Quantifying the benefits of reducing synthetic nitrogen application policy on ecosystem carbon sequestration and biodiversity

Overview of modeling framework

We have used a range of econometric, economic, and agricultural land surface models to analyze the factors driving land-use change in order to assess their ecological, agricultural, climatic and economic impacts. These multi-scale models differ in their methodologies, scale of interest, and resolution, but they are very complementary and could provide a unique opportunity to analyze public policy scenario effects on land-use and resulting changes in ecosystem carbon and biodiversity.

Among these models, the economic land use model Nexus Land Use (NLU)^29,30 and the agricultural supply-side model Agriculture, Recomposition de l’Offre et Politique Agricole (AROPAj)³¹ coupled with a spatial econometric model³² have allowed us to estimate the impact on EU land-use of a scenario involving a 50% reduction in N synthetic fertilizers compared to a baseline scenario. In the present study, we use these land-use scenarios to force ORCHIDEE-crop (Organising Carbon and Hydrology in Dynamic Ecosystems), an agricultural land surface model^16,33 and Projecting Responses of Ecological Diversity in Changing Terrestrial Systems (PREDICTS)³⁴, a biodiversity model to simulate, respectively, ecosystem C and biodiversity changes across the EU covering the domain 35.25°N and 69.25°N in latitude and 9.25°W and 34.25°W in longitude. The schematic (Fig. 1) provides a brief overview of the modelling framework applied in this study.

In order to link the land use output data from the AROPAj and NLU models with the ORCHIDEE-crop and PREDICTS models, the first step is to match land uses and crops between the models (see Table 1). AROPAj and NLU crops are classified into ORCHIDEE-crop plant functional types (PFTs): C3 winter and summer crops, C4 summer crop and C3/C4 natural grass (see “Model descriptions” section for a detailed description of ORCHIDEE-crop PFTs). The AROPAj and NLU crops are also classified into the PREDICTS crop types: annual, perennial, N-fixing. The AROPAj and NLU “rangeland” and “pasture” categories are found in PREDICTS but in ORCHIDEE-crop they are considered to fall within the C3 natural grassland PFT. Finally, NLU and AROPAJ forest and other natural areas are classified as “primary” natural areas (with low anthropogenic use) or “secondary” (intermediate to high anthropogenic environmental use) according to the land use map of these areas³⁵. For ORCHIDEE-crop, they are classified as natural forest PFTs. Note that the fallow areas described in AROPAj that are part of crops are classified as “grass” PFT in ORCHIDEE-crop and as “minimum” intensity annual crops in PREDICTS.

The land-use and land cover changes described in the following sub-section are used as inputs to ORCHIDEE-crop and PREDICTS from both the NLU and AROPAj models’ output.

Land-use change scenarios

Land-use changes in the EU are simulated for the present day using two scenarios: (1) a business as usual scenario (Baseline) and (2) a scenario involving a policy to reduce mineral nitrogen use by 50% from the Baseline (Halving-N). The land-use changes in Halving-N and Baseline are computed by both NLU and AROPAj models. In the latter model, the computed land-use changes result from coupling between AROPAj and a spatial econometric model. Since there are differences in the nature of the models (supply-side model versus partial equilibrium model) and their underlying data, the Baseline scenarios in the NLU and AROPAj frameworks are different. A detailed description of the differences and a discussion of their implications on the production and area of different land-uses is provided in Lungarska et al.³⁶. EU plant production is 370 and 383 MtDM (Million tons of Dry Matter) respectively based on the application of 12 TgN (Tera grams) of N fertilizer in AROPAj and NLU. Crops, grasslands, and forests cover respectively, 116, 57 and 234 Mha in NLU and respectively 94 (including fallow land), 38 and 142 Mha in AROPAj. In AROPAj and NLU, the 50% N reduction is achieved indirectly by increasing the N input price from present-day figures³⁶.

The land-use changes output from AROPAj and NLU are supplied as inputs to the ORCHIDEE-crop and PREDICTS models. The land-use changes are matched with corresponding plant functional types (PFTs) in ORCHIDEE-crop and land-uses in PREDICTS (see Table 1). “Model descriptions” section provides a detailed description of the ORCHIDEE-crop and PREDICTS models.

Model descriptions

Here, we describe the ORCHIDEE-crop and PREDICTS models that quantify the impacts of halving N fertilizer consumption in the EU. Table 2 presents a brief overview of the two models.

A detailed description of ORCHIDEE-crop: This model is a process-based agricultural land surface model that integrates crop-specific phenology based on Simulateur mulTidisciplinaire pour les Cultures Standard (STICS)^37,38. Carbon allocation is based on the plant-based hybrid model from the original ORCHIDEE allocation scheme³⁹ and a crop specific formulation of STICS providing leaf, root, and shoot biomass, grain maturity time, litter production, and litter and soil carbon decomposition. The harvest date is calculated after grains reach maturity⁴⁰. The ORCHIDEE-crop model has no explicit nitrogen cycle but accounts empirically for the effect of N fertilization by increasing the maximum Rubisco- and light-limited leaf photosynthetic rates as a function of the amount of N applied, using a Michaelis–Menten function⁴⁰. Also, ORCHIDEE-crop is calibrated against observations, which showed a good match between modeled observed aboveground biomass, crop yield, and daily carbon⁴⁰. This version of the model currently uses three crop PFTs: C3 winter, C3 summer and C4 summer. Forests are classified as Broadleaf, Needle leaf, Deciduous, Temperate and Boreal. Up to 11 non-cropland vegetation types can co-exist with crops on a grid point of the model, according to prescribed land cover information. A gridded simulation of ORCHIDEE-crop requires 30-min time step meteorological forcing (air temperature, specific humidity, incoming shortwave and longwave radiation, rainfall), which can be interpolated in time from gridded climate analysis data or atmospheric models. In this study, this model is used to quantify the ecosystem C variables.

A detailed description of PREDICTS: The PREDICTS database was collated by searching the published literature for studies where terrestrial biodiversity (including plants, fungi, vertebrates, and invertebrates) was sampled using consistent methods across multiple sites, which vary in the pressures faced. The land use and intensity of each site have been assessed and categorized in a consistent way^41,42,43. Authors of studies were contacted to ask for the raw biodiversity data where this was not already available^41,42. Most records in the PREDICTS database refer to the number of individuals of a species at a site; this makes it possible to compute a range of biodiversity indices. To estimate biodiversity responses to human impacts across such a global and heterogeneous dataset, linear mixed-effects models are used; random intercepts account for differences in biogeographic factors, sampling methodology and taxonomic focus, and the spatial layout of sites within studies. Using the PREDICTS database to assess the impact of human pressures on biodiversity assumes that space-for-time substitution is valid⁴⁴; it assumes that the sites have reached equilibrium and so the impact of pressures on biodiversity over time can be observed across space and that the relationship between biodiversity and drivers do not vary over time.

SR is calculated as the number of species at each site; it is a widely used measure of biodiversity and is both simple and intuitive. Responses of SR to land use and intensity were modelled using generalized linear mixed effects models and with a Poisson error structure; an observation-level random effect was included to account for overdispersion⁴⁵. This model is then used to project SR in each grid of a 0.5° map and expressed as a percentage of the SR level in primary vegetation from land use harmonization map³⁵.

To estimate BII change with land use and intensity, two models are required. Total abundance was first calculated as the sum of all individuals at each site; it was then rescaled within the study (so that the maximum within a study is 1) and was square-root transformed before modelling as a function of land use and intensity, to account for non-normality of the model residuals (a Poisson error structure could not be used as abundance data can include non-integer data e.g. densities). Inclusion of a random slope for land use within the study was supported (based on Akaike’s Information Criterion). Compositional similarity was then calculated as the asymmetric Jaccard index, comparing each baseline site (primary vegetation) with all other sites, and logit transformed with an adjustment of 0.01 (to account for non-normality of the model residuals). Compositional similarity was then modelled as a function of land use and intensity (coarsened so that only perennial crops were allowed to differ across intensities), including the environmental and geographic distance between sites as control variables, whose effects were permitted to differ among land use and intensity levels (these variables were cube-root and log-transformed respectively to improve residual distribution). To calculate BII, total abundance (expressed as a percentage of their level in primary vegetation) and compositional similarity (expressed as a percentage of their level in primary vegetation)⁴⁶ are projected for each grid of a 0.5° map; these two maps are then multiplied to give abundance-based BII¹⁹. The PREDICTS models include different levels of management (intensive, light or minimal) and different types of land cover (forest, pasture, rangeland, annual cropland, perennial cropland, and urban zones). The coefficients of these mixed-effect models and a detailed description of the link between the PREDICTS models and NLU are available in Prudhomme et al.⁴⁶. The spatial predictions of biodiversity were computed using a python pipeline, which was developed specifically for the PREDICTS project (https://github.com/ricardog/raster-project).

In our modeling framework, the impact of halving N fertilizer goes through two steps: (i) we calculate the effect of this reduction of N fertilizer on agricultural yield, and (ii) calculate the effect of the yield reduction on biodiversity. By keeping yield as a proxy of agricultural land use intensification as proposed in Prudhomme et al.⁴⁶, we include not only the direct effect of the reduction of N fertilizer on biodiversity but also the effects correlated to this reduction of N fertilizer such as the reduction of other chemical inputs (P and K fertilizers and pesticides). While the effect of the change in N fertilization on yield is calculated by the classical concave production function in agronomy²⁹, the effect of the change in yield is calculated by coupling the NLU land use model and the PREDICTS biodiversity model⁴⁶. For each category of crops (annual, perennial, leguminous), the coupling consists of estimating (using a Generalized Additive Model [GAM]) the share of each intensity class (minimum, light, intense) as a function of the average calorie yield based on the average crop yield maps from a plant growth model. The maps describing the share of land use intensities are from Newbold et al.¹⁹ Similarly for pasture, the share of each intensity class (light, intense) is estimated with the help of a GAM as a function of ruminant density.

Simulations

Our experimental design focuses on assessing the effects of a 50% reduction in present-day N fertilizer use levels across the EU. The choice of halving N fertilizer in EU agriculture is related to the “Farm to Fork” strategy, which puts forward the ambition for 2030 to reduce nutrient losses to the environment from both organic and mineral fertilizers by at least 50%. The results from NLU (and its nitrogen balance module) show that this level of reduction corresponds to a 50% reduction in nutrient losses (nitrogen and phosphorus) aimed by the Farm to Fork strategy as a part of the European Green Deal. AROPAj models exclusively the EU countries (in 2012, there were 28 member states) while NLU simulations cover the EU and the rest of the world (EU being a part of the European region as represented by the model). However, the N reduction policy implemented in the EU alone and the comparison of the results conducted only for the EU. All EU member states are considered but for some of them we present results. A total of four simulations corresponding to four land-use maps (two from AROPAj and two from NLU, see Fig. 1) are performed in the ORCHIDEE-crop model and also in the PREDICTS model. In addition to changes in the area of different land-uses, changes in mineral N input are accounted for in both models. However, changes in organic N input and crop rotations are not accounted for. In ORCHIDEE-crop 55% of the carbon harvested from croplands is exported but the remaining residues are returned to the soils.

ORCHIDEE-crop simulation details: the model simulations are performed over a domain covering the EU. Four idealized simulations are carried out using the ORCHIDEE-crop model by forcing present-day meteorological data (2006–2010), levels of N fertilizer (150 KgN/ha) and atmospheric CO₂ concentration (385 ppm). The four simulations include Halving-N and Baseline corresponding to AROPAj and NLU land-use scenarios (two ORCHIDEE-crop simulations per economic model). All four simulations start from the year 2010 climate and carbon cycle conditions with a recycled climate (2006–2010) for 150 years. For the year 2010, climate and carbon cycle conditions are obtained from the output of historical simulations. Historical simulations from the year 1901 to the year 2010 are performed for both AROPAj and NLU Baseline scenario land-use land cover maps. In addition, these historical simulations started from an equilibrium state of soil carbon, energy and water cycle variables corresponding to the year 1901. The 1901 equilibrium state is determined by running a 350-year spin-up simulation corresponding to a recycled climate (1901–1910). The observation-based climate forcing data from the Global Soil Wetness Project was only available starting from the year 1901. The drift in soil carbon over the last 100 years of the 350-year simulations is less than 1%. The equilibrium state simulations corresponding to the year 1901 were necessary to have stabilized biophysical and ecosystem C variables across the EU. Other forcing variables, e.g. atmospheric CO₂ concentration (296.57 ppm), N-fertilization rate (32 KgN/ha), harvest index (0.25), and also the phenology parameters for short-cycle variety winter and summer crops¹⁶ corresponding to the year 1901 were prescribed.

PREDICTS simulation details: the PREDICTS model represents changes in broad-sense biodiversity in different land-uses and intensities of land-use relative to a reference land-use (as the biodiversity metrics assessed include all terrestrial biodiversity for which data are present in the PREDICTS database including plants, fungi, vertebrates and invertebrates). Here the reference ecosystem is a primary natural ecosystem. Biodiversity changes are then reported as a percentage by dividing the obtained biodiversity levels by the level of biodiversity present in the primary natural ecosystem. This simulation is performed for each grid point on a map of the EU for land-use scenarios corresponding to Baseline and Halving-N for both economic models, AROPAj and NLU (Fig. 1).

Breakdown method for biodiversity and carbon changes

The Halving-N and Baseline scenarios provide contrasted land-use maps according to the assumptions of economic and land-use models³⁶. This results in different plant and animal production, and different land-uses at the European scale in each model. A price shock on inputs, as represented in the Halving-N scenario compared to the Baseline scenario, can induce (1) a spatial reallocation of production or (2) production changes⁴⁷. Here, we separate out the effects of these two mechanisms on biodiversity (species richness) and carbon indicators (NPP and soil carbon) by decomposing the overall environmental differences between the Halving-N and the Baseline scenarios. The breakdown is not possible for the BII indicator because this indicator is the product of two indicators: abundance and a similarity indicator of ecological communities.

First, we breakdown the carbon and biodiversity differences by land-use type. The breakdown for carbon is straightforward because the carbon changes are computed for each land-use. The biodiversity changes associated with each land-use are computed by setting no changes in the other PREDICTS model land-uses. The sum of the biodiversity changes for each land-use is thus equal to the overall change in biodiversity.

For each land-use i (forest, grassland and cropland), we separate out the carbon and biodiversity differences between the Halving-N and the Baseline scenarios into two effects in accordance with Eq. (1): (i) the carbon and biodiversity difference associated with the area difference—called “Area effect”, and (ii) the carbon and biodiversity difference associated with the difference in biodiversity and carbon sequestration per unit area—called “Intensity effect”. The “Area effect” corresponds to the change in carbon sequestration and biodiversity associated with a change in the land-use area. For example, a reduction in grassland area leads to reduction in the C sequestration and biodiversity associated with this area. The “Intensity effect” corresponds to a change in the C sequestration and biodiversity per unit area. For example, a reallocation of production toward places with high soil C content leads to an increase in the carbon stock per hectare or an increase in crop yield leads to a reduction in the biodiversity per unit of cropland. Thus, the “Intensity effect” corresponds to the effect of a production reallocation on C sequestration, and the effect of land-use intensity on biodiversity.

We use the Logarithmic Mean Division Index (LMDI) method, which breaks down the target values into several main influencing factors based on mathematical identity transformation⁴⁸ as follows.

(Delta {E}_{i}) is the difference in the environmental indicator between the Halving-N and the Baseline scenarios. Superscript ‘A’ denotes area effect and ‘I’ denotes intensity effect. Subscript ‘i’ denotes different land-use (e.g. forests, grassland, cropland etc.). (Delta {E}_{i}^{A}) is the difference in the environmental indicator between the Halving-N and the Baseline scenarios associated with the difference in area. (Delta {E}_{i}^{I}) is the difference between the Halving-N and the Baseline scenarios associated with the different intensity per unit of area of the environmental indicator.

({E}_{i}^{hN}) is the level of the environmental indicator in the Halving-N (superscript hN) scenario. ({E}_{i}^{b}) is the level of the environmental indicator in the Baseline (superscript b). ({A}_{i}^{hN}) is the area of land-use i in the Halving-N scenario. ({A}_{i}^{b}) is the area of land-use i in the Baseline

Equation (3) is same as Eq. (2) but for the intensity of the environmental indicator ({e}_{i}).

The breakdown of the differences in the environmental indicators is performed between the Halving-N scenario and the Baseline. A positive variation ((Delta {E}_{i}>0)) indicates a higher environmental indicator in the Halving-N scenario compared to the Baseline without implying any temporal variation since the scenarios compare the environmental indicator status in 2012 in the AROPAj and in the NLU land-uses. Conversely, a negative variation ((Delta {E}_{i}<0)) indicates a lower environmental indicator in the Halving-N scenario compared to the Baseline.

Source: Ecology - nature.com