Leveraging Weather Foundation Models for Hydrological Applications: Enhancing Hydrological Prediction through Sophisticated Decoder Design

In recent years, data-driven models have demonstrated remarkable performance in capturing complex land-atmosphere interactions. In particular, the emergence of weather foundation models, which are pre-trained on vast unlabeled meteorological datasets through self-supervision and can be applied to diverse downstream tasks, has introduced robust backbones capable of representing global atmospheric dynamics. However, fine-tuning these massive models to specific downstream hydrological tasks presents significant challenges. Full fine-tuning is computationally prohibitive, and even parameter-efficient fine-tuning methods, such as Low-Rank Adaptation, also have an amount of computational overhead over the large embedding dimensions of foundation models. Furthermore, modifying the backbone's weights can be a risk of catastrophic forgetting or destabilize the learned representations, which are essential for maintaining their physical consistency during iterative long-term forecasts.

To address these challenges, this study investigates a transfer learning approach that utilizes a weather foundation model backbone with lightweight decoders. This strategy allows the model to handle the robust feature space of the pre-trained backbone while maintaining computational efficiency and architectural stability. We design and evaluate two representative classes of lightweight decoder architectures that differ in their structural complexity and information integration strategy. The first decoder adopts a minimalistic mapping scheme that directly transforms the latent representations of the foundation model into hydrological estimates, allowing us to assess whether the backbone features alone contain sufficient information for soil moisture inference. The second decoder employs a more expressive architecture capable of capturing multi-scale spatial dependencies and structural coherence in the output fields. A key architectural distinction between the two decoders lies in their input configuration: the simpler decoder relies exclusively on backbone representations, whereas the more advanced decoder additionally incorporates prior hydrological state information to reinforce physical consistency and temporal continuity.

Our results indicate that both lightweight decoders successfully reconstruct patterns of hydrological variables (e.g., soil moisture), demonstrating that the weather foundation models' backbone contains sufficient information to infer hydrological variables effectively. This study highlights the immense potential of weather foundation models as a new paradigm for hydrological research, providing a stable and efficient pathway to achieve high-fidelity results without the need for exhaustive fine-tuning.