Predicting historic riparian vegetation in the Columbia River basin

We developed a GIS data set that depicts pre-settlement riparian vegetation in the Columbia River basin to guide stream restoration for endangered salmon. To do this, we first created a data layer of historic riparian vegetation information from survey notes that were taken mid 19th to early 20th century during the Public Land Survey System (PLSS) conducted by General Land Office (GLO). Our reconstructed riparian vegetation data include randomly sampled basin-wide data (drainage area >200,000 km2), as well as intensively reconstructed watershed-level data (>3,000 km2). Second, based on the reconstructed riparian vegetation points, which are arrayed along a 1-mile (1600 m) grid, we are developing statistical models to estimate potential historic riparian vegetation types (conifer, hardwood, willow-shrub, grass, sage) as well as the probability of occurrence of individual species at stream reach level (~ 200 m) in the basin. We examined environmental variables, such as mean annual precipitation, average minimum and maximum temperature, channel gradient, channel bankfull width, floodplain width, and fine sediment supply potential, against five vegetation types and found that precipitation and temperature discriminate vegetation groups. We also developed vegetation response curves against each variable using kernel density estimates to describe the probability of each vegetation type occurring across the range of each environmental variable. Using a decision tree, we found that reaches greater than 8 m bankfull width (bfw) tended to develop riparian vegetation that is distinctly different from upland vegetation, whereas in small streams the riparian and upland vegetation were similar. Therefore, we analyzed the two channel size classes separately. It is notable that this 8-m threshold is identical to the threshold of channel migration in the study area, which was identified in a previous study (Hall et al. 2007). We adopted linear discriminant analysis (LDA), support vector machine (SVM), random forest (RF), and k-nearest neighbor (KNN) to predict riparian vegetation types. Overall accuracy of models for large reaches was 68.3% (LDA), 73.0% (SVM), 71.4 (RF), and 68.3% (KNN), and 59.3%, 66.3%, 51.2%, and 47.7% for small reaches. We tested each model against data points outside of the training set, and found that overall accuracies were 47% to 48% for the large streams and 43% (all models) for small reaches. Even though overall accuracies were relatively low, we recognized that structure of error matrices reflected vegetation responses against environment variables. We are currently developing species occurrence models. We believe that, using the predicted vegetation group and species occurrence map along with vegetation response curves, we can reasonably estimate reference riparian condition in the Columbia River basin and our approach can be applicable to other areas in the US.