Note: This repository is still being completed.

DERE Dataset

Processed datasets for the accepted paper at the KDD AI4Science Track:

Knowledge-Guided Learning for Global Carbon Flux Prediction: Integrating High-Level Remote Sensing with Bottom-Up Physical Modeling

This dataset repository provides the processed .npz files used by the DERE codebase for global carbon flux prediction. The data support experiments for the proposed DERE framework, baseline models, and KGML comparison models.

DERE integrates process-based model simulations, high-level remote sensing observations, and in-situ flux measurements to predict carbon flux variables, including GPP, RECO, and NEE.

🧩 Overview

The dataset contains processed inputs, labels, normalization statistics, plant functional type information, age-weight labels, in-situ observations, and imputed in-situ labels used in the DERE pipeline.

The data are prepared for direct use with the corresponding scripts in the DERE code repository. Each experiment typically uses three types of files:

• data: model input and target data
• stat: normalization statistics
• pft: PFT labels, in-situ labels, age-weight information, or imputed labels

📁 Data Files for Each Script

DERE-main/Step01_DERE_Train_3PureModels_CompetitionModel.py
data: res_train4_test8_extract_4types_28years_update_with_NEE_Ra_RECO.npz
stat: data_stats_with_NEE_Ra_RECO.npz
pft:  pft_dataset_12mean_4types_28years_update.npz

DERE-main/Step02_DERE_Finetune_CompetitionModel.py
data: res_train4_test8_extract_28years_ageindependent_update_with_NEE_Ra_RECO.npz
stat: data_stats_with_NEE_Ra_RECO.npz
pft:  pft_dataset_12mean_28years_ageindependent_plus_ageweight_update.npz

DERE-main/Step03_DERE_Train_PFTModel.py
data: 1_res_train4_test8_allx_plus_ageweight_esapft_update.npz
stat: data_stats.npz
pft:  1_ED_PFT_train4_test8_1992_to_2020_update.npz

DERE-main/Step04_DERE_Finetune_with_InSitu.py
data: res_train4_test8_extract_4types_28years_{net}.npz
stat: data_stats_with_NEE_Ra_RECO.npz
pft:  pft_dataset_12mean_4types_28years_ESACCI_plusAW_{net}.npz

DERE-main/Step05_DERE_InSitu_imputation_CSDI-main/05_DERE_InSitu_imputation.py
data: res_train4_test8_extract_4types_28years_insituless_update_6networks_with_NEE_Ra_RECO.npz
stat: data_stats_with_NEE_Ra_RECO.npz
pft:  pft_dataset_12mean_4types_28years_ESACCI_plusAW_insituless_update_6networks.npz

DERE-main/Step06_DERE_Finetune_with_InSitu_imputation.py
data: res_train4_test8_extract_4types_28years_{net}.npz
stat: data_stats_with_NEE_Ra_RECO.npz
pft:  pft_dataset_12mean_4types_28years_ESACCI_plusAW_{net}_imputation.npz

For all Baseline scripts
data: res_train4_test8_extract_4types_28years_{net}.npz
stat: data_stats_with_NEE_Ra_RECO.npz
pft:  pft_dataset_12mean_4types_28years_ESACCI_plusAW_{net}.npz

For all KGML scripts
Pretraining:
data: res_train4_test8_extract_28years_ageindependent_update_with_NEE_Ra_RECO.npz
stat: data_stats_with_NEE_Ra_RECO.npz
pft:  pft_dataset_12mean_28years_ageindependent_plus_ageweight_update.npz
Finetuning:
data: res_train4_test8_extract_4types_28years_{net}.npz
stat: data_stats_with_NEE_Ra_RECO.npz
pft:  pft_dataset_12mean_4types_28years_ESACCI_plusAW_{net}.npz

🌐 Available Networks

The {net} field in the filenames refers to one of the following networks:

above
ameriflux
fluxnet
icos-ww
mix

📊 Data Format

All files are stored in NumPy .npz format. They can be loaded with:

import numpy as np

data = np.load("file_name.npz")
print(data.files)

The exact arrays contained in each file depend on the corresponding experiment script. Please refer to the DERE code repository for the expected keys and shapes.

🔗 Code Repository

The dataset is designed to be used with the DERE code repository:

https://github.com/ai-spatial/DERE

Before running the experiments, update the data paths in the corresponding scripts according to your local dataset location.

📚 Citation

If you use this dataset, please cite:

@inproceedings{xu2026knowledge,
  author    = {Shuo Xu and Zhihao Wang and Ruohan Li and Ruichen Wang and Lei Ma and George C. Hurtt and Xiaowei Jia and Yiqun Xie},
  title     = {Knowledge-Guided Learning for Global Carbon Flux Prediction: Integrating High-Level Remote Sensing with Bottom-Up Physical Modeling},
  booktitle = {Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2},
  year      = {2026},
  address   = {Jeju Island, Republic of Korea},
  publisher = {ACM},
  doi       = {10.1145/3770855.3818927}
}

📬 Contact

For questions or feedback, feel free to reach out:

• Shuo Xu — shuoxu98@umd.edu
• Yiqun Xie — xie@umd.edu