Introduction

Biogenic volatile organic compounds (BVOCs) are critical drivers of biosphere-atmosphere interactions, significantly influencing ozone formation, secondary organic aerosol (SOA) production1,2, and broader atmospheric chemistry3,4. These processes have profound implications for air quality, climate, and human health. However, despite their significance, accurately quantifying BVOC emissions across diverse ecosystems remains a persistent challenge. This complexity stems from the intricate interplay of biological, climatic, and anthropogenic factors5,6, compounded by the dynamic nature of vegetation and environmental conditions.

Existing BVOC models, such as the widely used Model of Emissions of Gases and Aerosols from Nature (MEGAN2.1)7, have made significant strides by providing a comprehensive framework for emission estimation. However, these models often rely on coarse spatial data and static parameters, which limits their ability to capture the heterogeneity of urban and natural landscapes, leading to inconsistent and inaccurate BVOC emission estimates across cities and regions8,9,10,11,12. Such limitations are particularly problematic in rapidly changing environments, such as urban areas13,14,15, where vegetation types, land use, and greening strategies are highly dynamic16,17,18. Furthermore, computational constraints hinder the application of high-resolution models over large geographic scales, leaving critical gaps in our understanding of BVOCs’ impacts on air quality and climate19,20.

To address these challenges, we introduce GEE-MEGAN, an innovative modeling framework (Fig. 1 and Fig. S1 in the Supplementary Information) that combines the strengths of MEGAN2.1 with the cloud-based computational power of Google Earth Engine (GEE)21,22. By integrating multi-source remote sensing data (e.g., MODIS, Landsat, Sentinel) and machine learning23,24, GEE-MEGAN dynamically refines critical model inputs (Fig. 1) such as leaf area index (LAI) and plant functional types (PFTs), enabling near-real-time, high-resolution BVOC emission estimates at spatial scales ranging from 10–30 m locally to global domains. These advancements in accuracy, resolution, and computational efficiency position GEE-MEGAN as a transformative tool for advancing research on biosphere-atmosphere interactions, particularly in understanding the complex feedback loops between BVOCs, climate and air quality.

Fig. 1: Comparison of GEE-MEGAN and traditional biogenic volatile organic compounds (BVOC) calculation processes.
figure 1

Each panel illustrates a set step in the calculation of BVOC emissions. A Calculate the standard emission factor grid map (EFMAP) in MEGAN2.1 based on plant functional types (PFTs) and emission factor (EF); B Derivation of leaf area index (γLAI) and environmental activity factor (γA) in MEGAN2.1; C Calculation of temperature (γT) and light (γP) activity factors using meteorological data in MEGAN 2.1; D Dynamic standard emission factor grid map construction in GEE-MEGAN, incorporating land use/cover change (LUCC) and wildfire data. E Updated calculation of γLAI and γA in GEE-MEGAN using high-resolution LAI data from Landsat.

Full size image

Results and discussion

Better prediction of observed LAI and BVOC fluxes

GEE-MEGAN accurately captured both daily and monthly emission trends at regional scales compared to conventional models at identical spatial resolutions. For example, hourly isoprene emission rates for August 2010 across South America showed that daily average emissions closely aligned with trends derived from the inventory from MEGAN model coupled with the MOdel for Hydrocarbon emissions by the CANopy (MOHYCAN) and constraint by Ozone Monitoring Instrument (OMI) Formaldehyde (HCHO) observation (MEGAN2.1-MOHYCAN-OMI, details in Fig. S2a)25, achieving a Spearman Rank Correlation Coefficient of R = 0.83 (p < 0.001, N = 31). Monthly regional emissions also exhibited minimal discrepancies when compared to the MEGAN default and MEGAN2.1 models for South America7,25,26, with differences ranging from 6.4% to 12% (Fig. S2c–h and Table S1). These results validate GEE-MEGAN’s consistency with the core dynamics of MEGAN2.1 across large temporal and spatial scales. Notably, GEE-MEGAN consistently produced isoprene emission estimates approximately 50% higher than those derived from OMI-MEGAN and Global Ozone Monitoring Experiment-2 (GOME2) based MEGAN model25,27. This discrepancy is consistent with recent studies suggesting that OMI-MEGAN underestimates emissions due to biases in the OMI formaldehyde (HCHO) column data28, which likely skew emission retrievals downward.

GEE-MEGAN utilizes high-resolution Landsat data for LAI estimation, outperforming conventional approaches that rely on coarser datasets, such as MODIS (Fig. 1E vs. Fig. 1B). When compared to observed LAI data (https://calvalportal.ceos.org/lpv-direct-v2.1), GEE-MEGAN’s Landsat-based LAI estimates reduced the mean absolute errors (MAE) by 12.5% at 500-m resolution and by 11.7% at a 3-km resolution relative to MODIS-based estimates (Fig. S3), and reduced the RMSE by 9.1% at 500-m resolution and by 3.6% at a 3-km resolution. Based on these refined LAI inputs, GEE-MEGAN further improves BVOC emission estimates, as evaluated in Table S2. In comparison to observed BVOC fluxes across diverse ecosystems in North America, South America, Europe, and East Asia29,30,31,32,33 (Table S2), GEE-MEGAN outperformed MEGAN2.1, reducing RMSE by 21.6–48.6% and improving correlation coefficients (R) from 0.83 to 0.90. The most significant improvements were observed in regions with timely captured dynamic canopy changes.

Enhanced spatial resolution in original and disturbed forests

GEE-MEGAN captures greater spatial heterogeneity in BVOC emissions (Fig. 2a), especially at forest edges affected by human activities, containing emission hotspots and dead-zones, which are often fragmented or averaged out in coarser-resolution datasets34,35. Unlike conventional models that rely on static parameters and coarse spatial data, GEE-MEGAN dynamically integrates climatic norms and land use/cover change (LUCC) data with high-resolution satellite observations (Fig. 1D, Table S3), producing a more accurate and responsive emission factor map (EFMAP) capable of detecting abrupt changes.

Fig. 2: Isoprene emission simulations over the Amazon region in August 2019 using GEE-MEGAN and MEGAN-default.
figure 2

a Isoprene emission at a 30-m resolution for the Amazon region using the GEE-MEGAN, which incorporated a total of 2,211 Landsat scenes. b Simulation results obtained with the MEGAN-default. c Relative difference computed as [(b) – (a)]/(b). d–f Total isoprene emissions and their differences for the 2° × 2° (220 km × 220 km) grids, uniformly distributed along the red line (Arc of Deforestation), blue line (inner Amazon edge) and yellow line (with reforestation) as indicated in (c).

Full size image

In the Amazon rainforest, a critical BVOC source contributing over a quarter of global isoprene emission36, GEE-MEGAN reveals substantial discrepancies in BVOC emissions when compared to the MEGAN default with LUCC data from MODIS. Coarser models often fail to account for critical impacts of forest loss, introducing potential biases of Amazonian BVOC emissions. During the high-BVOC-emitting month of August in 2019, for example, GEE-MEGAN estimated 13% lower Amazonian isoprene emissions than the MEGAN default (Fig. 2a, b). While the discrepancies of 3.0–8.3% occurred deep inside the forest, they were especially pronounced at forest edges by forest degradation or deforestation (Fig. 2c). In these areas, the MEGAN default model overestimated emissions by up to 77% (average 31%) in the fifteen 2° × 2° grid cells along the Arc of Deforestation35,37 (red line in Fig. 2c; e), while the overestimation became smaller, ranging 3–35% (average 9%) in the fifteen 2° × 2° grids closer to the forest interior (blue line in Fig. 2c; d). Moreover, the MEGAN default model underestimated isoprene emission by up to 250% in some southeastern Brazil grids (yellow line in Fig. 2c; f), where recent reforestation or post-fire vegetation recovery occurred38,39.

In paired 2° × 2° grids along the Arc of Deforestation (red grids) and approximately ~300 km into the inner forest (blue grids, Fig. 2c), the MEGAN default model showed an average isoprene emission difference of 6% (maximum 23%). In contrast, GEE-MEGAN estimated an average difference of 35% on average (maximum 68%), demonstrating its strength in capturing emission heterogeneity. Given that each grid cell represents an area of nearly 50,000 km2, such large under- or overestimation of reactive BVOC emissions can significantly misrepresent their roles and broader environmental impacts, especially at local and regional scales.

High-resolution BVOC emissions modeling in urban areas

GEE-MEGAN effectively captures the spatial variability of urban green spaces, enabling high-resolution BVOC emission modeling at 10–30 m resolution, resampled to 100 m. This ability is showcased in Fig. 3, where BVOC emissions for July 2019 in London, Paris, Los Angeles, and Beijing are quantified at finer urban scales by GEE-MEGAN, compared to the coarser 500-m resolution of the MEGAN default model. In Paris, the model identified key hotspots, and further evaluation using the BOXMOX chemical box model showed that GEE-MEGAN emissions achieved better agreement with observed isoprene concentrations than the default MEGAN model under identical ERA5 meteorological conditions9 (Fig. S4, Note S1 and Table S4). This result provides strong evidence that enhanced spatial detail in emission inputs improves model-observation consistency.

Fig. 3: Spatial distribution of isoprene emissions in Paris, Los Angeles, London, and Beijing simulations for July 2019.
figure 3

a, d, g, j Monthly mean isoprene emissions modeled by GEE-MEGAN, resampled to 100 m resolution; each panel includes an upper-right inset showing a magnified view of a representative urban sub-region. b, e, h, k Corresponding 500 m resolution isoprene fluxes estimated by the default MEGAN model, plotted over the same spatial extents and with identical inset areas for comparison. c, f, i, l Google Earth satellite imagery for the same regions (Maps data ©2019 Google). All panels in each city share a common geographic frame. The Paris Rive Gauche (PRG) site, marked in (j), is located on the seventh floor of the Lamarck B building at Université Paris Cité (30 m above ground level) and is equipped with Proton-Transfer-Reaction Mass Spectrometry (PTR-MS) for ambient isoprene monitoring.

Full size image

The high spatial heterogeneity captured by GEE-MEGAN, particularly in emission hotspots and high-value zones that are often fragmented or averaged in coarser-resolution models, consistently leads to higher emission estimates. For cities like London, Los Angeles, Paris, and Beijing, GEE-MEGAN estimated BVOC emissions for July 2019 to be 1.6-25 times higher than those predicted by the MEGAN default (Fig. 4). In Beijing, the larger discrepancy arises from fragmented vegetation being poorly resolved by lower-resolution land-cover inputs. GEE-MEGAN identifies an additional 2800.54 km² of vegetated area and a mean patch size of 0.67 km² within the same domain (Table S5), highlighting the critical role of spatial vegetation detail in driving urban BVOC emissions. This is primarily due to GEE-MEGAN’s incorporation of fine-scale vegetation components, such as isolated trees or smaller trees and shrubs40,41, which are often overlooked in coarser models (Table S6).

Fig. 4: Biogenic volatile organic compounds (BVOC) compositions in Paris, Los Angeles, London, and Beijing as modeled by MEGAN default and GEE-MEGAN.
figure 4

The base map is derived from Google Earth remote sensing satellite imagery (Maps data ©2025 Google). The bar depict the emissions of isoprene, monoterpenes, and sesquiterpenes for each city, focus on regions with nighttime light percentages above 80% (nighttime light data from the National Polar-orbiting Partnership, Visible Infrared Imaging Radiometer Suite). This threshold was used to define urban boundaries consistently across cities. For each city, the same spatial extent was used for both models to ensure direct comparison.

Full size image

A more specific comparison of BVOC emissions in New York on July 2, 2018 revealed that GEE-MEGAN estimated total emissions at 27 tons, nearly 30% higher than the 21 tons predicted by NYC-MEGAN12. This discrepancy arises from GEE-MEGAN’s superior ability to account for the continuous urban-to-suburban coverage of green spaces, addressing limitations in vegetation coverage used in NYC-MEGAN.

GEE-MEGAN provides a more accurate representation of the roles of BVOC emissions in urban atmospheric chemistry. In West Los Angeles, while the MEGAN default model estimated zero BVOC emissions for most of the 251 communities in August 2019 (Fig. 5a), using compound-specific OH reactivity (OHR) coefficients and SOA yield factors20, GEE-MEGAN identified BVOC emissions that were, on average, about 23.0 times higher (Fig. 5b), along with 23.4 times greater OHR and 25.1 times higher SOA formation potential (SOAFP) (Fig. S5). More recent state-of-the-art airborne flux measurements in Los Angeles during summer 202119,20 revealed that biogenic terpenoids accounted for ~15% of VOC emissions, but contributed over 50% of OHR and ozone/SOA formation potentials. In comparison, GEE-MEGAN estimated these BVOCs contributed to 3.5% of VOCs, 26.3% of OHR, and 29.4% of SOAFP, while the MEGAN default model severely underestimated these contributions (1% of VOCs, ~15% of OHR and SOAFP; Fig. S6). Interestingly, significant gaps still exist between observations and GEE-MEGAN estimates. This could be due to biased emission factors or it may reflect the substantial contributions of volatile chemical products (VCPs) to terpenoid emissions in Los Angeles. This suggests that more accurate BVOC emission estimates also enable better quantification of the contribution of temperature-sensitive VCPs42 to total terpenoid emissions, and help disentangle their temperature-driven variability20.

Fig. 5: Comparisons of biogenic volatile organic compounds (BVOC) emissions and community-level spatial distributions modeled by MEGAN default and GEE-MEGAN for West Los Angeles during August 2019.
figure 5

a Total BVOC emissions modeled by MEGAN default at a resolution of 500 m. b Total BVOC emissions modeled by GEE-MEGAN at a finer resolution of 30 m. Red boundary lines indicate community boundaries within West Los Angeles (https://lacounty.maps.arcgis.com/home/index.html). c Community-average daily BVOC emissions modeled by GEE-MEGAN. d Community-average daily BVOC emissions modeled by MEGAN default. e Differences in community-average daily emissions between GEE-MEGAN and MEGAN default, calculated as (c) minus (d).

Full size image

Implications for air quality and climate

GEE-MEGAN provides a solid foundation for investigating the complex interplay between BVOCs and climate by offering near-real-time high-resolution emission modeling. BVOC emissions from forests are the largest source of reactive organic gases released into the atmosphere43,44, profoundly influencing radiative ozone (a potent greenhouse gas in the lower atmosphere) formation, aerosol growth, cloud formation, and atmospheric oxidative capacity, which controls the lifetime of methane (CH4) and thereby radiative forcing4,45,46,47. GEE-MEGAN captures BVOC emissions more accurately and efficiently across heterogeneous landscapes, particularly at forest edges where human disturbance is more pronounced. While GEE-MEGAN and the default MEGAN model yield similar isoprene emissions in pristine forests, large discrepancies emerge at forest edges, such as along the Brazilian Amazon’s Arc of Deforestation35. For example, during the high-BVOC-emitting month of August 2019, emission discrepancies are more pronounced in areas affected by forest degradation or deforestation (Fig. 2c), highlighting the need for high-resolution emission estimates, especially along global forest edges. Although pathways by which BVOCs influence climate are more complex than currently recognized48, emission estimation with high spatiotemporal resolution as provided by GEE-MEGAN is a prerequisite for better delineating human perturbations to atmospheric compositions and climatic impacts of BVOCs at local, regional and even global scales.

GEE-MEGAN’s high-resolution, near-real-time emission modeling capabilities can help refine climate predictions by providing insights into the strong and complex feedbacks between climate and BVOC36. As temperature-driven BVOC emissions increase under global warming20,49, GEE-MEGAN can offer critical insights into how ecosystems respond to and feedback on climate change43,50,51,52 by providing more accurate and timely estimates of BVOC emissions driven by both natural and anthropogenic forces, such as deforestation and afforestation. Additionally, GEE-MEGAN supports the assessments of climate resilience strategies53,54,55, such as urban greening and afforestation/reforestation initiatives56,57,58, ensuring that these efforts balance the cooling benefits of increased vegetation with minimization of BVOC-related warming and pollution. This dual focus on environmental and climatic impacts optimizes sustainability measures, helping mitigate air quality degradation while maximizing carbon sequestration and heat mitigation benefits.

GEE-MEGAN also equips urban planners and even non-experts with powerful tools to assess the air quality impacts of urban greening initiatives. As BVOCs increasingly influence urban atmospheric chemistry due to declining anthropogenic emissions and expanding greening spaces20,42,59, accurately capturing their emissions in urban areas, however, becomes challenging because of complex vegetation patterns and dynamic land-use changes. By resolving emissions at finer spatial scale, down to the street and community level, GEE-MEGAN substantially outperforms the 500-m resolution MEGAN default model. As illustrated by the examples from Los Angeles discussed above, GEE-MEGAN provides a more precise understanding of the contributions of biogenic emissions to total urban terpenoid emissions, which is essential for evaluating their roles in air pollution, including ozone and SOA formation19,20,42. This ability enables more nuanced exposure assessments by accounting for interactions between biogenic and anthropogenic sources across diverse finer-scale microenvironments in densely populated areas58,59. Additionally, GEE-MEGAN’s near-real-time modeling capability supports adaptive air-quality management in response to events such as heat waves16,60,61,62 or rapid urbanization63, ensuring that green infrastructure projects maximize environmental benefits while minimizing unintended air pollution consequences.

Despite its advancements, GEE-MEGAN does not yet fully account for complex environmental factors influencing BVOC emissions64, such as soil moisture, nitrogen availability, evapotranspiration, and biotic stresses. Integrating these factors through GEE’s multisource, multiscale data capabilities and advanced machine learning techniques could further refine its predictions. Although ERA5-Land hourly meteorology (~11 km) was used in this study to ensure global applicability and consistency, our results indicate that improving the spatial representation of vegetation yields substantial gains in BVOC emission accuracy. Nevertheless, incorporating high-resolution meteorological reanalysis products or AI-driven weather prediction models65,66 could further enhance GEE-MEGAN’s capacity to simulate urban-scale processes, offering deeper insights into BVOC-related dynamics in air quality and climate change.

In summary, GEE-MEGAN bridges critical gaps in BVOC emission modeling by combining scalability, precision, and real-time capabilities. These features make it a transformative tool for improving our understanding of roles played by biogenic emissions in a rapidly urbanizing and warming world, addressing both air quality and climate challenges. By providing actionable insights for sustainable urban planning and climate mitigation, GEE-MEGAN sets a benchmark in biosphere-atmosphere interaction modeling.

Methods

GEE-MEGAN model

The GEE-MEGAN framework integrates the MEGAN2.17 biogenic emission model with the computational capabilities of GEE, providing a flexible and scalable platform for estimating BVOC emissions across diverse spatial and temporal scales and bridging the gap between high-resolution emission modeling and large-scale atmospheric analyses.

Data source module

This module aggregates inputs from a range of data sources to enhance model accuracy and adaptability: Fifth generation European Centre for Medium-Range Weather Forecasts atmospheric reanalysis (ERA5) climate data, MODIS-derived LAI, and reconstructed PFTs are used for large-scale analysis, and higher-resolution data such as Landsat imagery are employed for improving LAI precision, particularly in urban and regional studies (see Table S3 and Note S2). In addition to ERA5, GEE-MEGAN can incorporate other meteorological datasets such as Modern-Era Retrospective Analysis for Research and Applications version 2 (MERRA-2), the Climate Forecast System Version 2 (CFSv2), or the Global Land Data Assimilation System (GLDAS-2.1), and also support user-defined inputs such as WRF model output or CESM projections (Table S7).

Preprocessing module

To ensure consistency and reliability, this module processes the input through: (1) Temporal filtering: Aligning data to the required timeframes for simulations; (2) Spatial reprojection: Standardizing data projections to maintain spatial consistency; and (3) Resampling: Adjusting spatial resolution to match the scale of the analysis. Further details are provided in Note S3.

BVOC online calculation module

This module implements the core MEGAN2.1 computational framework within GEE, including canopy energy balance (Fig. S7) for simulating canopy conditions to estimate BVOC emissions, environmental correction factors for adjusting emissions based on environmental variables such as temperature and light, and standardized EFs7 for applying standardized emission rates for various vegetation types. Despite improving GEE-MEGAN computational efficiency and enabling near-real-time simulations, GEE-MEGAN produces emission activity factors and outputs consistent with MEGAN2.1 when using identical inputs (see Fig. S8 and Note S4).

Output and post-processing module

This module supports both online and local data processing, allowing data export for integration into atmospheric chemistry models. Outputs are available in formats compatible with subsequent analyses, with additional details outlined in Note S5.

Scalability and customization

GEE-MEGAN provides three operational scales tailored to different applications: local scale (10–30 m) that utilizes high-resolution remote sensing and machine learning (e.g., Random Forest algorithm) to refine land use and cover change classification (LUCC), enhancing accuracy in urban and regional estimates, regional scale (500 m) that balances resolution and computational efficiency for broader geographic analyses, and global scale (500–5000 m) that employs coarser datasets to support global-scale modeling of BVOC emissions (see Table S3).

Box modeling

To evaluate isoprene concentrations and validate GEE-MEGAN estimates, atmospheric simulations were performed using the BOXMOX model67, based on the Model for Ozone and Related Chemical Tracers (MOZART-4) chemical mechanism. The simulations focused on the Paris Rive Gauche site in southeastern Paris, where Proton-Transfer-Reaction Mass Spectrometry measurements of ambient isoprene concentrations were available for comparison with GEE-MEGAN outputs9. Key inputs for the BOXMOX simulations included meteorological data (2-m air temperature, boundary layer height, and photolysis rates derived from the ERA5 reanalysis data) and anthropogenic emissions (pollutant data such as nitrogen oxides, VOCs and other species obtained from Copernicus Atmosphere Monitoring Service).

Refinement of PFTs, emission factor maps (EFMAP) and LAI in GEE-MEGAN

PFTs (Table S8) in GEE-MEGAN were derived using a combination of the Köppen-Geiger climate classification map68 (Table S9) and the MODIS MCD12 land cover dataset69 (Fig. S9 and Table S10), following established methodologies used in the evolution of the Community Land Model (CLM)70. This integration leverages climate reanalysis data and MODIS land use categories to ensure accurate and consistent PFTs mapping. Detailed derivation approaches are provided in Table S11 and Note S6, with comparison of multiple PFTs classifications shown in Fig. S10.

For local modeling at 30 m and 10 m resolution, high-resolution datasets from Landsat and Sentinel satellites were used, with a supervised machine learning algorithm, specifically a random forest algorithm (ee.Classifier.smileRandomForest), to update near-real-time land cover mapping (Note S7 and S8). The random forest was configured with 50 decision trees, and the maximum depth and other parameters were set to default values. Training labels were generated using a combination of datasets, including MODIS MCD1269, GLC_FCS30D71, Global Sentinel-2 10 m Canopy Height72, EAS 10 m Annual Land Cover73, and Finer Resolution Observation and Monitoring of Global Land Cover 10 m (FROM-GLC10) datasets74. Each dataset was initially mapped to the Annual International Geosphere-Biosphere Programme classification69, and a hybrid sampling approach was employed to generate training labels. This approach combines spatially uniform sampling, which selected 5000 points randomly across the target area, and stratified sampling, which selected 1000 points from each LUCC classification category.

The predictor variables used in the random forest model were derived from Landsat and Sentinel satellite bands (visible, infrared, and near-infrared) as well as from calculated indices, including the normalized difference vegetation index, enhanced vegetation index, and soil adjusted vegetation index75. Historical data were filtered to create training labels and corresponding attributes, which were then used to infer land use classification based on near-real-time remote sensing data. Finally, the predicted LUCC data were divided into PFTs based on canopy height and climate rules, and these were subsequently used as inputs for GEE-MEGAN.

Dynamic EFMAP

The dynamic EFMAP module adjusts BVOC emission factors in real-time to account for environmental fluctuations. This dynamic refinement enhances the accuracy of BVOC emission estimates, as detailed in Note S7 in the supplementary materials.

LAI

The LAI estimation was improved by combining MODIS-derived LAI data with Landsat-based LAI interpolated on matching dates (Note S9). This hybrid approach reduces temporal gaps and enhances spatial resolution, achieving 9.1% reduction in RMSE at 500-m resolution and 3.6% reduction in RMSE at 3-km resolution. The refined LAI method was validated against the global DIRECT ground-based LAI dataset (Fig. S3A), ensuring reliability and accuracy for both local and global applications.

Multiscale simulation framework

To balance near-real-time applicability across spatial scales, GEE-MEGAN uses consistent meteorological inputs from ERA5-Land hourly while tailoring land-use and vegetation inputs across 100 m, 500 m, and 5 km scales, with varying data sources for LUCC and LAI to reflect spatial heterogeneity in emissions. For scales smaller than 500 m, high-resolution Landsat data were utilized to calculate LUCC and PFTs, while the LAI was derived using an enhanced MCD15 method with Landsat observations. At the 500 m scale, LUCC data were sourced from MCD12Q1 dataset, and LAI data were obtained using the MODIS MCD15A3H product. At the coarser 5 km scale, LUCC inputs were extracted from the MODIS MCD12C1dataset, and LAI was modeled using the LAI/FPAR Climate Data Record (CDR)76. At coarser spatial resolutions (e.g., 500 m), urban green space may be underrepresented or misclassified as impervious surfaces, resulting in negligible urban BVOC emissions. Accurate estimation of emissions in urban areas therefore requires high-resolution LUCC and LAI inputs, which GEE-MEGAN supports. The spatial resolution of GEE-MEGAN outputs aligns with the resolution of the input datasets and can be flexibly selected depending on the application. For urban or rapidly changing landscapes, we recommend using high resolution inputs (10–30 m); for regional or global assessments, coarser inputs (500 m to 5 km) are more practical and computationally efficient.

Validation

The accuracy of GEE-MEGAN was evaluated by comparing emission estimates with those of MEGAN2.1 across multiple sites and scales. Validation metrics included R², root mean squared error (RMSE), MAE, and Pearson correlation, for a comprehensive assessment of the model’s performance. These metrics were also applied to assess improvements in LAI estimates.

The GEE-MEGAN framework (Fig. S1) significantly improves the computational efficiency and scalability of BVOC emission modeling by leveraging the cloud-based capabilities of GEE21,22. This innovation enables near-real-time, high-resolution simulations across large spatial domains while minimizing both computational demands and data transfer requirements. For instance, using a consistent input dataset for South America (60° S–13° N, 35–90° W) at a 500-m resolution and operating on a local Intel Xeon Gold 6230R CPU (2.1 GHz), MEGAN2.1 required 736.9 s per species to complete the simulation while GEE-MEGAN achieved comparable results in just 7.2 to 9.1 s per species, delivering an 81- to 102-fold increase in processing speed; moreover, GEE-MEGAN reduced data transfer by 97–99% compared to traditional models, facilitating efficient processing even for large-scale or high-resolution applications.

GEE-MEGAN accurately replicates the core dynamics of MEGAN2.1 while significantly improving efficiency and scalability7. Simulations using input datasets from North America (https://bai.ess.uci.edu/megan/data-and-code/megan21, last retrieved September 14, 2024) revealed high consistency between MEGAN2.1 (Fig. S8a, e, i, m) and GEE-MEGAN (Fig. S8b, f, g, n), with a correlation coefficient of 1.0 and a total error of less than 0.1 across 16,576 data points. Minor discrepancies (Fig. S8d, h, l, p) were observed due to differences in computational environments but remained within acceptable margins (RMSE < 3.0 × 10−2; Fig. S8 and Table S12).