The identities and spatial extents of land-use and land-cover classes were drawn from the global HILDA+ dataset (Winkler et al., 2020, 2021) at 0.01° (Table 1). The cropland areas were refined into specific crop categories by integrating the CROPGRIDS v1.08 dataset (Tang et al., 2024), which provided primary crop data for maize, wheat, soy, rice, and starchy roots at 5 arcmin resolution for the year 2020. To determine management intensity for these crops we utilised a crop-specific N fertiliser dataset (Adalibieke et al., 2023). Pasture land management intensity was derived from total N fertilizer datasets (Tg N yr⁻¹) (Tian et al., 2022) at 5 arcmin. The HILDA+ forest management (managed/unmanaged) dataset (Winkler et al., 2021) was combined with a 100 m resolution forest management dataset (Lesiv et al., 2022) to determine forest management intensity. To account for large-scale photovoltaic (PV) deployment worldwide, we integrated a global PV panels dataset at 20 m spatial resolution (Li et al., 2025), ensuring that solar energy installations were spatially represented in the land-use dataset. We utilised some input datasets corresponding to different reference years than 2020, particularly irrigation and forest management data from 2015. These datasets were considered as they provide the most consistent and currently available spatially-explicit information. To harmonise these inputs, we assumed that large-scale spatial patterns of irrigation and forest management remained relatively stable over short periods of time at the global scale. Thus, the final dataset represented an approximation for the year 2020.

Country-wise FAO crop use data were applied to categorise land-use into food-feed-bioenergy uses using random distribution ensuring spatially-explicit classification at the national level. To characterise socioeconomic contexts, we used the global gridded socio-economic dataset of Perkins et al. (2025) at 1 km resolution. Some of the input variables used to construct the land management intensity dataset included socioeconomic and accessibility related variables in their underlying methodologies, which however differed from the variables used in this study in terms of spatial resolution, temporal reference and role. The passive variables of the inputs of the management intensity dataset relied largely on country-level or coarse-resolution inputs and earlier reference years, whereas the analysis presented here utilised the spatially-explicit and harmonised socioeconomic datasets specifically for the year 2020 (Perkins et al., 2025). These variables were explicitly used as explanatory factors in a separate regression framework, and were not used as direct inputs into the construction of the management intensity dataset. We utilised four globally available socioeconomic indicators representing complementary dimensions of development and accessibility and two biophysical variables – aridity index and elevation. Gross domestic product (GDP) per capita was used as a proxy for financial capital, reflecting economic capacity. Population density captures aspects of human capital and demographic pressure on land systems, influencing labour availability and land management intensity. Market accessibility represents infrastructural and manufactured capital, approximating access to markets and services that facilitate intensive land management. Income inequality, measured using the GINI coefficient, was included as an indicator of social structure and distributional inequality. GINI may influence decisions related to the adoption of land management practices as it is determined by the ability to access to the land, capital, and agricultural inputs Together, these variables provided a parsimonious representation of socioeconomic conditions. In addition, we included aridity index and elevation to account for important biophysical constraints on land management. The aridity index represented long-term moisture availability, where lower values indicate more arid conditions and higher values indicate more humid environments. This is important for understanding the feasibility of intensive and irrigated land management. Elevation was included as a proxy for terrain-related constraints such as climate limitations, which influence the intensity of land management practices. Finally, we compared the land management intensity dataset with a land-use decision-making dataset (Malek and Verburg, 2020) to assess the correlation with the management practices and land-use decision-making types. More details on the datasets used are provided in Table 1.

To map the primary crop types – wheat, maize, rice, oil-N-fixing crops (soy), and starchy-roots (potatoes and tubers) – we used the overall cropland extent from the HILDA+ dataset and then integrated crop specific datasets from CROPGRIDS v1.08 (Tang et al., 2024) to spatially allocate each crop type. The crop-specific harvested area layers were extracted and harmonised to a common spatial grid, with strict enforcement of identical resolution, extent, and coordinate and projection systems across all crop types. Cropland cells with non-zero harvested area for at least one crop were treated as cropland grid cells. A binary presence-absence map was created for each crop type, indicating whether a specific crop was present in a given grid cell. These binary maps were then aggregated into a single dataset to represent the distribution of all crop types. Missing or negative values were assigned as zero. To address overlapping crop types within a grid cell, we allocated the crop type with the highest harvested area as the dominant crop, ensuring one-cell one-crop representation. To ensure consistency across the dataset, we reprojected the data onto the 0.01° resolution by aggregating neighbouring cells and assigning the majority crop type within each aggregation unit.

To assess the management intensity associated with each crop type, we utilised a crop-specific N fertiliser, manuring, and irrigation dataset (Table 1; Adalibieke et al., 2023; Mehta et al., 2024). Fertiliser application rates (kg N ha⁻¹ yr⁻¹) were derived for wheat, maize, rice, soybean, and starchy roots. For each crop, we considered various N fertiliser types including Urea, nitrate fertilisers (ammonium nitrate (AN) and calcium ammonium nitrate (CAN)), compound fertilisers (ammonium phosphate (AP), N K compounds (NK), N P K compounds (NPK), other NP (ONP)), AA&NS (anhydrous ammonia (AA) and N solutions (NS)), other synthetic fertilisers (other N straight (ONS) and ammonium sulphate (AS)) and MA&CR (manure (MA) and crop residues (CR)). These fertiliser components were aggregated at the grid cell level to obtain total N application rates for each crop type. Prior to aggregation, values were screened to remove implausible negatives and extreme outliers. The total crop-wise N fertiliser rates (originally at 5 arcmin resolution) were resampled to the 0.01° grid using the nearest-neighbour resampling approach, and thus all finer resolution cells within each coarse grid cell inherit the averaged value. To determine the management intensity, we classified the crop-specific N fertiliser rates dataset based on absolute N fertiliser thresholds applied to the dominant crop in each cell. Nitrogen application rates were classified into three classes of intensity based on the relevance of their application for agriculture as described by Kleijn et al. (2009): very extensive (<50 kg N ha⁻¹ yr⁻¹), extensive (50–100 kg N ha⁻¹ yr⁻¹), and intensive (≥100 kg N ha⁻¹ yr⁻¹) (Overmars et al., 2014; Dou et al., 2021). Missing fertiliser values were conservatively treated as zero application and classified as very extensive management. The absolute thresholds were applied consistently across all regions and crop types to ensure comparability of management intensity classes at the global scale. We did not include other chemical inputs such as pesticides because pesticide use is often associated with agricultural intensification and tends to co-vary with fertiliser application rates (Pergner and Lippert, 2023). Including both variables would therefore introduce redundancy. The irrigation dataset at 0.05° resolution (Mehta et al., 2024) was also used to distinguish intensive management with and without irrigation. The irrigation dataset at 0.05° resolution (Mehta et al., 2024) was also used to distinguish intensive management with and without irrigation. The information on irrigation presence and absence was first extracted to generate a spatial binary map, the grid cells with an availability of irrigation being assigned value 1, while those without irrigation were assigned value 0. This irrigated and rainfed information was then combined with the intensive class of the N fertiliser application map, prepared in an earlier step, in order to divide the intensive management class into two; intensive irrigated and intensive rainfed.

The spatial extent of pasture land was derived from the HILDA+ dataset, ensuring consistency with the broader land-use and land cover classification (though note that the underlying data is lower-resolution for pasture than for other land use types). To create the management intensity dataset, we combined data for synthetic (NO3-, NH4+) fertiliser and manure (organic N), giving total N fertiliser for pastures (Table 1). To convert the total N data into application rates, the fraction of managed pasture within each grid cell was calculated by spatially aggregating the managed pasture mask.

Grid-cell areas were computed as a function of latitude and total N fertilisers were converted to application rates per unit pasture area (kg N ha⁻¹ yr⁻¹). Cells without pasture were excluded from the calculation. Pasture management intensity was classified into three intensity categories based on the N application rates: very extensive (<50 kg N ha⁻¹ yr⁻¹), extensive (50–100 kg N ha⁻¹ yr⁻¹), and intensive (≥100 kg N ha⁻¹ yr⁻¹). The absolute thresholds were applied consistently across all regions and pastures to ensure comparability of management intensity classes at the global scale. Similar to cropland, the pasture fertiliser inputs (5 arcmin) were resampled to map onto a 0.01° grid.

The forest categories considered in this study included needle leaf evergreen, broad leaf evergreen, needle leaf deciduous, broad leaf deciduous, and mixed forest. Forest management intensity was determined by combining the HILDA+ forest management dataset (Winkler et al., 2021) with the forest management dataset of Lesiv et al. (2022), both of which were projected at 0.01° resolution using a majority method. Forest areas were classified into intensive, extensive and unmanaged based on the level of human intervention and management practices. Planted forests were considered as intensively-managed, naturally regenerating forests with signs of management, such as selective logging, or other forestry interventions, were considered as extensively-managed (Lesiv et al., 2022) and forests with no substantial signs of active human intervention (i.e., no or minimal direct management inputs) were considered as natural forests with minimal or no active management.

Spatial information on PV farms for the year 2020 was aggregated from high resolution raster data (20 m) to the harmonised land management intensity dataset at 0.01° resolution (Li et al., 2025). Individual PV tiles were processed sequentially and reprojected onto the 0.01° grid using average resampling. The PV fraction of underlying high-resolution cells covered by PV installations, resulted in a continuous PV coverage fraction between 0 and 1. Using these data, we calculated a total PV land area of 8432 km². PV cell allocation was restricted to non-water, non-urban, and non-forest areas and allocated by ranking candidate cells by descending PV fraction and allocating PV cells until the total PV area was achieved using latitude-adjusted cell areas.

Finally, the land management intensity dataset for year 2020 was produced on 0.01° global grid using a deterministic, rule-based mapping scheme designed to preserve broad land-use and land cover classes from HILDA+ while adding finer details on the management intensity. The land management intensity dataset relies on some low-resolution input datasets (e.g., crop and pasture fertiliser and irrigation), while some of the input datasets were at finer resolutions (e.g., forest management, forest types, etc.). Therefore, the final dataset represented a harmonised integration of varying underlying input data. Water and urban were mapped directly from HILDA+, while cropland, pasture and forest classes were assigned with the management intensity levels derived from respective land management layers. Conservative management intensity was applied where management information was missing or invalid. The cropland cells were assigned to an “other crops” class unless defined crop types and management intensity levels were available. Managed pastures were assigned as very extensive management unless classified otherwise. Forest cells were assigned to extensive management for the respective forest type, when the management intensity information was not present.

We performed a systematic sensitivity analysis by varying the thresholds used to define the management intensity classes (baseline case: 50 kg N ha⁻¹ yr⁻¹ and 100 kg N ha⁻¹ yr⁻¹) for a range of possible threshold combinations i.e., by varying the lower threshold between 40 and 60 kg N ha⁻¹ yr⁻¹ and the upper threshold between 80 and 120 kg N ha⁻¹ yr⁻¹ resulting in nine combined scenarios. For each combination, the spatial distribution of management intensity was quantified (i) the global share of cropland under each management class and (ii) spatial agreement with the baseline thresholds (50 kg N ha⁻¹ yr⁻¹ and 100 kg N ha⁻¹ yr⁻¹). The spatial agreement was computed as the percentage of grid cells for which assigned management intensity class remained unchanged relative to the baseline classification. The sensitivity analysis showed that the management intensity classification remained broadly stable across alternative threshold combinations (Fig. A4; Tables A1–A2). For cropland systems, spatial agreement with the baseline classification ranged between 89.7 % and 96.5 %, while pasture systems showed even higher agreement ranging from 97.7 % to 99 %. Although changes in thresholds altered the relative shares of very extensive, extensive, and intensive classes, the overall spatial patterns remained largely consistent.

To examine the association between socioeconomic and biophysical conditions and land management intensity, we employed multinomial logistic regression at the grid cell level. This analysis was based on an explanatory design-setup for individual land sectors i.e., cropland, pasture and forest. All the grid cells, that had valid management intensity and explanatory variables values were considered in the analysis. The same approach was applied separately within each World Bank income group. To evaluate that the socioeconomic variables do not inculcate multicollinearity, a pairwise correlation coefficient and the variance inflation factor (VIF) were computed for all the variables and the land-use types. Separate models were developed for cropland, pasture and forest systems, with management intensity (very extensive, extensive, and intensive) as the dependent variables. Very extensive management was used as the reference category in the model. Independent variables comprised GDP per capita, Gini index (income inequality), market accessibility and working age population, giving indicators of the primary socioeconomic capital groups of financial, social, manufactured and human capital (Perkins et al., 2025). The biophysical variables comprised aridity index (Zomer et al., 2022) and elevation (ASTER GDEM V3, 2018). All these datasets were harmonised to the same spatial resolution of 0.01° and for the year 2020 as the land management intensity dataset. These predictors were standardised prior to the use in the regression model to allow a direct comparison across the variables. The individual regression models were developed to analyse the variations both globally and in the World Bank income groups (lower-middle income, upper-middle income, high-income and low-income). We also computed the average marginal effects, which quantify the change in the predicted probability of each management intensity category associated with one standard deviation change in each variable. Marginal effects were computed for all explanatory variables and management intensity classes across the land systems and income groups. In addition, to assess where socioeconomic drivers explain land management intensity well, and where they do not, we computed spatial residuals between observed land management intensity and predicted class probabilities derived from the multinomial logistic regression models. For each land sector and management intensity class, residuals were calculated as the difference between the observed binary management class membership and the model-predicted probability at each grid cell. Positive residuals indicate locations where a management intensity class occurs more frequently than predicted by socioeconomic variables alone, while negative residuals indicate under prediction.

We performed a robustness check by using a spatial thinning method to evaluate whether spatial clustering has a role in the regression relationships. We performed an analysis using progressively coarser resolution samples such as 0.10, 0.25, and 0.50° to compute the regression for an individual set-up. The direction and magnitude of marginal effects were compared against the original resolution model (0.01°). The outcome indicated that marginal-effect relationships retained their direction across progressively coarser spatial samples (Figs. A8–A10; Table A6). In particular, GDP, market access, aridity index, and elevation effects remained largely stable across cropland, pasture, and forest systems, although effect magnitudes varied under coarser thinning. Mixed responses occurred mainly for selected pasture and forest relationships, indicating that some associations are more sensitive to spatial aggregation. However, such influences were very limited. Overall, the results suggest that the main regression patterns are not solely driven by spatial clustering of neighbouring grid cells.

To assess the consistency between land management intensity and land-use decision-making behaviour, we harmonised management intensity classes with decision-making types derived by Malek and Verburg (2020). The mapping was based on the conceptual correspondence between decision-making strategies, which reflected actors' objectives, capacities, and constraints, and observable land management outcomes. Decision-making types associated with market orientation, capital access, and production optimisation including professional intensifiers and professional commercialists were mapped to intensive or extensive management classes (Table 2). These decision-making types corresponded to active engagement in land-use markets and adaptive management strategies, which were inferred as either high-input intensive systems or input-efficient extensive systems depending on biophysical and economic conditions. Decision-making types emphasising environmental concerns or multifunctionality, such as eco-agriculturalists, were mapped primarily to extensive management, reflecting lower input use and greater reliance on ecosystem processes. In contrast, decision-making types constrained by limited resources or risk tolerance such as subsistence-oriented smallholders and survivalists were mapped to very extensive management (Table 2). Several decision-making categories represented co-presence or hybrid strategies, where multiple behavioural types coexist within the same spatial unit. For these cases, we applied multiple management intensity classes (e.g. intensive, extensive, and very extensive). Decision-making classes with low classification confidence or insufficient information were excluded from the comparison. Spatial comparisons were then conducted at the grid-cell level. A match was defined when the observed land management intensity at a given grid cell fell within the set of possible management intensity classes associated with the corresponding decision-making type. This flexible matching approach allows for one-to-many relationships between behaviour and management outcomes, reflecting the fact that similar decision-making strategies may lead to different management intensities depending on local environmental, and institutional conditions. To evaluate the spatial consistency between the land decision-making types and the land management intensity, pixel-wise comparison metrics were computed including total matched (hit) cells, mismatches, and overall accuracy. Additionally, confusion metrics summaries were computed. It is to note that the decision-making dataset is derived using proxy variables such as GDP, irrigation and accessibility from earlier datasets and at a coarser spatial resolution. Similar variables were used in this study to perform the regression analysis which differed in terms of their role, spatial representation, and time coverage. The comparison between the newly developed land management intensity dataset and the decision-making types should therefore be interpreted as broad consistency between behavioural patterns and management intensity.

The global land management intensity dataset (Fig. 2) revealed spatial heterogeneity in cropland, pasture and forest management intensity at 0.01° resolution, enabling a spatially consistent global assessment of land management patterns (figures below are expressed as percentages of the entire global terrestrial surface, unless otherwise indicated) (Saxena et al., 2026). Approximately 79 % of the Earth's land area was found to be managed to some degree, ranging from very extensive to intensive management across cropland, pasture and forest systems. The remaining 21 % of the total land area is predominantly unmanaged and consists largely of sparsely and non-vegetated areas, and ice/snow covered areas with minimal human intervention. Intensive management occupies about 22 % of the global managed land area (i.e., cropland, pasture, and forestry), while extensive and very extensive management together accounted for 78 % (Table A3). Despite perceptions of widespread intensive management, most managed land globally remains under extensive or very extensive management (Table A3). Although, management intensity varied substantially across land-use and land-cover sectors. Croplands accounted for around 10.5 % of the global terrestrial surface, yet contained the highest share of intensive management among all land-use and land-cover sectors (Fig. 3). Of global cropland, 44 % was managed intensively comprising around 30 % irrigated management and 14 % rainfed management (Fig. A1). Intensive and extensive croplands were concentrated in densely populated and highly industrialised regions, where infrastructure, fertiliser use, and irrigation were widely available (Fig. A2).

The results revealed a sectoral contrast, where intensive management remained concentrated in cropland systems, while pasture and forest systems remained predominantly extensive or very extensive despite their large spatial extent. At the aggregate level, pasture and forest land dominated managed land areas, and together accounted for more than two-thirds of the terrestrial surface. Pasture land occupied around 36 % of the global terrestrial surface, but was predominantly managed at low intensity, with only 10 % under intensive management, 44 % under extensive management, and 46 % under very extensive or minimal management (Figs. A1 & A5). Forests covered around 31.3 % of the global land surface, of which only around 5 % was managed intensively, 40 % managed extensively and 55 % remained unmanaged (mainly in boreal and tropical regions) (Fig. A1). The mapped global distributions of land management intensity aligned closely with independent reference datasets (Fig. A5, Table A3), with deviations (on an area-basis) generally within ±3 percentage points of FAO and IPCC estimates, supporting the robustness of the mapped dataset.

Land management intensity revealed distinct and non-uniform relationships with world income groups, world regions, and countries across land-use sectors (Figs. 4–5 & A1).

Multinomial logistic regression models were used to quantify the relationship of socioeconomic (economic development i.e., GDP per capita, inequality i.e., GINI, infrastructure accessibility i.e., market access, and population density) and biophysical variables (elevation and aridity index – higher values are more humid and lower values more arid) with the probability of different land management intensity classes across cropland, pasture and forest systems. A pairwise correlation and the variance inflation factor (VIF) for all the variables and land-use types were computed to check for multicollinearity. A weak to moderate correlation (-0.09 to +0.37) and low VIF (+1.1 to +3.0) values were found to indicate the absence of multicollinearity (Fig. A6). Non-linear relationships were identified between socioeconomic variables and land management intensity, with differences in direction and magnitude across land-use and land-cover classes and world income groups (Table A4, Figs. A7–A9). Marginal effects were computed to quantify the changes in predicted class probabilities associated with a one-standard deviation increase in each socioeconomic variable (Fig. 6).

A comparison between the land management intensity dataset presented in this study and the decision-making types of Malek and Verburg (2020) (Table A3) provided insights into the behavioural dimensions of land management. The decision-making categories represented diverse farmer and land manager behaviours, ranging from survivalist and subsistence smallholders to professional intensifiers and eco-agriculturalists. Figure 7 shows the global spatial distribution of consistency with one or more plausible management intensity classes and mismatch between land management intensity and land-use decision-making types (more details in Methods). Across all land systems, a total of 32 822 016 valid grid cells were compared, of which, 22 074 882 grid-cells (68 %) showed spatial consistency with one or more plausible management intensity classes, while 32 % showed divergence. This level of consistency indicated that, despite conceptual differences, spatial patterns of land management intensity are broadly consistent with important dimensions of land-use decision-making at the global scale. Large regions of consistency were observed in intensively managed cropland systems, including parts of North America, Europe, East Asia, and South Asia, where high-input management aligns with decision-making types characterised by commercialisation and intensive management. Consistency was also evident in regions dominated by very extensive or extensive land systems, such as parts of boreal Eurasia, central Africa, and arid and semi-arid regions where low-input management corresponds to decision-making types characterised by limited intervention. In contrast, spatial mismatches were concentrated in heterogeneous landscapes, where land-use decisions and management intensity vary over small spatial extents. These included regions with mixed land-use systems, transitional management regimes, or overlapping decision-making strategies. In such cases, similar decision-making types may correspond to multiple management intensities or management outcomes may diverge from dominant behavioural classifications.