Feature development

This page tracks a few larger ideas that are still open for future development. These are not part of the core v1 release, but they are promising directions if you are looking for a substantial contribution.

Automatic shoreline extraction

Delta workflows currently require a user-supplied shoreline. While the strategy for creating shorelines is documented in Shoreline creation, a built-in shoreline generator would make delta workflows much easier to automate.

Several published methods, such as opening-angle approaches, exist for this problem. In practice, the challenge is less about generating a coastline-like line and more about producing a robust line that intersects outlet links cleanly and works well with RivGraph’s pruning logic.

Lake and wetland connectivity

Many deltas, especially Arctic ones, are rich in lakes and wetlands. Their connectivity has implications for routing, transport times, and biogeochemical exchange. Extending RivGraph to represent connected non-channel water bodies remains an interesting but difficult problem.

An example mask with labeled channel network (gray) and lakes (white).

A future implementation would likely allow a second mask, or additional labels within a mask, to represent lakes or wetlands connected to the channel network. The hard parts are topology and directionality: such features can behave as sources, sinks, storage, or bidirectional connectors depending on the application.

Machine-learned flow directions

The current scheme for setting flow directions is rule-based and physically motivated. A machine-learned alternative could be interesting, especially if it used the existing algorithms as feature generators or weak labels.

The main challenge is that flow-direction assignment is constrained: interior sinks and sources are generally not allowed, and cycles must be handled carefully. Any learned approach would likely need to preserve those network-level constraints rather than treating each link independently.

Flow-direction uncertainty

Some links have genuinely uncertain or bidirectional flow. One possible extension would be to represent that uncertainty explicitly rather than exporting only a single best-guess adjacency matrix.

A useful approach might be to export a family of plausible adjacency matrices, ideally paired with a strategy to prune links that have little influence on flux routing so the uncertainty representation stays manageable.