Probabilistic machine learning for noisy labels in Earth observation.

Area of Science:

• Earth Observation (EO)
• Machine Learning (ML)
• Geospatial Data Analysis
• Data Science

Background:

• Label noise significantly degrades supervised ML model performance in Earth Observation (EO).
• Reliable ML solutions are crucial for high-impact EO applications.
• Existing methods often fail to account for domain-specific noise sources in EO data.

Purpose of the Study:

• To leverage probabilistic ML for modeling input-dependent label noise in EO.
• To quantify data uncertainty in EO tasks, addressing unique noise characteristics.
• To develop and assess uncertainty-aware ML models for improved EO applications.

Main Methods:

• Training uncertainty-aware probabilistic models across diverse EO applications.
• Utilizing a dedicated pipeline to evaluate model accuracy and reliability.
• Assessing performance against standard deterministic ML approaches.

Main Results:

• Uncertainty-aware models demonstrated superior performance over deterministic models in most EO datasets and metrics.
• Rigorous evaluation validated the reliability of predicted uncertainty estimates.
• Enhanced interpretability of model predictions was achieved through uncertainty quantification.

Conclusions:

• Modeling label noise and incorporating uncertainty quantification are vital for robust EO solutions.
• Probabilistic ML offers a promising direction for trustworthy AI in Earth Observation.
• The study paves the way for more reliable and interpretable ML applications in the EO domain.