Structured Physical Attributes Enable Efficient Foundation Models for Land-Surface Prediction

Foundation models for structured tabular and time-series data have demonstrated strong zero-shot and transfer capabilities across general-purpose prediction tasks, yet their applicability to domain-governed physical systems remains underexplored. We present StefaLand, a structured foundation model pretrained on tabular landscape attributes and meteorological time-series across 8,634 global catchments, targeting two core Earth system prediction tasks: streamflow forecasting and soil moisture estimation. StefaLand employs a transformer-based masked autoencoder with cross-variable group masking to learn interactions among physically related attribute groups, and is adapted to downstream tasks through lightweight residual adapters. Against strong baselines including TabPFN, TiRex, AlphaEarth embeddings, and supervised LSTM models, StefaLand achieves consistent gains in spatial generalization under strict holdout regimes, reducing streamflow RMSE by approximately 20\% over supervised baselines in ungauged basin experiments. Our results demonstrate that domain-specialized pretraining on structured physical attributes provides stronger cross-domain transfer than general-purpose tabular foundation models, and offers a computationally accessible alternative to large-scale vision-based Earth observation models.