Streamflow prediction gets a new starting point: Basin2Vec skips observed flow and reaches 3.5 times the strongest baseline's median on unseen basins
Synopsis
Basin descriptions have come either from fixed attribute tables such as CAMELS, which lose within-basin spatial variation, or from Earth-observation embeddings pretrained on pixels or coordinates, which need post hoc aggregation. Basin2Vec instead treats each basin-year as an entity, encodes basin-aligned gridded climate and physiographic data with a mask-aware convolutional encoder, and uses a contrastive objective to pull together same-basin, nearby-year, overlapping and nested basins. On 671 CAMELS-US basins it gives average relative temporal gains of up to 8.3% across five forecasting architectures and reaches 3.5 times the median predictive performance of the strongest baseline in zero-shot transfer to unseen basins.
Interpretation
A basin-level descriptor can now be learned directly from basin-aligned gridded hydroclimatic and physiographic data, without a handcrafted attribute table and without any observed streamflow. Regional hydrological models have typically compressed each basin into a fixed-length vector of CAMELS or GAGES-II attributes, and statistics such as mean elevation or forest fraction cannot express spatial variation inside the basin; CAMELS-SPAT keeps that spatial information but forces downstream models to process high-dimensional rasters. On 671 CAMELS-US basins, trained on 1990–2009 and evaluated on 2010–2014, Basin2Vec ranks first in both NSE and KGE for all five architectures LSTM, TFT, TimeXer, Timer-XL and iTransformer, with average relative gains of up to 8.3%.
The encoder maps each basin's monthly meteorological rasters and quasi-static physiographic rasters to one fixed-length vector while explicitly retaining the basin's physical scale. Because every basin is resampled to the same fixed patch, the inputs no longer carry basin area, so scale information has to be restored by a separate objective. A mask-aware convolutional encoder pools only inside the basin boundary with masked average and attention pooling, a four-head Transformer links the 12 months of the dynamic branch, and learned branch weights fuse it with the quasi-static branch; an auxiliary head regresses standardized log drainage area from the fused representation, and metadata enter through gated conditioning. Pretraining uses 1985–2024 data for 9,067 GAGES-II basins in BASINGRID.
The geometry of the embedding space is shaped by hydrological relations among basins rather than by treating every other sample as a negative. Standard contrastive learning treats only two augmented views of the same entity as positive, which is insufficient for basin-year entities that may be related by basin identity, nearby years or spatial overlap. The contrastive objective makes other years of the same basin and overlapping or nested basins positive, measuring overlap with the Szymkiewicz–Simpson coefficient; for non-overlapping negatives it downweights repulsion by metadata similarity and applies a metadata-dependent negative margin. In ablations the full objective beats every relation ablation on both temporal and spatial metrics.
Perspective
The demonstrated scope is the contiguous United States: pretraining uses gridded data for 9,067 GAGES-II basins in BASINGRID from 1985 to 2024, and downstream evaluation uses 671 CAMELS-US basins, with temporal generalization trained on 1990–2009 and evaluated on 2010–2014 and spatial transfer assessed by six-fold ungauged-basin cross-validation. The intended readers are researchers in hydrological forecasting and Earth-observation representation learning; descriptors stay frozen during downstream training, and only the input width of the conditioning vector changes.
The 8.3% and 3.5 times figures are arithmetic summaries of reported basin-median scores across architectures, not ratios of prediction errors, and they do not represent statistical significance. KG-SSL is compared only in temporal generalization because building its descriptor requires observed streamflow from the target basin, which rules it out of zero-shot transfer; the ablations assess component contributions and do not constitute a coefficient search or an optimal setting.
