Basin2Vec: Learning Basin-Centric Embeddings for Scalable Hydrological Modeling

Muhammad Talha; Xiaofeng Liu; Yuan Yang; Pang-Ning Tan; A. Pouyan Nejadhashemi

cs.LGNov 3, 2026

Basin2Vec: Learning Basin-Centric Embeddings for Scalable Hydrological Modeling

Authors: Muhammad Talha, Xiaofeng Liu, Yuan Yang, Pang-Ning Tan, A. Pouyan Nejadhashemi

Organizations: Michigan State University · University of Michigan · University of California, San Diego

Abstract

Reliable streamflow prediction across time and in ungauged basins requires representations that capture hydrologically relevant basin characteristics. Existing methods rely on either fixed handcrafted attributes or general-purpose Earth observation embeddings not tailored to basin structure. We introduce Basin2Vec, a self-supervised framework that learns compact basin representations from hydroclimatic and physiographic data without observed streamflow. Basin2Vec organizes representation learning around basin scale and relationships across space and time. Across multiple forecasting architectures, it consistently outperforms predefined attributes and general-purpose geospatial embeddings in temporal generalization and zero-shot spatial transfer. Basin2Vec provides average relative gains of up to 8.3% temporally and achieves 3.5x the median predictive performance of the strongest baselines on unseen basins. These results demonstrate the effectiveness of basin-centric pretraining for transferable hydrologic prediction. Code is available at https://github.com/Muhammad-Talha-MT/basin2vec

Figures & tables

Panel of Table 1 — Table 1: Comparison of basin representations across three design requirements. Scores indicate High, Moderate, or Low support.

Figure 1: Overview of Basin2Vec. The scale-aware encoder maps monthly and quasi-static basin-aligned grids, masks, and metadata to normalized basin-year embeddings $z_{b,y}$ through mask-aware encoding, multimodal fusion, and area regularization.

Figure 2: Basin relations used by the relation-aware contrastive objective.

Figure 3: Representation diversity. Normalized effective rank across 9,067 basins. Higher values indicate greater representation diversity.

Figure 4: Metadata consistency.

\ell_{2}

-distance distributions for metadata-similar and metadata-dissimilar non-overlapping pairs. Dashed lines denote means.

Figure 5: Per-basin KGE CDFs for temporal generalization using selected downstream architectures.

Figure 6: Spatial-split performance across basin-size classes.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Group	Variable	BASINGRID type	Basin2Vec input	Unit	Source
Hydroclimate	Precipitation	Daily	Monthly dynamic	mm/day	Daymet
	Minimum temperature	Daily	Monthly dynamic	∘ C	Daymet
	Maximum temperature	Daily	Monthly dynamic	∘ C	Daymet
	Downward shortwave radiation	Daily	Monthly dynamic	W/m 2	Daymet
	Snow water equivalent	Daily	Monthly dynamic	kg/m 2	Daymet
	Vapor pressure	Daily	Monthly dynamic	Pa	Daymet

Appendix

Table A.2: BASINGRID variables used for Basin2Vec pretraining. Native daily variables are aggregated to monthly resolution before pretraining. Source definitions follow BASINGRID [ 33 ] .

Figure A.1: Spatial coverage of the 9,067 BASINGRID basins used for Basin2Vec pretraining. Colors denote HUC-2 hydrologic regions, and major rivers are shown in light blue.

Variable	Quantity
log_area_km2_z	Drainage area
log_perimeter_km_z	Basin perimeter
log_equivalent_diameter_km_z	Equivalent diameter
log_bbox_width_km_z	Bounding-box width
log_bbox_height_km_z	Bounding-box height
log_bbox_diagonal_km_z	Bounding-box diagonal

Appendix

Table A.3: Standardized basin metadata fields supplied to Basin2Vec. The final three overlap-summary fields are constant zero in the training metadata file; the remaining 15 fields are nonconstant.

Group	Attributes
Climate	p_mean , pet_mean , p_seasonality , frac_snow , aridity , high_prec_freq , high_prec_dur , low_prec_freq , and low_prec_dur
Topography	elev_mean , slope_mean , and area_gages2
Vegetation	frac_forest , lai_max , lai_diff , gvf_max , and gvf_diff
Soil	soil_depth_pelletier , soil_depth_statsgo , soil_porosity , soil_conductivity , max_water_content , sand_frac , silt_frac , and clay_frac
Geology	carbonate_rocks_frac and geol_permeability

Appendix

Table A.4: CAMELS-US attributes used as the hand-crafted basin descriptor baseline. The descriptor contains 27 variables.

Setting	Protocol
Task	Daily streamflow prediction, lead $0$ , horizon $H=1$
Forcing windows	$L\in\{120,365\}$ days
Dynamic inputs	Five Daymet meteorological forcings through target day $t$
Static input	One fixed basin descriptor
Discharge input	None
Temporal development	Same 671 basins, 1990–2009

Appendix

Table A.5: Downstream streamflow-prediction protocols.

Architecture	Hyperparameter	Temporal	Spatial
LSTM	Hidden dimension	128	256
	Recurrent layers	2	2
Attention-based	Model dimension	128	128
	Attention heads	8	8
	Encoder layers	4	2
	Feed-forward dimension	512	512

Appendix

Table A.7: Model hyperparameters used in downstream experiments.

Figure A.2: Example reference-basin similarity maps for RQ1. Each panel shows the cosine similarity between one selected reference-basin descriptor and all Basin2Vec basin descriptors. The reference basin is outlined in black. Similar basins occur both near and far from each reference, indicating that the learned neighborhoods are not determined solely by geographic proximity.

Figure A.3: Basin-level performance distributions for RQ2 temporal generalization. Rows correspond to forcing-window lengths

L=365

and

L=120

, and columns correspond to downstream forecasting architectures. The CDFs report NSE, while the accompanying boxplots summarize KGE across the same evaluation basins.

Figure A.4: Basin-level performance distributions for RQ3 zero-shot spatial transfer at

L=365

. The top row reports NSE and the bottom row reports KGE across held-out basins. Columns correspond to downstream forecasting architectures, and curves compare Basin2Vec, AlphaEarth, CAMELS attributes, TESSERA, and SatCLIP.

Hydrology

See all 26 papers

Utilizing Earth Foundation Models to Enhance the Simulation Performance of Hydrological Models with AlphaEarth Embeddings
Jan 4, 2026Pengfei Qu, Lei Ye, Wenyu Ouyang +6Streamflow Forecasting Time Series Forecasting
School of Computer Science and Technology, Dalian University of Technology, Dalian, China · School of Infrastructure Engineering, Dalian University of Technology, Dalian, China · School of Information and Communication Engineering, Dalian University of Technology, Dalian, China +2
Evaluating Transformer and LSTM Frameworks for Prediction in Ungauged Basins
Jun 1, 2026Taye Akinrele, James Halgren, Noorbakhsh Amiri Golilarz +2Streamflow Forecasting Transformer
Dept. of Computer Science, The University of Alabama, Tuscaloosa, USA · Alabama Water Institute, USA
Learning global river system hydrodynamics with a topology-informed foundation model
Feb 25, 2026Hancheng Ren, Gang Zhao, Shuo Wang +12Hydrology Streamflow Forecasting
College of Water Sciences, Beijing Normal University, Beijing, China. · School of Geography and the Environment, University of Oxford, Oxford, UK. · Department of Transdisciplinary Science and Engineering, Institute of Science Tokyo, Tokyo, Japan. +6

Jan 4, 2026J/K move · Enter open · S save

Utilizing Earth Foundation Models to Enhance the Simulation Performance of Hydrological Models with AlphaEarth Embeddings

Pengfei Qu, Lei Ye, Wenyu Ouyang, Chi Zhang+5

School of Computer Science and Technology, Dalian University of Technology, Dalian, China · School of Infrastructure Engineering, Dalian University of Technology, Dalian, China · School of Information and Communication Engineering, Dalian University of Technology, Dalian, China+2

Streamflow Forecasting Time Series Forecasting

Open paper AskPDF arXiv

Accurate prediction in ungauged basins (PUB) requires effective transfer across hydrologically diverse regions. While deep learning has advanced large-sample hydrological modeling, its cross-regional generalization depends in part on the static basin descriptors used to guide model behavior. Earth Foundation Models (EFMs) provide a promising alternative by learning environmental representations from multi-modal satellite data. In this study, we evaluate the 64-dimensional AlphaEarth Foundations (AEF) embeddings as a new form of static basin representation and assess their value for streamflow prediction under both in-sample (IS) and out-of-sample (OOS) settings using the CAMELS dataset. Results show that AEF embeddings offer higher OOS robustness than CAMELS attributes, despite nearly identical IS performance. Mutual-information analysis reveals statistical associations between AEF embeddings and conventional catchment attributes related to terrain, vegetation, and rooting characteristics. We further investigate similarity-based donor-basin selection for PUB using CAMELS attributes and AEF embeddings. Across 100 target basins, AEF static inputs generally improved PUB performance. With AEF inputs, similarity-based donor selection outperformed random selection for smaller donor pools, but its advantage diminished as pool size increased. Adding more donor basins did not consistently improve performance. Leave-one-cluster-out experiments showed AEF input benefits depended on basin partitioning, suggesting that transferability remains sensitive to the catchment conditions represented during training. Overall, this study demonstrates that Earth Foundation Model embeddings provide a transferable representation for PUB prediction and underscores the importance of basin representation quality and donor pool design, motivating future research toward unified hydrological foundation models.

Abstract

Hydrology

Explore similar work

Utilizing Earth Foundation Models to Enhance the Simulation Performance of Hydrological Models with AlphaEarth Embeddings

Evaluating Transformer and LSTM Frameworks for Prediction in Ungauged Basins

Learning global river system hydrodynamics with a topology-informed foundation model