Watershed Modeling

Established workflows for the generation of a watershed model from a DEM requires a minimum of 7 processing steps which can be generally described in three stages, preprocessing, conditioning, and watershed modeling. 

Preprocessing is the simplest aspect of this workflow, from a technical perspective, however it can be far from straightforward. Determining a study extent and finding the appropriate data sets can be one of the most difficult challenges of any study. Selecting a study area that is either too large or too small can severely limit what questions can be answered by your analysis. In the context of a watershed analysis, it may not be immediately obvious how large your study area needs to be without prior knowledge. For example, while the primary focus of this project was to define the watersheds of the Olympic Peninsula, the study area had to be substantially extended to the South to sufficiently capture the Chehalis River watershed (Figure 10, large teal colored watershed). Even so, a segment of the watershed was unintentionally cut off in the East by the study extent. If future research were to focus on the Chehalis River specifically, significant data re-processing would be required. Additionally, while it may be tempting to select the highest resolution data available, this comes at an additional cost. File storage and downloading times are directly correlated to data resolution, and, perhaps most significantly, data processing times can increase exponentially with increased resolution. Using ArcGIS Pro software on a Windows 11 desktop with an AMD Ryzen 7 5700G processor and 16 GB of RAM, multiple of the processing steps outlined above took greater than 1 hour to run, and several processing steps had to be strategically scheduled to run overnight. Depending on the study extent and the resources available, using the highest resolution data may not be advisable, or indeed even possible.

Hydrologically conditioning a DEM is perhaps the most critical part of creating a watershed model. Simply put, conditioning comprises identifying and subsequently filling sinks. A sink is a pixel (or group of pixels) which are lower in elevation than all the pixels surrounding it. They are so-named because they resemble a sink without a drain. Natural sinks can and do exist, but they are more likely to be errors within the DEM. While it is true that lake bottoms are natural sinks, it is not typically true that there is no outflow for most natural lakes. Moreover, the objective of a hydrologically conditioned DEM is not to accurately model lake bathymetry, but instead to model water flow, which is more accurately represented by water surface elevation rather than lake bottom. Therefor, it is critical that sinks are identified and filled prior to performing a watershed analysis. In some instances, it may not be appropriate to fill every sink, and it is important to know the context of your specific study area.

Only once you have conditioned your DEM can you proceed to developing your watershed model. At minimum, a watershed model requires a flow direction raster and pour points, which describe the flow of water across a surface and river outlets respectively. In some cases, you may know (or be given) pour points a priori. However, it is often not immediately apparent where the appropriate pour points lie. The majority of the workflow described in figure 8 is dedicated to locating and creating pour points. Assuming that no additional data is available beyond the input DEM it is possible, although not trivial, to find the pour points.

Firstly, flow direction and flow accumulation intermediate layers must be calculated. Then, from the flow accumulation layer, a rasterized model of streams can be created using a conditional statement. Depending on the study area and resolution of the flow accumulation raster, the minimum accumulation value for a stream will vary. For instance, a stream in an arid region with a high resolution dataset will require a much higher accumulation value before forming than a very wet area using low resolution data. Depending on the precision needed for your study it may be possible to compare your stream raster to existing basemaps using a variety of threshold values (as shown here), otherwise detailed ground truthing through field surveys may be required. Once you have determined your threshold flow accumulation value and created your streams raster, you may optionally calculate the stream order using either the Strahler or Shreve methods. Both methods have their benefits, however the Shreve method requires substantially longer processing times, therefor I opted to use the Strahler method for this analysis. While not strictly necessary, this is a highly recommended processioning step because it adds additional detail to the streams raster which can be used to easily identify tributaries, further categorizing and differentiating streams and rivers. Similarly, while not mandatory, it is highly advisable to convert your streams raster to a vector feature class. Vectors are not only a more pleasing and intuitive graphical representation of streams, they also support snapping, which will massively aid in the creation and placement of pour points. Unfortunately, in the absence of a highly detailed coastline polygon or polyline layer, it is not easy to automate the generation of pour points. This step requires careful examination of the streams data layer, and comparison to existing base maps will likely be required. Compromises and simplifying assumptions will likely be required to limit the number of pour points in a given study. For example, watersheds in this study area were limited to outlets of Strahler stream order 5 or greater. This reduced the number of pour points to be created from thousands to 227, while still providing adequate coverage of the study area. Nevertheless, creation of 227 precisely placed pour points was not a trivial undertaking. It required hours of careful, detail oriented work with substantial fine-tuning of point placement. Too few pour points may not sufficiently cover the study area, while too many pour points will be arduous to create and may result in errors in the subsequent watershed analysis. Once pour points have been created, they are then rasterized and snapped to the associated flow direction raster. Finally, the pour points raster is combined with the flow direction raster to generate a watershed model.

The Olympic Peninsula is a highly complex, mountainous region dominated by mixed conifer forests. It is over 9000Km2, largely rural, and much of the land is federally protected. Notably, it is home to the Olympic National Park, which alone consists of 18 different watersheds (not including the park beaches). Among the watersheds identified in this study are the Chehalis River watershed, the second largest in the state, and the Skokomish watershed, which will be explored in detail in Raster Analysis - Part 2.

The watershed model developed in this exercise will be utilized and further evaluated to create a flood risk analysis within the Skokomish River watershed, and is expected to be of ongoing use in future research projects.