Husnain Asif: Shared Latent Representations for Kilometer-Scale Multivariate Climate Fields
Learning representations of climate data remains challenging at km-scale resolution, where near-surface fields exhibit sharp gradients, localized extremes, land-ocean contrasts, and coupled variability across physical variables. Existing methods are largely designed for coarse global fields, single-variable settings, or task-specific prediction pipelines, leaving it unclear whether a single latent state can represent heterogeneous regional climate fields with scientific fidelity. We introduce CSVAE, a variational autoencoder-based architecture for learning shared latent representations of km-scale climate fields. The encoder uses static geographic fields and cyclical temporal metadata via multi-scale Feature-wise Linear Modulation, while the decoder receives only the latent tensor, with no auxiliary inputs or skip connections. This asymmetric design forces topographic, seasonal, spatial, and cross-variable structures to pass through the shared bottleneck. To handle heterogeneous near-surface variables, the model combines variable-aware monotone preprocessing with a reconstruction objective that targets local amplitude, spatial organization, and distributional range. On 0.04° regional reanalysis fields over Australia, CSVAE preserves fine-scale spatial structure, marginal distributions, tail variability, spectral properties, and multivariate dependence across six near-surface variables. The learned state enables direct latent-space analysis without full reconstruction. These findings show that km-scale regional climate fields can be encoded as a compact shared latent state, extending learned climate representations beyond storage-oriented compression and task-specific prediction models.