Scalability of causal deep learning methods to complex Earth observation data

Determine the scalability of causal deep learning approaches—specifically causal representation learning and meta-learning of structural causal models—to complex, noisy, high-dimensional Earth observation datasets, assessing whether these methods can be effectively applied beyond synthetic or controlled settings.

Background

Within the discussion of causality in machine learning, the paper highlights that recent advances such as causal representation learning and meta-learning of structural causal models have primarily been validated on synthetic or highly controlled datasets. This raises concerns about whether these methods scale to the complexity and noise inherent in real-world Earth observation data, which often involves high-dimensional, multi-modal time series with spatial and temporal heterogeneity.

The authors frame this as a critical gap for environmental applications, noting that robust causal methods capable of handling such data are necessary to achieve generalization under distribution shift. Their own approach leverages J-PCMCI+ for causal discovery in supraglacial lake time series, but the broader question of scaling causal deep learning techniques to complex Earth observation remains explicitly open.