CarbonBench: A Global Benchmark for Upscaling of Carbon Fluxes Using Zero-Shot Learning

Aleksei Rozanov; Arvind Renganathan; Yimeng Zhang; Vipin Kumar

cs.LGMar 10, 2026

CarbonBench: A Global Benchmark for Upscaling of Carbon Fluxes Using Zero-Shot Learning

Authors: Aleksei Rozanov, Arvind Renganathan, Yimeng Zhang, Vipin Kumar

Organizations: Department of Computer Science & Engineering, University of Minnesota – Twin Cities, Minneapolis, MN, USA

Abstract

Accurately quantifying terrestrial carbon exchange is essential for climate policy and carbon accounting, yet models must generalize to ecosystems underrepresented in sparse eddy covariance observations. Despite this challenge being a natural instance of zero-shot spatial transfer learning for time series regression, no standardized benchmark exists to rigorously evaluate model performance across geographically distinct locations with different climate regimes and vegetation types. We introduce CarbonBench, the first benchmark for zero-shot spatial transfer in carbon flux upscaling. CarbonBench comprises over 1.3 million daily observations from 567 flux tower sites globally (2000-2024). It provides: (1) stratified evaluation protocols that explicitly test generalization across unseen vegetation types and climate regimes, separating spatial transfer from temporal autocorrelation; (2) a harmonized set of remote sensing and meteorological features to enable flexible architecture design; and (3) baselines ranging from tree-based methods to domain-generalization architectures. By bridging machine learning methodologies and Earth system science, CarbonBench aims to enable systematic comparison of transfer learning methods, serves as a testbed for regression under distribution shift, and contributes to the next-generation climate modeling efforts.

Figures & tables

Figure 1. Overview of the CarbonFluxBench setup. Eddy covariance flux observations provide sparse footprint-level targets, while remote sensing and meteorological data serve as gridded inputs for spatial generalization.

Figure 2. Summary of data sample quality in CarbonFluxBench . The upper panel shows the empirical distribution of $NEE\_VUT\_USTAR50\_QC$ across all samples. The lower left panel reports mean quality values aggregated by IGBP class, with bars indicating one standard deviation. The lower right panel shows the same aggregation for Köppen climate classes.

Figure 3. Temporal coverage of CarbonFluxBench , showing the distribution of samples by year (left) and by month (right).

Figure 4. The distribution of site number by different IGBP (left) and Köppen (right) types, illustrating the large imbalance between geographical context of the available ground-true observations.

Figure 5. The average flux aggregates by Köppen climate type with bars representing standard deviation (left axis) and the number of site per climate type (right axis).

Table 1. Train-test site partitions for IGBP vegetation (a) and Köppen climate (b) stratification schemes. All splits ensure representation of every category in both training and test sets despite severe class imbalance.

Table 7

Split	Target	$R^{2}\uparrow$	RMSE $\downarrow$	nMAE $\downarrow$
Köppen	GPP	0.124	2.958	0.806
	RECO	$-$ 0.174	2.303	0.960
	NEE	0.133	1.733	0.891
IGBP	GPP	0.114	2.920	0.797
	RECO	$-$ 0.228	2.212	0.985
	NEE	$-$ 0.005	1.750	0.972

Table 3. Performance of the process-based MiCASA model on CarbonFluxBench . Results show median site-level metrics across test sites. MiCASA predicts NPP (not GPP) and

R_{h}

(not RECO), which are related but distinct physical quantities.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Table 9

Table 10

Figure 6. Geographic distribution of 474 eddy covariance sites in CarbonFluxBench, colored by Köppen climate classification.

AmeriFlux
AR-CCg ( Posse, 2025 )	AR-TF1 ( Kutzbach, 2025 )	BR-CST ( Antonino, 2025 )	BR-Npw ( Vourlitis et al., 2026 )	CA-ARB ( Todd and Humphreys, 2025 )	CA-ARF ( Todd and Humphreys, 2026 )	CA-BOU ( Garneau, 2026 )	CA-Ca1 ( Black, 2025a )	CA-Ca2 ( Black, 2025b )
CA-Cbo ( Staebler, 2022 )	CA-CF1 ( Papakyriakou, 2026 )	CA-DB2 ( Knox, 2025 )	CA-DBB ( Christen and Knox, 2025 )	CA-EM1 ( Badiou, 2025 )	CA-ER1 ( Wagner-Riddle, 2025 )	CA-HPC ( Sonnentag and Marsh, 2026 )	CA-LP1 ( Black, 2026 )	CA-MA1 ( Amiro, 2025a )
CA-MA2 ( Amiro, 2025b )	CA-MA3 ( Amiro, 2025c )	CA-Qc2 ( Margolis, 2025 )	CA-SCB ( Sonnentag and Quinton, 2025 )	CA-SCC ( Sonnentag and Quinton, 2026 )	CA-TP1 ( Arain, 2026 )	CA-TP3 ( Arain, 2025a )	CA-TPD ( Arain, 2025b )	CL-SDF ( Perez-Quezada and Armesto, 2025 )
MX-Tes ( Yepez and Garatuza, 2025 )	US-A32 ( Billesbach et al., 2026a )	US-A74 ( Billesbach et al., 2026b )	US-Akn ( Leclerc, 2026 )	US-ALQ ( Desai et al., 2026k )	US-ARM ( Biraud et al., 2025 )	US-ASH ( Anderson, 2026a )	US-ASM ( Anderson, 2026b )	US-Bar ( Richardson and Hollinger, 2026 )
US-Bi1 ( Rey-Sanchez et al., 2025 )	US-Bi2 ( Rey-Sanchez et al., 2026 )	US-BRG ( Novick, 2026b )	US-BZB ( Euskirchen, 2026a )	US-BZF ( Euskirchen, 2025a )	US-BZo ( Euskirchen, 2026b )	US-BZS ( Euskirchen, 2025b )	US-CdM ( Bowling et al., 2025 )	US-CF1 ( Phillips et al., 2026 )
US-CF2 ( Huggins, 2026a )	US-CF3 ( Huggins, 2026b )	US-CF4 ( Huggins, 2026c )	US-CGG ( Oikawa and Fenster, 2026 )	US-CS1 ( Desai, 2026a )	US-CS2 ( Desai, 2026b )	US-CS3 ( Desai, 2026c )	US-CS4 ( Desai, 2026d )	US-CS5 ( Desai, 2026e )

Appendix

Table 6. AmeriFlux sites used in this study and their dataset citations. We cite individual AmeriFlux sites as it is required by the data policy.

Cross-Domain Generalization

Cross-domain generalization studies how machine-learning models retain reliable performance when data, sensors, populations, or tasks differ from training.

See all 178 papers

FLUXtrapolation: A benchmark on extrapolating ecosystem fluxes
May 19, 2026Anya Fries, Jacob A Nelson, Martin Jung +2Distribution Shift Robustness OOD Generalization
Seminar for Statistics, ETH Zürich Zürich, Switzerland · Max Planck Institute for Biogeochemistry Jena, Germany
Task Aware Modulation Using Representation Learning for Upsaling of Terrestrial Carbon Fluxes
Mar 10, 2026Aleksei Rozanov, Arvind Renganathan, Vipin KumarClimate Science Physics-Informed ML
University of Minnesota, Twin Cities
A Footprint-Aware, High-Resolution Approach for Carbon Flux Prediction Across Diverse Ecosystems
Dec 1, 2025Jacob Searcy, Anish Dulal, Courtney Mathers +6Remote Sensing
Department of Data Science, University of Oregon · Department of Computer Science, University of Oregon · Institute of Ecology and Evolution, University of Oregon +3

May 19, 2026J/K move · Enter open · S save

FLUXtrapolation: A benchmark on extrapolating ecosystem fluxes

Anya Fries, Jacob A Nelson, Martin Jung, Markus Reichstein+1

Seminar for Statistics, ETH Zürich Zürich, Switzerland · Max Planck Institute for Biogeochemistry Jena, Germany

Distribution Shift Robustness OOD Generalization Forecasting Benchmarks Spatiotemporal Forecasting

Open paper AskPDF arXiv

We introduce FLUXtrapolation, a benchmark for extrapolating ecosystem fluxes under progressively harder distribution shifts. Ecosystem fluxes are central to understanding the carbon, water, and energy cycles, yet they can only be measured directly at sparsely located measurement towers. Producing global flux estimates therefore requires training models on observed sites using globally available covariates and predicting in unobserved regions, that is, upscaling. Flux upscaling is a challenging domain generalization problem that is affected by a shift in covariate distribution across climates, ecosystem types, and environmental conditions, as well as by conditional shift: important drivers remain unobserved at global scale. We provide a quantitative analysis of both these shifts in

P_X

and

P_{Y\mid X}

. FLUXtrapolation is designed based on domain expertise on flux upscaling: it defines temporal, spatial, and temperature-based extrapolation scenarios and evaluates performance across held-out domains, temporal aggregations, and tail errors. In a pilot study, we find that baselines perform similarly under median hourly RMSE, but separate under the proposed tail-focused and multi-scale evaluation. FLUXtrapolation therefore poses a realistic and thus relevant challenge for machine learning methods under distribution shift; at the same time, progress on this benchmark would directly support the scientific goal of improving flux upscaling.

Abstract

Cross-Domain Generalization

Explore similar work

FLUXtrapolation: A benchmark on extrapolating ecosystem fluxes

Task Aware Modulation Using Representation Learning for Upsaling of Terrestrial Carbon Fluxes

A Footprint-Aware, High-Resolution Approach for Carbon Flux Prediction Across Diverse Ecosystems