exchange capacity; agrology: crop area, crop share of cropland, country fertilizer per hectare, total fertilizer use, holding size, nutrient removal, irrigation share, machinery use; economics: crop prices, global fertilizer prices, fertilizer production-distance proxies, GDP per capita, education expenditure, population pressure (Coello et al., 2024). A key environmental aggregation equation used was: $_5$6 where the environmental variable for a country-crop-year was computed as a crop-area-weighted mean over grid cells.

Two gradient-boosted tree models were trained: eXtreme Gradient Boosting (XGB) and HistGradientBoosting (HGB). These were chosen because they handle nonlinear relationships and missing values well. The models were trained using nested 2 × 5 cross-validation, with grid search over hyperparameters. Performance was assessed with Pearson correlation coefficient / $_5$7, MAE, MSE, and RMSE (Coello et al., 2024).

HGB performed slightly better overall and was used for the downstream validation and interpretation. Reported test performance was N: MAE $_5$8, RMSE $_5$9, $_2$0; P$_2$1O$_2$2: MAE $_2$3, RMSE $_2$4, $_2$5; K$_2$6O: MAE $_2$7, RMSE $_2$8, $_2$9 (Coello et al., 2024). These substantially outperformed a naive mean predictor.

Because the models were trained on global compiled data that can differ from national fertilizer totals, the authors then adjusted the predictions to match FAOSTAT country-level fertilizer consumption, after removing the portion used on grasslands and fodder crops. The adjustment formula was: $_2$0 In words: the model’s crop-level predictions were scaled so that, when multiplied by harvested area and summed across crops, they matched the national fertilizer total that FAOSTAT reports for cropland use (Coello et al., 2024).

4. Spatial downscaling and gridded map generation

The map generation pipeline combined the EARTHSTAT 2000 crop maps for spatial crop allocation, HYDE 3.3 historical cropland maps for changes over time, FAOSTAT country harvested area totals, and the country-level ML predictions (Coello et al., 2024). This was done in three steps.

First, the authors created gridded harvested area by crop and year. EARTHSTAT provides crop-specific harvested area in 2000. HYDE provides historical cropland area through time. The authors used these together to construct gridded harvested-area maps for every year from 1961 to 2019. For rice and non-rice crops, they used separate rules. The basic logic was: preserve 2000 crop spatial patterns, scale them through time using HYDE’s cropland dynamics, and use local neighborhood averaging when new cropland appeared (Coello et al., 2024).

Second, the gridded crop-area map was adjusted so that, when summed over each country, the grid cells matched FAOSTAT harvested area totals: $_2$1 This ensured consistency between the spatial allocation and the country totals (Coello et al., 2024).

Third, fertilizer mass or intensity was computed per grid cell: $_2$2 So each grid cell received a fertilizer amount based on how much of the crop is grown there, which country occupies the cell, and the country-level fertilizer application rate for that crop (Coello et al., 2024).

A plausible implication is that CropGlobe should be interpreted as a spatially explicit reconstruction constrained by country totals and historical land-allocation assumptions, rather than as a direct measurement product at 5-arcmin resolution. The paper’s description of the downscaling logic supports that interpretation.

5. Validation, drivers, and reported historical patterns

The dataset was validated in two complementary ways: 1) Model-level validation using nested cross-validation, and 2) External validation against national databases from USA, UK, India, Sweden, Philippines, New Zealand, and Pakistan (Coello et al., 2024). The validation used MAE and MAPE, and for some countries direct time-series plots were shown. The broad conclusion was that the dataset reproduces national patterns reasonably well, typically with MAE values around 5–40 and MAPE values around 10%–50%, although performance varied by country and crop. The worst discrepancies were seen in countries like Sweden and New Zealand, where national data were sparse and/or the training data were limited.

The predictions were also interpreted with SHapley Additive exPlanations. Across N, P$_2$3O$_2$4, and K$_2$5O, Agrological variables dominated the predictions. The strongest positive drivers were country fertilizer per hectare and crop nutrient removal per hectare. The strongest negative driver was crop nutrient content (Coello et al., 2024). For P$_2$6O$_2$7 and K$_2$8O, GDP per capita was the most important socioeconomic factor, whereas for N, global crop price mattered more. Environmental variables played a smaller role, but the notable ones were soil pH, soil organic carbon stock, and aridity index.

The paper also highlights broad historical tendencies revealed by the dataset. N use increased strongly for wheat, maize, rice, and especially vegetables. P$_2$9O$_5$0 increased across all major crops, again especially vegetables. K$_5$1O increased strongly in maize and oil palm, with oil palm showing one of the largest jumps. Geographically, fertilizer use shifted from being concentrated in the USA and Europe in the 1960s to being increasingly concentrated in Asia by the 2010s, especially China and India (Coello et al., 2024).

These findings frame CropGlobe not only as a repository of estimates but also as an empirical basis for studying intensification. The paper emphasizes several uses for the dataset: food security; climate change and environmental modeling; causal analysis; fertilizer price prediction; forecasting; and other modeling applications where yield, ecological impact, or fertilizer pricing are the outputs rather than fertilizer use itself (Coello et al., 2024).

6. Position within crop-global data and modeling research

CropGlobe belongs to a broader class of crop-global resources that combine machine learning, remote sensing, and process-based reasoning, but its object of estimation is fertilizer application rather than crop type, yield, or management policy. This differentiates it from county-level crop yield datasets such as CropNet, described as “the first terabyte-sized, publicly available, and multi-modal dataset” for climate change-aware crop yield predictions across the contiguous United States (Lin et al., 2024), and from reinforcement-learning environments such as CropGym, which is a standardized OpenAI Gym interface for fertilization management using process-based crop growth models (Overweg et al., 2021).

It also differs from global crop-mapping systems that focus on label scarcity and geospatial transfer. For example, GEDI- and Sentinel-2-based work has produced annual, field-scale, 10 m maps of tall and short crops at the global scale using sparse structural labels and local optical models (Tommaso et al., 2022), while related work showed that GEDI profiles provide much more invariant features across geographies compared to spectral and phenological features detected by passive optical sensors (Tommaso et al., 2021). In a later and distinct study, the name “CropGlobe” was used for a global crop type dataset with 302,052 pixel-level samples from 8 countries across 5 continents, designed specifically to study geographic invariance in remote sensing-based crop classification (Tong et al., 3 Sep 2025). That later usage refers to a different resource centered on crop-type transferability rather than fertilization intensity.

This naming overlap can create confusion. In the fertilization literature, CropGlobe denotes a reconstructed, long-term, crop-specific nutrient-application dataset (Coello et al., 2024). In the crop-classification literature, CropGlobe denotes a benchmark for cross-country, cross-continent, and cross-hemisphere transfer (Tong et al., 3 Sep 2025). The shared name reflects a common concern with globe-scale agricultural information, but the data modalities, labels, and scientific objectives are different.

Within this broader landscape, CropGlobe’s distinctive role is to supply a historical, crop-resolved fertilization layer that can be coupled to environmental modeling, yield analysis, and decision systems. This suggests a natural complementarity with process-based or hybrid crop models that explicitly represent nutrient, water, and climate interactions, including frameworks for crop management optimization (Overweg et al., 2021), deployable RL-based management systems under partial observability (Wu et al., 2024), and hybrid global gridded crop models that combine process-based structure with learned components for improved yield predictions under climate stress (Lin et al., 23 Dec 2025).