🌍 CarbonGlobe Dataset

Accepted to NeurIPS 2025 (Datasets & Benchmarks Track)

• Paper is available on OpenReview: https://openreview.net/pdf?id=M07aAJKH8B
• Project page is available on NeurIPS: https://neurips.cc/virtual/2025/loc/san-diego/poster/121701.
• Datasets are available on Kaggle: https://www.kaggle.com/datasets/zhihaow/carbonglobe.

Overview

CarbonGlobe is a comprehensive, ML-ready dataset for modeling and forecasting the forest carbon cycle. It integrates heterogeneous Earth system variables into a consistent spatiotemporal grid and provides standardized scenario-based evaluations and benchmark implementations to accelerate research across climate science, environmental monitoring, and ecological modeling.

Key Features

• Global coverage at 0.5°
  First global-scale ML-ready dataset for monitoring and forecasting forest carbon dynamics.
• Multi-decadal span (40 years)
  Enables long-term trend and variability analysis.
• 100+ variables
  Harmonized inputs from meteorology, CO₂, soils, vegetation, and ancillary layers.
• Scenario-based splits
  Training/testing protocols resembling real applications (e.g., climate zones, forest age).
• Benchmarks & metrics
  Strong baselines (LSTM, Transformer family, DeepED, etc.) and problem-driven metrics for carbon forecasting.

Source Datasets

All inputs are from open sources. Please follow original licenses and citation guidelines.

• Meteorological (MERRA-2)
  NASA GMAO MERRA-2 (Gelaro et al., 2017).
  Source: https://gmao.gsfc.nasa.gov/reanalysis/MERRA-2/
  Citation guidance: https://gmao.gsfc.nasa.gov/reanalysis/MERRA-2/citing_MERRA-2/

• CO₂ (NOAA CarbonTracker)
  Peters et al., 2007.
  Source: https://gml.noaa.gov/ccgg/carbontracker/

• Soils (ROSETTA)
  Montzka et al., 2017. CC BY 3.0.
  DOI: https://doi.pangaea.de/10.1594/PANGAEA.870605

• Climate Zones (Köppen–Geiger)
  Beck et al., 2018. CC BY 4.0.
  Data: https://figshare.com/articles/dataset/6396959/2

Benchmarks

We provide reproducible baselines covering classical sequence models, knowledge-guided emulators, and recent transformer variants:

• LSTM — Standard recurrent model for time-series forecasting [Graves, 2012]
• LSTNet — CNN+RNN for short/long dependencies [Lai et al., 2018]
• DeepED — Knowledge-guided ED emulator (error-accumulation mitigation) [Wang et al., 2023]
• Transformer — Vanilla self-attention [Vaswani et al., 2017]
• Informer — ProbSparse attention for efficiency [Zhou et al., 2021]
• DLinear — Seasonal/trend decomposition with linear heads [Zeng et al., 2023]
• Crossformer — Cross-time/feature attention [Zhang et al., 2023]
• TimeXer — Decoupled inter-target & input-target modeling [Wang et al., 2024]

Problem-driven metrics (e.g., RMSE, MAE, delta error, cumulative error) are included to capture both step-wise and long-horizon behavior.

📚 Citation

If you use CarbonGlobe in your research, please cite:

Zhihao Wang, Lei Ma, George Hurtt, Xiaowei Jia, Yanhua Li, Ruohan Li, Zhili Li, Shuo Xu, Yiqun Xie.
CarbonGlobe: A Global-Scale, Multi-Decade Dataset and Benchmark for Carbon Forecasting in Forest Ecosystems.
In Proceedings of the 39th Conference on Neural Information Processing Systems (NeurIPS), 2025.

📬 Contact

For questions or feedback, feel free to reach out:

• Zhihao Wang — [email protected]
• Yiqun Xie — [email protected]