Evidential Deep-Learning Emulator for Land Surface Models with Uncertainty Quantification and Seasonal Sensitivity Analysis

This repository provides code and workflows to train, evaluate, and analyze an evidential deep neural network (EDNN) emulator for CLM5 (Community Land Model v5) leaf area index (LAI) using a perturbed-parameter ensemble (PPE). The EDNN maps 32 CLM5 parameters to seasonal LAI anomalies and outputs a Normal–Inverse–Gamma predictive distribution, providing the mean and decomposed aleatoric and epistemic uncertainties. The workflow supports seasonal evaluation and FAST (Fourier Amplitude Sensitivity Test) analyses to quantify main and interaction effects across seasons.

What's included

• Data preprocessing for CLM5 PPE outputs and feature engineering (e.g., cyclical time encodings)
• Reproducible member/time splits, training scripts, and evaluation (R², RMSE, calibration via reliability/PIT)
• Uncertainty quantification and seasonal sensitivity analysis (FAST) with plotting utilities
• Notebooks and scripts to reproduce figures and key results
• Environment and configuration files for end-to-end runs

Intended use

• Rapid surrogate modeling for parameter exploration and calibration
• Quantifying predictive and uncertainty in land-surface modeling
• Assessing seasonal parameter sensitivities that control LAI variability

Data & licensing

• PPE data and required inputs are referenced in the repository documentation (see README.md / Data Availability).
• Licensing terms are provided in LICENSE.

How to cite

Please cite the Zenodo record for this software (concept DOI) and, for exact reproducibility, the version-specific DOI corresponding to the release you used. If a CITATION.cff is present, follow its guidance for author names and affiliations.