Skokomish River Flood Risk Analysis

Reclassification

All risk factor layers had to be reclassified to a common, unitless scale prior to being used in the flood risk classification (Figure 3). Firstly, the Skokomish River watershed raster was reclassified to a binary scale, with 1 being within the watershed and 0 being outside. This is equivalent to a detailed watershed boundary polygon, which was used to further clip subsequent layers. Next the hydrologically conditioned DEM was reclassified to a 1 to 10 scale with 5 classes (even integers) corresponding to the Jenks natural breaks method with lower elevations equal to 10 and highland equal to 2. Similarly, the slope layer was reclassified to a 1 to 10 scale with 5 classes (even integers) also using the Jenks natural breaks method. The steepest slopes were classified as 2 and the flattest slopes equal to 10. The land cover classification was given as a categorical raster with a 30-meter spatial resolution and 21 unique classes contained within the study area. A new flood risk field was added to the raster attribute table and flood risk was classified to a 1 to 10 scale (even integers) based on my educated assessment of ecological flood risk according to table 1. 

Prior to classifying the proximity to stream risk factor the distance had to be calculated. Using the Euclidean Distance tool and the streams raster layer, the proximity to the nearest stream was calculated across the entire study area. All locations within the study extent were within 167 meters of a stream, with a mean distance of only 13 meters. Again, the distance to stream layer was reclassified as with the DEM and slope layers, with the cells nearest to a stream equal to 10 and those farthest away equal to 2.

The MNDWI and NDBI layers were then used to identify and classify open water bodies within the study area. While the MNDWI is designed to identify open water, it can also highlight developed areas, particularly within a highly forested area. Therefore, in order to further increase the usable range of the MNDWI and especially to remove incorrectly classifying urban areas the NDBI was subtracted from the MNDWI, resulting in a modified MNDWI as seen in Equation 3. Water bodies were then selected from the modified MNDWI layer using a conditional statement selecting values ≥ -0.1 and the subsequent raster layer was reclassified into a binary scale, with open water equal to 0. Additionally, the binary water bodies layer was smoothed, and minimal water bodies were removed using a 3x3 rectangle, majority focal statistic. 

From disasters to crime to disease, using raster datasets can be an efficient and powerful way to model risk across space. Rasters are used to map a wide variety of continuous data. In addition to data derived from DEMs, rasters are also frequently used for remotely sensed data such as satellite imagery, radar, and LiDAR. Rasters can also be used to map non-continuous, categorical data such as land cover and political preference. Even phenomena that are typically mapped using vector data (points, lines, and polygons) are able to be mapped using rasters. Point data such as river outlets (i.e. pour points, see Raster Analysis - Part 1), lines such as streams, and even complex polygons such as a watershed can all be modeled using a binary raster. While nearly all phenomena could be modeled using rasters, a key limitation of raster datasets is resolution. Similar to zooming into a digital image, rasters have an inherently finite resolution. Increasing the scale of a raster layer does not add new information. While it is possible to resample raster layers to different scales using techniques such as interpolation and/or aggregation, these techniques must be used with caution as they do not change the underlying dataset and are liable to be over interpreted. In contrast, vector data sets are capable of being infinitely precise, and data can easily be modified to meet the required precision of a given study.

In this case study, we have seen how raster data can be used to break down and map complicated phenomena such as flooding into a variety of risk factors. Those risk factors can then be classified, weighed, and combined using raster math to map relative risk levels across a large and complex study area. This analysis utilized six raster datasets derived from either (1) a hydrologically conditioned DEM (processed in Raster Analysis - Part 1), (2) Landsat satellite imagery, or (3) categorical land cover data. Each of these datasets had to be reclassified to a common, unitless scale of relative risk. Finally, raster math was used to determine the overall flood risk as well as to calculate various intermediate layers such as MNDWI and NDBI (Equations 1 to 4).

Importantly, I choose not to include precipitation in this analysis. This was primarily for two reasons. (1) Although average precipitation data is readily available across the US from projects such as Oregon State University's, Parameter-Elevation Regression on Independent Slopes Model (PRISM) project (Kittel, et. al., 1997), available data is much coarser than the other datasets used in this analysis, typically 4Km resolution or greater. (2) Furthermore, because the study area is a single watershed, the entire study area is hydrologically connected. In the absence of highly technical and detailed hydrologic modeling use of precipitation data as a risk factor is likely to underestimate flooding risk in the lowlands, because, the risk model has no way to account for precipitation falling in the uplands of the catchment flowing down into the valley. While hydrologic modeling is a fascinating and valuable subject, it requires a highly robust dataset including data such as river transects, wetted width, riverbed composition, etc., which are not currently publicly available for the Skokomish River. Therefore, although precipitation is a known risk factor which is commonly used in flood risk analysis, in this case it was not found to add any additional utility to the study.

The Skokomish is a wild river. In one of the rainiest regions of the US, with most of its tributaries lying within mountainous Federal land, and emptying into a wide, flat river valley. Additionally, the history of extensive logging in the upper watershed is thought to be a major contributor to sediment loads within the river (Army Corps of Engineers, 2015). This additional sediment transport contributes to river channel destabilization, bank erosion and channel migration, all of which increase flooding risk (USGS, 2023). Furthermore, according to the University of Washington's, College of the Environment, Climate Mapping for a Resilient Washington project, the average annual precipitation for Mason County, WA is expected to increase by as much as 8.8% over the coming decades. With all these risk factors, it is no wonder that the Skokomish River is the most flood prone river in the state. 

Despite the known risk, the Skokomish Valley is still an important economic area. The thick, fertile, glacial soils support healthy and productive rural agriculture, and the remote forested mountains attract outdoor enthusiast for both recreation and its unique lifestyle. There is also a relatively large area of the lowlands owned by the Skokomish Indians, and the entire region is a refuge for a wide variety of fauna, including endangered species. All of these stakeholders, public, private, and wild are impacted by annual flooding. Obvious costs associated with flooding are blocked access and the potential for infrastructural damage. For example, the most important access roads to the valley (West Bourgault Road and Skokomish Valley Road) are regularly blocked by floodwaters and both US HWY 101 and WA HWY 106 lie within areas of extreme flood risk. If either of these major highways flooded, it could cut off access to tens of thousands of residents in the Olympic Peninsula, potentially requiring a detour of hundreds of miles. Less obvious costs incurred by flooding include increased salmon mortality due to stranding (Figure 5). 

Using raster analysis to map flood risk in the absence of highly technical hydrodynamic modeling can help public and private stakeholders understand their own risk and liability, helping people to make informed decisions and focus their efforts. While the risk of river flooding is a fact of life for residents of the Skokomish Valley, the community has adapted to annual flooding, and they are not about to let a little floodwater stop them enjoying their unique way of life.