The Birth of RivGraph

In 2016, I walked into the postdoc office at the Saint Anthony Falls Laboratory to see deltas on display as 8.5 x 11 sheets of paper taped together with hand-drawn markings all over. My colleagues Alejandro Tejedor and Anthony Longjas were trying to whip these deltas into shape for a series of papers (1, 2) introducing a new way to think about deltas: through the lens of their channel networks. As I looked at their handiwork, I couldn’t help thinking that there must be a better way. My own research at the time had led me to develop RivMAP, a Matlab toolbox for analyzing the morphodynamics of meandering rivers using binary masks. And so RivGraph was born as a small set of Matlab scripts with very limited functionality.

I finished my PhD with RivGraph in a barely-formed state, and followed my adviser, Efi, to UC-Irvine for a short postdoc. Alex and Anthony were pushing the deltas work even further, and I kept adding scripts to RivGraph. The inevitable idea was hatched to analyze dozens of deltas worldwide, and eventually became a successful proposal. In the meantime, I took a postdoc position at Los Alamos National Laboratory with an eye toward global river morphodynamics. My arrival at LANL marked the beginning of my Python journey, and soon I had converted all the RivGraph scripts to Python. My research needs led to a fuller development of RivGraph and brought braided rivers into the mix, and sometime in 2017, RivGraph officially became RivGraph.

I cut my Python teeth while developing RivGraph, so there is a degree of inherited clunkiness and inefficiency baked in. Initially, there were also some implements (such as automatically adding artificial nodes to parallel edges) that were tailored to somewhat particular use cases. However, as the user base continues to grow, improvements and upgrades have been implemented to meet their needs. Please add your requests to the mix and report any bugs you find using Github’s issue tracker.

What’s different about RivGraph?

RivGraph fills a void in coding space by providing a full-featured package for working with masks of river channels. In my opinion, most of RivGraph is just a convenient collection of already-existing functionality like CRS handling, path management, etc. That said, there are three somewhat novel components of RivGraph that are not available elsewhere, or at least not available in Python.

The first of these is the walking algorithm that breaks a skeleton into its constituent links and nodes. While there is similar functionality available in GIS packages, RivGraph ensures that the resulting skeleton is parsimonious–i.e. contains as few nodes as possible while fully preserving the topology of the mask. This is achieved with help from a double-convolution, where the first convolution identifies possible branchpoints and the second reduces branchpoint clusters to a minimum-required set.

The second novelty RivGraph offers is its ability to automatically set flow directions. This is trivial for some networks (like Wax Lake Delta), but many channel networks are wild beasts. RivGraph’s solution is correspondingly complicated, but does a pretty good job. We published the method, its validation, and its implications.

The third novel component of RivGraph is its abilty to generate an along-channel mesh that approximately follows a river’s centerline while transecting the centerline approximately perpendicularly. This was a surprisingly tricky function to get right, and I’m not even sure it’s there yet. The first iteration of this appeared in RivMAP, and there have probably been 3-4 method changes before settling on the current version, which uses Dynamic Time Warping (thanks Zoltan et. al for introducing this to me, although I still contend that it’s inappropriate to use for measuring channel migration rates) to iteratively map vertices on buffered centerlines away from the original.

Is RivGraph only for channel networks?

While RivGraph is designed around channel networks, it contains a smattering of tools that can be useful across a broad range of analyses. For instance, the script contains tools that will convert any binary mask to a vectorized skeleton, not just river channel networks. There are a number of image processing tools in that I use frequently in other projects, like a Matlab-like implementation of regionprops() for measuring blob properties of a binary image. I have personally found one of the most broadly useful tools in RivGraph is write_geotiff, which does what it says.