Data for UFLUX-ensemble Journal Publication

UFLUX-Ensemble (v1) Journal Publication Data and Workspace This dataset provides the complete workspace file system used to generate the figures and analyses presented in the manuscript. It also includes both the input and output files used for training the UFLUX machine learning algorithm and for producing the UFLUX-ensemble data products. Source code: https://github.com/songyanzhu/UFLUX-ensemble_scidata Trained models: Due to their large size, the trained models are hosted separately: MODIS-related models: https://zenodo.org/uploads/17256064 AVHRR-related models: https://zenodo.org/uploads/17256071 SIF-related models: https://zenodo.org/uploads/17256073 Note: To use the models, download and unzip them into the 1_models folder. For demonstration purposes, we only provide grid data from satellites and climate records for a single timepoint due to file size limitations. These datasets are openly available from sources such as Copernicus and NASA; the manuscript includes direct links, which will be added here once the paper is published. The 3_output_products folder is intentionally left empty, as it is the designated directory for storing the UFLUX-ensemble outputs. With the provided models and grid data, you can fully reproduce the results. The models were trained using the deepforest-21 algorithm (https://github.com/LAMDA-NJU/Deep-Forest). Please ensure that the corresponding Python package is installed before running the workflows. We also welcome requests to train and provide additional datasets using alternative configurations (e.g., different satellite products, climate reanalyses, or spatiotemporal resolutions). These can be shared free of charge for research purposes and/or for the broader public interest worldwide. For queries or requests, please contact: [e-mail] OverviewThe UFLUX ensemble (v1) dataset offers multiscale fluxes , generated using machine learning models. It integrates satellite-based vegetation proxies (e.g., vegetation index and solar-induced fluorescence) with climate reanalysis like ERA5, and is trained against eddy covariance observations (e.g., FLUXNET, ICOS, and Ameriflux). It includes five core flux components: Gross Primary Production (GPP) Ecosystem Respiration (RECO) Net Ecosystem Exchange (NEE) Sensible Heat Flux (H) Latent Energy Flux (LE) Background and MethodologyThe Unified FLUXes (UFLUX) initiative is a data-driven, machine learning-based platform designed to upscale eddy covariance (EC) flux measurements from tower sites to the global scale. It aims to answer pressing questions about how effectively terrestrial ecosystems are managed under climate change. Key innovations of UFLUX include: Consistent Upscaling Framework: Harmonizes flux upscaling across spatial/temporal scales and multiple flux types (GPP, RECO, etc.) using deep decision tree-based methods, better suited than conventional neural networks for EC flux data. Hybrid Explainable ML: Combines black-box ML with ecological interpretability through residual learning, offering both predictive power and new scientific insight (UFLUXv2). Uncertainty Quantification: Employs sampling space completeness to assess model uncertainty in a transparent, robust manner. Multisource Integration: Leverages complementary strengths of vegetation proxies (e.g., NIRv, SIF) and climate data (e.g., ERA5) to represent carbon dynamics more comprehensively than single-source approaches. Superior Gap-Filling: Originally developed as a global EC flux gap-filling tool, UFLUX improves accuracy by up to 30% and reduces uncertainty by as much as 70% compared to traditional methods. High Performance: Achieves strong predictive accuracy, with global-scale R² > 0.8 for RECO and ≈0.9 for GPP, while being computationally efficient enough to run on a standard laptop. Community Adoption: Already used by other global upscaling projects, highlighting its reliability and impact. ApplicationsUFLUX is ideal for studying the interactions between land management, climate change, and carbon fluxes, particularly in improving global estimates of GPP and RECO by addressing biases in EC measurements. Resources UFLUX Website: https://sites.google.com/view/uflux Code Repository: https://github.com/soonyenju/uflux Code Repository for Journal Publication: https://github.com/songyanzhu/UFLUX-ensemble_scidata Technical & Descriptive Publication: https://doi.org/10.1080/01431161.2024.2312266