Uncertainty quantification for probabilistic machine learning in earth observation using conformal prediction

Area of Science:

• Earth Observation (EO)
• Machine Learning
• Uncertainty Quantification

Background:

• Machine learning on EO data is crucial for international accords but suffers from unreliable uncertainty quantification.
• Existing methods often fail to provide statistically valid uncertainty estimates for EO datasets.
• Only 22.5% of reviewed EO datasets included uncertainty information, with prevalent unreliable techniques.

Purpose of the Study:

• Introduce conformal prediction as a statistically sound method for uncertainty quantification in EO.
• Address the need for reliable uncertainty estimation in EO data processing.
• Facilitate the integration of uncertainty quantification into existing EO machine learning workflows.

Main Methods:

• Developed Google Earth Engine native modules for conformal prediction, bringing computation to the data.
• Applied conformal prediction to diverse EO applications, including regression and classification tasks.
• Reviewed existing EO datasets to assess current uncertainty quantification practices.

Main Results:

• Conformal prediction provides statistically valid prediction regions applicable to any machine learning model and data distribution.
• The developed Google Earth Engine modules enable efficient, data-native uncertainty quantification.
• Demonstrated the versatility and scalability of conformal prediction across various EO applications and scales.

Conclusions:

• Accessible tools for conformal prediction will drive wider adoption of rigorous uncertainty quantification in EO.
• Enhanced uncertainty quantification improves the reliability of downstream EO applications, including monitoring and decision-making.
• The developed modules simplify the integration of uncertainty quantification into traditional and deep learning models for EO.