Transferability of data-driven models trained in data-rich regions to unknown domains

Ascertain the transferability of data-driven hydrological models trained in data-rich regions to geographically and climatologically unknown target regions when the target region’s climate, hydrology, and environmental characteristics are absent from the training dataset.

Background

While data-driven models can perform well in data-rich scenarios, the paper highlights uncertainty about their ability to generalize to locations with climates and environments not represented in training data.

This unresolved issue affects the reliability of global-scale predictions, particularly for ungauged or remote basins with distinct hydro-climatic regimes.

References

However, the transferability of models trained in data-rich regions to unknown regions remains uncertain, especially when the characteristics of the climate, hydrology, and environment of the target location are not included in the training dataset.

— Approaches for enhancing extrapolability in process-based and data-driven models in hydrology  (2408.07071 - Shi, 2024) in Section 1 Introduction

Although we test on rural regions, we cannot yet validate whether these findings transfer to other regions that lack the ground-truth labels or the geospatial inputs necessary to train the model. We see this as a limitation and note it as an important area for future work. Applying this model to other livability or socioeconomic benchmarks outside the Netherlands would let us test whether the preference of AlphaEarth and multi-embedding combinations would hold on out-of-distribution areas that these embeddings were meant to be validated on.

— Applying foundation model embeddings towards urban livability evaluation  (2609.09429 - Khot et al., 8 Sep 2026) in Limitations subsection, Section 5 (Conclusion)

Reviews of deep-learning flood mapping identify generalization to unseen cases and uncertainty treatment as open problems.

— A Controlled Evaluation of Model Rankings and Input Reliance in Surface Water Segmentation  (2608.30895 - Phunjanna et al., 31 Aug 2026) in Section 2, Related Work, paragraph “SAR water mapping and ancillary information”