Feature Development

There’s always more to be done! The features listed below are suggestions that require some significant amount of development. If you’re interested in tackling, or helping tackle, any of these, please email j………k@gmail.com and we’ll add you to the RivGraph slack channel to start discussions.

Automatic Shoreline Extraction

Currently, users must provide their own shoreline file for each delta they wish to analyze. While the strategy for creating shorelines is documented, a preferable option would have RivGraph automatically generate shorelines. There have been a number of published softwares that attempt to solve this problem (e.g. opening-angle method), but in practice these have been found to be too slow and/or need finegaling to interface with RivGraph.

Lake/wetland connectivity

Many deltas, especially Arctic ones, are lake-dense, and the connectivity of these lakes has implications for biogeochemical cycling and transport times of water and sediment. We have spent a few weeks developing a rivgraph-lakes branch that attempts to resolve lake connectivity. However, it’s a difficult problem and we didn’t quite cross the finish line.

image-20220107142427137

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

In this formulation, a lake mask would be provided in conjunction with the river mask, or lakes would be labeled distinctly in a river mask, as shown by the above mask. The difficulties here lie in the number of ways lakes can be connected; the figure shown is a rather simple case, but sometimes lakes themselves intersect both the channel network and the ocean/draining body. The many possibilities make simply resolving their position within the channel network a formidable task.

The second major issue arises when trying to set flow directions. Are lakes sources, sinks, or both? The flow directionality algorithms need to know! And they need to be adapted accordingly. We have made significant progress on this feature, but it is not ready for Production.

Another point to mention here is that while we are focusing on lakes here, the concept may be more generally applicable. For example, if a user also had a mask of wetlands–or any objects that are connected to the channel network, this framework could handle those cases.

Machine Learned Flow Directions

The current scheme for setting flow directions in the links of a channel network is quite complicated. It is “physically-based” in the sense that the rules defining flow directions have physical bases. There are two “hard” constraints–1) there can be no interior sources or sinks and 2) there can be no cycles (although this can be violated if RivGraph cannot find a cycle-less solutions).

The paper (and attached dataset) linked above contains plenty of “training data” that could be used to take a machine learning approach to set flow directions. Many of the algorithms in place, for example determining if a link is part of a “main channel,” or synthetically-generated slopes, can be used to generate the features of an AI approach. It’s not clear to me how you would enforce the constraints listed above, but I’ll bet it could be done. Another difficulty might be that in the “physically-based” scheme, directions are set iteratively starting with the most certain. By setting a link, that information is useful in setting its neighbors’ directions, and so-on. Continuity can only be exploited when only one link (in a group of adjacent links) has an unknown direction. I would therefore guess that a ML model would have to follow a similar iterative path.

However, who knows? An AI approach could be faster, especially for the larger deltas. Post-processing corrections would also be useful (for either approach, really).

Flow Direction Uncertainty

Many links’ flow directions are not certain and may in fact be bidirectional. One approached proposed to handle this is: rather than export a single adjacency matrix with RivGraph’s “best guess,” we could export a family of adjacency matrices. Alej Tejedor has done some recent work on the concept of “effective resolution”–i.e. the resolution of the underlying mask at which graph-based flux partitioning is significantly impacted (as we coarsen the mask). This concept could be combined with the family of adjacency matrices to prevent the family from being huge. In other words, “effective resolution” could prune links that don’t contribute much to flux routing–and these links are typically the most uncertain.