Research map / Large language models / AI and society
CarbonSense: A Multimodal Dataset and Baseline for Carbon Flux Modelling
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces CarbonSense, a standardized multimodal ML-ready dataset of eddy covariance carbon fluxes, meteorological predictors and MODIS imagery from 385 sites, plus the EcoPerceiver Perceiver-style baseline that beats an XGBoost stand-in for SOTA.
Problem
Data-driven carbon flux modelling (DDCFM) lacks a standardized dataset for comparing models, and current SOTA relies on tabular methods that do not exploit the multimodal nature of biophysical data.
Approach
A Perceiver-based multimodal architecture (EcoPerceiver) with windowed cross attention over time and causal self-attention, using Fourier encoding of tabular inputs and modality masking for missing data, released alongside a fused, ML-ready eddy covariance + MODIS dataset.
Key results
- EcoPerceiver beats XGBoost on 12 of 15 IGBP types in NSE, e.g. SAV NSE 0.7368 vs 0.5802 and DNF 0.4336 vs 0.2803.
- XGBoost wins slightly in ENF (0.7765 vs 0.7694), WET (0.4530 vs 0.4137) and WAT, where both models score worse than predicting the mean.
- Ablation: vanilla transformer (NSE mean 0.557) is clearly worse than the full EcoPerceiver (0.669) on truncated mean NSE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EcoPerceiver instead of XGBoost | Carbon flux prediction across IGBP ecosystem types · CRO, CSH, CVM, DBF, DNF, EBF, ENF, GRA, MF, OSH, SAV, SNO, WAT, WET, WSA | NSE: 12 out of 15 IGBP types; XGBoost performs better in WET, WAT, and ENF | Mixed |
| add causal masking (ablation) | EcoPerceiver · CarbonSense test sites, excluding WAT | NSE truncated mean: 0.654 -> 0.669 | Helped |
| add observational dropout (ablation) | EcoPerceiver · CarbonSense test sites, excluding WAT | NSE truncated mean: 0.634 -> 0.669 | Helped |
| add Fourier encoding (ablation) | EcoPerceiver · CarbonSense test sites, excluding WAT | NSE truncated mean: 0.624 -> 0.669 | Helped |
| handle missing inputs within the model rather than using gapfilled values (ablation) | EcoPerceiver · CarbonSense test sites, excluding WAT | NSE truncated mean: 0.635 -> 0.669 | Helped |
| add image data (ablation) | EcoPerceiver · CarbonSense test sites, excluding WAT | NSE truncated mean: 0.635 -> 0.669 | Helped |
| use WCA and CSA rather than a vanilla transformer encoder (ablation) | EcoPerceiver · CarbonSense test sites, excluding WAT | NSE truncated mean: 0.557 -> 0.669 | Helped |
Surprising findings
- Both XGBoost and EcoPerceiver perform worse than predicting the mean on water bodies (WAT), with NSE values of -11 and -14.4.
- A vanilla transformer beats EcoPerceiver on WAT (NSE -9.68 vs -14.40).
Limitations
- Strong geographic and ecosystem imbalance; Africa, Central Asia and South America underrepresented.
- EcoPerceiver underperforms XGBoost in boreal ecosystems (ENF, WET) and WAT; hyperparameter search for deep model was limited by compute.
- Ablations run with a single seed due to compute limits.
Experimental setup
Datasets: CarbonSense FLUXNET 2015 ICOS 2023 release ICOS Warm Winter release AmeriFlux 2023 release MODIS MCD43A4 MODIS MCD43A2 ERA5
Benchmarks: CarbonSense
Metrics: NSE RMSE MSE
Baselines: XGBoost Random Forest Linear Model Vanilla Transformer
Builds on: Perceiver NeRF Fourier encoding Swin Transformer XGBoost
Closest papers by meaning
- Multi-modal Learning for Geospatial Vegetation Forecasting (CVPR 2024 · similarity 0.83)
- Terra: A Multimodal Spatio-Temporal Dataset Spanning the Earth (NeurIPS 2024 · similarity 0.83)
- REACT: Physically and Chemically Consistent Reconstruction of Marine Active Tracers (NeurIPS 2026 · similarity 0.83)
- LLMCarbon: Modeling the End-to-End Carbon Footprint of Large Language Models (ICLR 2024 · similarity 0.82)
- CATransformers: Carbon Aware Transformers Through Joint Model-Hardware Optimization (NeurIPS 2025 · similarity 0.82)
- OxyGenerator: Reconstructing Global Ocean Deoxygenation Over a Century with Deep Learning (ICML 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.