We determined the relative abundance of different PFTs in each Köppen climatic zone by convolving the European Space Agency's Land Cover Climate Change Initiative (ESA LC-CCI) Global Plant Functional Types Dataset (v.2.08) from Harper et al. (2023) with the Köppen Climate Classification System maps from Rubel et al. (2017). We also considered the estimates of C3/C4 grass leaf area fraction generated in the NACP MsTMIP simulations (Global 0.5-degree Model Outputs in Standard Format, Version 2.0, from Huntzinger et al., 2021) to separate the Grasses PFTs (S1 in the Supplement). These probabilities are used to determine the number of species belonging to a PFT when a Scene is generated.

For each climatic zone, we selected 15 random locations (60 sites total) where we downloaded two years (2020–2022) of ERA5-Land hourly meteorological data (S2 in the Supplement). Accumulated radiation was recomputed to instantaneous rates (Wm-2) and accumulated precipitation at hourly intervals. The selected variables were prepared as the inputs of the model Soil Canopy Observation, Photochemistry and Energy fluxes (SCOPE) (Van Der Tol et al., 2009) as in Li et al. (2023) (Table S2 in the Supplement).

BOSSE randomly generates sets of plant traits (i.e., radiative transfer model variables) for each pixel of a Scene, each representing a species individual or group of identical individuals. This can be constrained by PFT-dependent bounds or consider empirical relationships between traits (Table ). BOSSE preserves the covariance between most foliar traits by drawing them from a Gaussian Mixture Model (GMM) trained over a 16 703 samples database generated by combining and gap-filling different spectral libraries and the TRY database (Kattge et al., 202) (S3 in the Supplement). The maximum carboxylation at optimal temperature (Vcmo) rate is then predicted as a PFT-dependent function of chlorophyll content (Cab). Most structural parameters are sampled from uniform distributions within global or PFT-dependent bounds. Canopy height (hc) is derived as a function of leaf area index (LAI).

For computational feasibility, BOSSE uses emulators (i.e., 2-hidden layer neural networks; Gómez-Dans et al., 2016) of SCOPE model (Van Der Tol et al., 2009) to simulate RS imagery and most of the ecosystem functions (Table ). The emulators predict hyperspectral R, hyperspectral F, and LST imagery as a function of plant traits, soil properties, meteorological data, and sun-view geometry (S4–S6 in the Supplement). Additional emulators retrieve optical traits (RS-based estimates of plant traits) from sun angles and R, which can also be tailored to specific RS missions (i.e., EnMAP, DESIS, and Sentinel-2 MSI) by convolving first R to their spectral response functions. An additional emulator predicts gross primary production, evapotranspiration, transpiration, sensible heat and soil heat fluxes, net radiation, and total and green light-use efficiency; ecosystem respiration is approximated with the semi-empirical model of Migliavacca et al. (2011) and used to compute net ecosystem productivity. Additionally, BOSSE computes friction velocity for the later computation of ecosystem functional properties (Gomarasca et al., 2024)

The emulators were trained over a database of simulations generated with SCOPE 1.74 using the full model (including the energy balance and photosynthesis) to ensure that spectral variables were linked to the simulated plant physiology. Vegetation traits were randomly generated (1) using a GMM trained over a plant trait database (Sect. ) and (2) using a Latin Hypercube Sampling approach. Moreover, we developed a semi-parametric model (a 2D interpolator) predicting soil resistance for evaporation from the pore space as a function of the relative soil moisture (Wr, the ratio volumetric soil water content (SMp) to soil field capacity (θfc)) and a parameter determining the sensitivity of the response (S5 in the Supplement). Realistic meteorological conditions were randomly drawn from an additional GMM trained over a 10⁵-size dataset subsampled from the BOSSE meteorological dataset (Sect. ).

We did not only test the performance of the RS emulators to predict different RS variables on independent datasets but also assessed the impact of these uncertainties on the derived functional diversity metrics (Table S4 in the Supplement). These analyses suggested the emulator's predictions align with the expected uncertainties of the different RS products and that the fact of using emulators does not bias or jeopardize the computation of functional diversity metrics. Similarly, the emulators' capability in predicting ecosystem functions was within the uncertainties expected from eddy covariance measurements (Table S5 in the Supplement). While the emulators carry certain epistemic uncertainty, the underlying models also carry such uncertainty. As the models, the emulators establish sufficiently formal relationships between the plant traits and the RS variables used to estimate functional diversity metrics and the fluxes used to compute ecosystem functional properties, thus enabling the benchmark of methods exploring the relationships between those variables. BOSSE also allows adding random uncertainty to the simulated RS variables to match specific Gaussian noise levels.

BOSSE uses the Growing Season Index (GSI) phenological model (Forkel et al., 2014) that defines vegetation phenology as a function of its response to water availability (Fig. a), light (i.e., incoming shortwave radiation, Fig. b), as well as cold (Fig. d) and heat (Fig. e), determined by air temperature. The combination of these functions over local meteorology (Fig. c) leads to the combined response of each species (Fig. f). GSI predicts vegetation trait values at any time as a function of the boundaries assigned to each pixel and meteorology (S7 in the Supplement). BOSSE assigns the PFT-dependent values parametrized by Forkel et al. (2014) and adds random Gaussian noise to generate variability between species and between traits. The GSI runs over the 30 d averaged meteorological values, and the rate of change of the GSI index is limited for each PFT to avoid unrealistic sudden changes in the vegetation state.

The BOSSE class is initialized by passing two Python dictionaries containing the simulation options (Table ) and paths necessary to load all the models and constants required and to store the outputs.

During the initialization, BOSSE loads all the necessary datasets and models and preallocates 3D matrices representing the 2D Scene with the third dimension containing predefined values for all the plant, soil, and meteorological variables the model uses. BOSSE also establishes upper and lower bounds for each variable.

First, BOSSE sets the PFT-dependent information regarding:

The PFT and relative abundances corresponding to the selected climatic zone

Next, it loads the following models:

The GMM used to draw random samples of correlated plant traits.

Once BOSSE has loaded all the necessary variables and models, it can initialize a given Scene, which is controlled by the options provided when the BOSSE class is initialized (Table ). The only necessary input argument is an integer number that identifies the meteorological dataset to be loaded. This number is used by default to seed the BOSSE class random number generators to ensure reproducibility; alternatively, the user can provide a different seed number and a low bound for the maximum mean LAI simulated in the Scene. By default, it is larger than or equal to 1.0 m2m-2 to generate Scenes with a certain amount of vegetation. However, this value can be lowered to represent barely vegetated Scenes. The random number generator is seeded several times during the Scene initialization to maximize the comparability of the species' plant traits for the scenes generated with different spatial patterns.

First, the time series of ERA5-Land meteorological variables are loaded for the site. Units are converted to those of the SCOPE model when necessary. Since ERA5-Land timestamps are UTC, the first and the last day might be incomplete in local time; therefore, these are gap-filled data from the nearby day, so they feature 24 hourly values at local time, which is the time used by BOSSE. Next, unfeasible negative values are interpolated and, if persistent, truncated. Wr, vapor pressure deficit, and timestamps in different formats (e.g., Day of the Year) are then computed. Finally, BOSSE calculates daily and 30 d averages and midday values and stores them in separate data frames; also, daily potential evapotranspiration is calculated using the Python package “pypyet” (Vremec et al., 2023) according to Penman-Monteith (1965).

The random number generator is seeded for the first time. Then, BOSSE randomly generates unique values of the SCOPE soil reflectance model parameters and field capacity (θfc, –) for the Scene. θfc is set within 100 %–150 % of the maximum volumetric moisture content (SMp, %] of the time series, and water availability (Wp, –) is updated according to the new θfc.

Next, BOSSE uses the default PFT-dependent parameters of the phenological model and Wp to calculate the GSI time series of each PFT, utilizing the default coefficient values from Forkel et al. (2014). BOSSE filters out the PFTs that are poorly adapted to the Scene's climate. To do so, it compares the percentile 95 (or 5 for the warm temperature) of each meteorological variable with the corresponding GSI “base” parameter that determines the inflection point of the phenological response function to meteorology (see S7 in the Supplement). BOSSE removes the PFTs from the Scene where the most extreme meteorological conditions do not cross this point, assuming these are too stressful for such species to develop in competition with other species. If no PFT meets these criteria, BOSSE keeps only the most competitive PFT (reaching the largest phenological index values).

Then, some or all of the remaining PFTs are randomly selected for the Scene, and individuals of different species are assigned to the scene pixels. BOSSE runs at a 100 % spatial resolution; therefore, each pixel is occupied by a single phenotype (a single individual or identical individuals of the same species), meaning a unique set of vegetation parameters. The number of species is randomly sampled from the range between 1 and the input “S_max” (maximum species richness). BOSSE assigns these species to the PFTs and, together with foreseen relative abundances and the spatial pattern configured (“clustered”, “intermediate”, or “even”), the Neutral Landscape model (Etherington et al., 2015) distributes them in space. Neutral Landscape produces spatial patterns of a continuous variable that is then classified to determine the pixels belonging to each species, trying to respect the foreseen abundances, which must be recomputed later. BOSSE filters this spatial pattern generated using a Gaussian kernel of standard deviation randomly selected between [0.5, 1.5] pixels. This smoothed map is then used to scale the soil reflectance model variables (B, lat, and lon), generating a spatial variability of soil properties that is partly correlated with the species' location.

Plant trait values are randomly generated, either using the GMM for the leaf radiative transfer variables (N, Cab, Car, Cant, Cs, Cw, Cdm) truncated within PFT or species-dependent bounds (S3 in the Supplement) or randomly sampling within those bounds (LIDFa, LIDFb, LAI, lw, and mBB) (Table ). Moreover, Vcmo is computed as a function of Cab using PFT-dependent relationships from Luo et al. (2019), and canopy height (hc) is computed using the allometric relationship with LAI and PFT-dependent scaling factors described in Jones (1998).

BOSSE determines the phenological maximum and minimum values for the plant traits per pixel, considering that the bounds of some plant traits are PFT-dependent (e.g., LAI, Cab, and mBB). In each scene, BOSSE randomly generates a set of plant traits and a range of variability between 10 % and 40 % of the global PFT range. The bounds above and below this “central” value are used to generate random sets of mean plant traits for the species of each PFT (Fig. a). Then, the interspecific variability is calculated from the weighted average of the plant traits of all the species included, and the intraspecific variability is determined assuming it ranges between 20 % and 40 % of the interspecific variability (Fig. 1 in Albert et al., 2010) (Fig. b). Using each initial mean species plant trait values and the intraspecific variability, the plant traits of each species individual (i.e., pixel) are randomly determined. This process can bias the resulting species' mean from the initial mean value. Also, any value outside the global bounds of the above and below-average site ranges are truncated if they exceed the global trait bounds. Finally, each pixel's phenological maximum and minimum values are produced by adequately assigning the largest or the lowest value to the maximum or minimum phenological moment (Fig. c). For example, pigments and structural variables maximize in the green peak (GSI = 1), but senescent pigments content (Cs) are assumed minimum during that period. Also, BOSSE ensures that LAI, Cab, carotenoids (Cca), and anthocyanins (Cant) content of deciduous species equal zero at the bottom of the phenological period (GSI = 0) and reduces the range of variability of all variables for the evergreen PFTs. It also ensures that the Cs only develop after the green peak (maximum GSI of the year). At any time (t) of the simulations, the plant trait values are determined by interpolating between these phenological bounds with the corresponding phenological state determined by the GSI. When LAI = 0, plant traits are set to 0, which improves the emulators' performance. The whole process is represented for a given individual in Fig. d.

Then, the GSI phenological model is run using 30 d averages of the meteorological variables. Starting from the PFT parameter values, different parameters are generated for each species and plant trait by adding random Gaussian noise (5 % for the slope (sGSI) and the base (bGSI), and 1 % for the legacy sensitivity (τGSI), S7 in the Supplement). This allows species of the same PFT to behave similarly but prevents all the plant traits from synchronizing precisely.

BOSSE runs GSI using the incoming shortwave radiation and air temperature from ERA-5 Land. Following Forkel et al. (2014), water availability must be calculated from the relative soil water content, a PFT-defined maximum transpiration value (Table 1 in Gerten et al., 2004), potential evapotranspiration, and a PFT-dependent potential canopy conductance value from Leuning et al. (2008).

Moreover, to prevent unrealistically fast changes in the vegetation's phenological status, BOSSE includes some PFT-dependent arbitrary limits to the rate of change of the GSI index of each of the response functions of the model (Table S6 in the Supplement). During the calculation of the GSI index, these values truncated any change of the index, taking into account the time step length of the change.

BOSSE uses the model described and parametrized by Migliavacca et al. (2011) to predict ecosystem respiration (Reco) and takes the PFT-dependent parameters and their associated standard errors reported in the abovementioned manuscript (S8 in the Supplement). The model scales Reco with the maximum seasonal leaf area index, which is determined at this point for each pixel. This is done by multiplying the maximum species-dependent GSI value assigned to LAI and the LAI upper bound assigned to each pixel.

Then, for the rest of the model parameters, BOSSE first generates the interspecific variability by random sampling parameter values of a Gaussian distribution with a mean equal to the adjusted PFT-dependent value and standard deviation equal to the reported standard error in Table 5 in Migliavacca et al. (2011). Once the model parameters are defined per species, intraspecific variability is generated by adding random noise drawn from a uniform distribution, assuming traits' intraspecific variability ranges between 20 % and 40 % of the interspecific variability (Fig. 1 in Albert et al., 2010).

BOSSE uses the precomputed plant trait bounds and the GSI values to generate plant trait maps at any time (t) of the meteorological time series. Each species and trait GSI value (GSIsp,PT∈[0,1]) is used to linearly interpolate the corresponding trait value between the bounds corresponding to the upper (PTupper) and lower (PTlower) points of the phenology, assigned for each trait and pixel (pix) of the Scene (Eq. ). The fact that all the necessary variables are precomputed during initialization ensures the repetition of these simulations. Plant trait map resolution is degraded using a regular grid, mimicking field sampling schemes. 1PTt,pix=PTlower,pix+(PTupper,pix-PTlower,pix)×GSIt,sp,PT,

BOSSE uses instantaneous meteorological data and vegetation and soil traits to predict instantaneous gross primary production (GPP), total latent heat flux (λE), canopy latent heat flux (λEcanopy), sensible heat flux (H), net radiation (Rn), soil heat flux (G), light use efficiency (LUE), and LUE computed with the photosynthetically active radiation absorbed by chlorophyll (LUEgreen) using a SCOPE dedicated emulator. Negative GPP, LUE, and LUEgreen are allowed to prevent biases in averages and ecosystem functional properties, which is acceptable since these values are present in observational datasets due to uncertainty. The analysis of the simulated fluxes might be subject to quality analysis and filtering, as the observations.

In addition, Reco is computed also using the model described in Migliavacca et al. (2011). The model is parametrized to predict daily Reco, using daily mean air temperature, daily GPP, and monthly precipitation. Therefore, the model is fed with the values averaged for those periods around the timestamp where respiration is computed. Reco is converted to µmolCm-2s-1 to calculate net ecosystem productivity (NEP = GPP-Reco); additionally, Reco,15°C at the reference temperature of 15 °C is also computed. Finally, BOSSE also calculates the friction velocity (u∗) following Wallace and Verhoef (Wallace and Verhoef, 2000), as done in the SCOPE model. The ecosystem functions can be provided as a map or as the average value for the Scene. The functions can be provided as the average value of the Scene, as well as maps so that other integration approaches (e.g., climatology footprints; Kljun et al., 2015) can be applied.