Impact assessment of missing data in model predictions for Earth observation applications
Published in IEEE International Geoscience and Remote Sensing Symposium (IGARSS), 2024
DOI: 10.1109/IGARSS53475.2024.10640375
Code: github.com/fmenat/missingviews-study-EO
This work investigates the impact of missing temporal and static data sources on machine learning predictions in Earth Observation. Across four datasets covering classification and regression tasks, the study evaluates how different multi-view modeling strategies respond when individual data sources become unavailable due to factors such as clouds, noise, or satellite failures. Results show that some approaches are inherently more robust to missing data, with Ensemble fusion achieving up to 100% prediction robustness. Missing-data scenarios are found to be substantially more challenging for regression than classification, while optical data emerges as the most critical individual view, with its absence causing the largest degradation in predictive performance.
Recommended citation: Mena, Francisco, et al. "Impact assessment of missing data in model predictions for Earth observation applications." IGARSS 2024-2024 IEEE International Geoscience and Remote Sensing Symposium. IEEE, 2024.
Download Paper | Download Slides