Drainage Pattern Recognition Data for Graph Transformer Models

A dataset for recognizing five typical drainage patterns—rectangular, pinnate, dendritic, parallel, and fanned—using vector river network data converted to directed graph structures. The data was used to train a Gated Sparse Graph Transformer (GSGT) model, achieving an accuracy of 95.72% on a US dataset and 92.32% in a cross-regional test on New Zealand data. The dataset was authored by hao Zhang and last updated on June 4, 2026.

Use Cases

• Training graph neural networks for drainage pattern classification based on hierarchical topology and reach geometry features.
• Improving watershed hydrological simulations based on recognized drainage pattern types.
• Aiding geological structure exploration by analyzing correlations between lithology and drainage patterns.
• Benchmarking new algorithms for river network analysis against the reported GSGT model performance.

Strengths

• Model performance is quantified with specific accuracy scores: 95.72% on US data and 92.32% on New Zealand data.
• The dataset integrates multiple feature types, including hierarchical topology, reach geometry, basin structure, topography, and lithology.
• The dataset is licensed under CC-BY-4.0, allowing for open sharing and modification.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• The dataset size is 40.7 MB, which is relatively small and may limit the complexity of models it can support.

Provenance

Source

figshare

Collection Method

Vector river network data converted to directed graph structures.

Freshness

Last updated 2026-06-04 11:08:04; freshness should be verified.

Geography

Includes data from the United States and New Zealand.