Automated mapping and extraction of Arctic water tracks
On hillslopes underlain by permafrost, water flowpaths control slope stability and release of sediment and carbon stored in soil. Low-order channels in some arctic landscapes take the form of water tracks, curvilinear features defined by subtle topographic lows and vegetation variation. Since spacing of liquid flowpaths across icy surfaces is modulated by thermodynamics, and channelization in landscapes is controlled by the balance of advective and diffusive processes, mapping these flowpaths and characterizing their spacing should reveal the thermo-mechanical forces acting on permafrost hillslopes. My work shows that water tracks in warming permafrost are areas of high topographic disturbance ('21 AGU poster), portending gullying and ecosystem disturbance. However, water tracks hydrologically behave as channels (quickly transmitting water across the surface of a landscape) but geomorphically behave more like hillslopes (they do not carve concave channel valleys), making them difficult objects to delineate with algorithms for identifying channel heads developed in temperate landscapes. Also, dense shrub growth around channels may bias elevation values in ArcticDEM, the photogrammetrically derived topographic model. See my flow routing page for a deep dive.
The goal of my work is to automate the identification and extraction of flowpaths in soil-mantled Arctic landscapes and to perform analyses on the skeletonized network. This work would be transformational because (1) there is currently no pan-Arctic map of water tracks (we don't even know where these things are, and why they are in some places but not others!) and (2) there is currently no algorithm for identifying water track "heads" or continuous flowpath. 'Computer vision" has been successfully applied to automated detection of thermokarst and polygonal ground, and water tracks have been remotely sensed and categorized in the past, but no similar effort exists to automate the mapping of water tracks as flow networks. My initial work shows that satellite-derived measures of plant and surface moisture, in addition to modeled flow accumulation, distinguish water tracks from the surrounding hillslope. The Google Earth Engine App below demonstrates how spectral indices may factor into an algorithm (toggle datasets' visibilities and opacities with "Layers"):
NDVI - Vegetation index, shown red (brown/bare) to green (green plants). Derived from 3 m resolution 4-band Planet imagery.
NDMI - Plant moisture index, shown brown (dry) to teal (moist). Derived from Sentinel-2, 20 m resolution. Seasonal standard deviation shown as white (low std) and black (high std).
NDWI - Surface wetness index, shown yellow (dry) to blue (wet). Derived from Sentinel-2, 10 m resolution. Seasonal standard deviation shown as white (low std) and black (high std).
RBG imagery - Planet 3 m resolution.
Flow accumulation - Modeled flow accumulation with flow direction defined as steepest downwards slope on planar triangular facets on a block centered grid (d-infinity), modeled from 2m topographic model from ArcticDEM, a photogrammetry-derived product.
InSAR velocity - land surface displacements averaged across late summer 2014-2019, shown no movement (dark purple) to lots of movement (yellow). Derived from Sentinel-2 and ALOS on a 15 m resolution, produced by Simon Zwieback.
August 2022 - progress update
Summer research student Rebecca Risch (Dartmouth '25) and I have been making great progress stress-testing my data pipeline, collecting imagery and creating image labels. We are learning about the best Planet Labs products for image segmentation, how senescence timing varies across the Arctic and its impact on water track visibility, and how to effectively troubleshoot coding issues! Thanks to Rebecca's determination a clearer picture has emerged for my research plan:
I'm happy to report that Rebecca has worked through the first column of Phase I, working out the kinks along the way. She's now labeling our imagery, which is cool because determining where water tracks are and aren't is a valuable dataset in and of itself! Once we work though labeling the ~25 watersheds I've randomly selected from my catalog of 100,000, we'll start on model training.
Update! 2/3/2021
I have been using Daniel Buscombe and Evan Goldstein's Doodler and Segmentation Zoo tools with great success on my Planet RGB data! This is just some out-of-the-box model parameters trained on a mere 18 images at 256x256 pixels and 3 m resolution. Time to queue up some podcasts and scribble on the tundra!
Current capabilities:
My collaborator at Appalachia State has students hand-mapping the locations of water tracks in Google Earth (polygons surrounding water tracks), which could provide a training dataset for supervised binary image segmentation/object detection in satellite imagery. A network skeleton could then be extracted from segmented water track pixels. However, we do not know how transferrable image segmentation is pan-Arctic if diagnostic vegetation patterns vary.
Cost path analysis may provide adequate network maps, though it uses a priori designation of certain layers as more important than others, and I still don't know how to map (or decide!) where water tracks "start".
Proposal:
I want to test unsupervised or semi-supervised image segmentation and perhaps semantic segmentation in order to construct the skeletonized networks. I am curious to know how an algorithm might "choose" a water track channel head as well as a likely flow path. I have been investigating geomorphic implementation of U-net architecture for segmentation in oblique imagery, but having multispectral data as data layer inputs is a slightly different use case. I also know that I can weight the structure of objects in images if I am to detect a network rather than just blobs, but I don't know enough about these models to know where to start!
I also know that some geomorphologists have used graph theory to map sediment and water cascades, which seems like an applicable analysis for my work. Since I don't know much about network analysis it would be good to make connections with folks who do!