Impact Assessment of Missing Data in Model Predictions for Earth Observation Applications

Earth observation (EO) applications involving complex and heterogeneous datasources are commonly approached with machine learning models. However, there isa common assumption that data sources will be persistently available. Differentsituations could affect the availability of EO sources, like noise, clouds, orsatellite mission failures. In this work, we assess the impact of missingtemporal and static EO sources in trained models across four datasets withclassification and regression tasks. We compare the predictive quality ofdifferent methods and find that some are naturally more robust to missing data.The Ensemble strategy, in particular, achieves a prediction robustness up to100%. We evidence that missing scenarios are significantly more challenging inregression than classification tasks. Finally, we find that the optical view isthe most critical view when it is missing individually.