ADECEES: Anomaly DEtection of CO2 Emissions via Ensemble Segmentation

Accurate and spatially explicit monitoring of anthropogenic CO2 emissions is essential for independent verification of mitigation efforts and understanding their atmospheric impact.

ADECEES is a generative anomaly detection method designed for the unsupervised identification of CO2 point sources from satellite-derived XCO2 (column-averaged dry-air mole fractions of CO2) maps. By learning the distribution of healthy, background atmospheric variability, ADECEES provides the framework to globally isolate statistically significant concentration enhancements.

Key Features

• Unsupervised Detection: Identifies localized emission sources directly from atmospheric concentration fields, overcoming challenges related to background variability and transport-driven spatial structures.
• Generative Framework: Utilizes a partial diffusion model trained exclusively on background atmospheric states.
• High-Resolution Tracking: Evaluated using a global daily XCO2 dataset, derived from OCO-2 observations.
• Temporal Monitoring: Capable of discriminating between operational and inactive periods for specific sites (e.g., coal mines) and detecting spatiotemporal changes linked to plant closures, units upgrades, or commissioning events.

Methodology

The ADECEES framework operates through three core components:

1. Partial Diffusion Model
2. Ensemble-Based Reconstruction: Generates multiple reconstructed background states for a given observation to account for uncertainty and establish a robust baseline.
3. Residual Segmentation: Compares the actual satellite-derived map against the reconstructed ensemble background to isolate and segment statistically significant anomalies (emissions).

Data Requirements

This framework is built to process satellite-derived XCO2 maps. The default implementation is configured for OCO-2 observational data processed to a 0.03° × 0.04° gridded resolution. The dataset can be found here.
Link to the dataset paper: paper

Installation

Clone the repository and install the required dependencies:

git clone https://github.com/ar1619/ADECEES.git
cd ADECEES
pip install -r requirements.txt

Quick Start

Training using available configuration and sample data

python diffusion_training.py 1

Run anomaly detection between start_date and end_date (modify accordingly)

python compute_variance.py 1 start_date end_date positive

Citation

@article{Rakotoharisoa2026ADECEES,
  title={Diffusion-Based Anomaly Detection for Satellite Monitoring of CO2 Point Sources},
  author={Rakotoharisoa, Andrianirina and Cenci, Simone and Arcucci, Rossella},
  journal={Available at SSRN 6368100},
  year={2026}
}

License

This project is licensed under the MIT License - see the LICENSE file for details.