Explore the learned landscape of US river basins.
Drag the timeline to explore an individual year.
Basin2Vec encodes climate, land cover, soil, terrain, and population into a 64-dimensional representation. This atlas loads the full pretrained vectors for all 9,067 basins and 40 years, 1985–2024.
Embeddings use the first three principal components as RGB color. A fixed PCA basis keeps colors comparable across years within a representation. PCA is a visual summary, not the similarity metric.
Similarity is cosine similarity in all 64 dimensions, from −1 to +1. The map stretches the 5th–95th percentiles for contrast; the legend shows the actual range and rankings always report raw cosine scores. The selected basin is excluded from its own neighbors.
Change is 1 − cosine(1985, 2024). It measures a change in the learned representation, not a measured environmental trend or a causal attribution.
Projection · z is the final contrastive embedding, also distributed in the public Python package. Hydrologic · h is the representation before projection. Each period mean is computed independently in its own space and L2-normalized.
A map click checks every display polygon at that location and lists all hits, smallest first. Smaller basins are drawn above larger ones. The Overlaps tab shows recorded strong overlaps: intersection area divided by the smaller basin’s area, calculated in EPSG:5070. A score near 1 suggests near containment. Missing links do not prove that basins are disjoint.
This dataset covers the contiguous United States; it does not include Alaska, Hawaii, or territories. Boundaries are simplified for display, so edge selection is approximate. Station names and regions come from GAGES-II. Vectors retain float32 precision and load by year. No synthetic basins or embeddings are used.