"scale dependent" predicts WUA (actually WUA per linear ft, to eliminate effects of arbitrary COMID break locations) (ft) versus streamflow (cfs). Includes all the predictors including those related to the stream's flow magnitude (drainage area, etc.). Simplest to use and interpret.
"scale independent" predicts HSI (fraction of inundated area that is suitable habitat) versus normalized streamflow (cfs relative to the reach's mean annual flow). Does not include predictors that are more related to flow magnitude than habitat suitability, such as drainage area. This model may be advantageous in that it's focused on predicting the habitat question only while controlling for the scale of the drainage.
Three statistical forms:
linear regression
linear regression with lasso feature selection (uses L1 regularization to eliminate the less important predictors, then re-runs the model with only the remaining predictors)
random forest regression - machine learning approach, definitely produces the best fit, just need to make sure we are not over-fitting
To do:
[ ] Still needs plenty of tuning/optimization of the selection of parameters, transformations and interactions.
[ ] Adding more training data (more sites) will greatly improve the model
[ ] Adding missing predictor vars: aquatic, ndvi, valley width
WIP, testing predictive modeling capabilities.
Two variations of model:
"scale dependent" predicts WUA (actually WUA per linear ft, to eliminate effects of arbitrary COMID break locations) (ft) versus streamflow (cfs). Includes all the predictors including those related to the stream's flow magnitude (drainage area, etc.). Simplest to use and interpret.
"scale independent" predicts HSI (fraction of inundated area that is suitable habitat) versus normalized streamflow (cfs relative to the reach's mean annual flow). Does not include predictors that are more related to flow magnitude than habitat suitability, such as drainage area. This model may be advantageous in that it's focused on predicting the habitat question only while controlling for the scale of the drainage.
Three statistical forms:
To do: