This page will hosts the supervised machine learning models that predict wood jam dynamics. Specifically, these models predict the probability of a wood jam:
- Mobilizing (jam either moves downstream or loses enough wood to no longer be considered a jam in the same location it was previously)
- Losing wood
- Accumulating wood
- Contracting (decrease in volume independent of losing pieces)
- Expanding (increase in volume independent of gaining pieces)
As of December 2018, there isn’t enough data to build a robust predictive model. However, to illustrate the potential of this tool, we have included a preliminary, example model below, which we also discuss in Scott et al. (2018, preprint) and Scott et al. (2019, post-print) . We present this model with the disclaimer that this preliminary model should not be used to make predictions about real wood jams, as it is insufficiently accurate to do so.
We (Scott, Wohl, and others) are continuing to collect data and will resurvey over 320 natural and engineered wood jams by Fall 2019. While this additional data will help improve the model, it will eventually need data input from more diverse rivers than we have time to access. If you are interested in this tool, we encourage you to read the Data Collection page and Scott et al. (2018, preprint) or Scott et al. (2019, post-print) to learn how to collect data and submit it to the database. Contributing to WooDDAM is quick and easy, requiring minimal field work and facilitated by the user interface on this website. By submitting data to the database, you will help make these models more accurate and ready for real-world usage faster. Please feel free to contact Dan Scott with questions.
Statistical Model Summary
Please see Scott et al. (2018, preprint) and Scott et al. (2019, post-print) for a detailed explanation of how these statistical models are built. Here, we only summarize the key characteristics of the models.
The supervised machine learning models of wood jam dynamics utilize multiple logistic regression to predict the probability of occurrence for each of the 5 binary variables listed above. Note that these modes of change are not exclusive, with the exception of mobilization (e.g., a jam could both accumulate and lose wood). Each model represents a single logistic regression that uses multiple predictor variables (e.g., whether the jam touches the bed, whether the jam is pinned on a relatively immobile object) to estimate the probability of a given form of change. Along with probabilistic predictions, each model is presented with odds ratios for each predictor variable to aid in interpreting why a given prediction was made.
As new data is added to the database, we will regularly check and update models to maximize their predictive performance. As more data becomes available, we will eventually host all 5 predictive models on this page, enabling users to upload a single .csv file containing wood jam characteristics and receive an output file with predictions of wood jam dynamics under variable flow scenarios for each wood jam.
Please see below for a preliminary look at the model to predict wood jam mobilization.
Preliminary Model of Wood Jam Mobilization
DISCLAIMER: This model is for educational purposes only and we discourage the use of the predictions provided at this time. The database is currently too limited to train robust predictive models, but will likely be sufficient to produce robust models soon. In evaluating the model below, for instance, it only predicts mobilization correctly when mobilization actually occurred around 25% of the time. With more data, especially from diverse rivers and structurally varied jams, we hope model performance will improve substantially. Of course, that’s a hypothesis: help us test it by contributing data to the database!
Because this model is shown as an example only, it does not allow users to make predictions for multiple jams at once (this will be enabled once the model is sufficiently robust). Please use the checkboxes below to customize the example wood jam, then view the resulting prediction of how likely it is to mobilize.
A prediction is given for flows near and above bankfull stage. To date, we haven’t observed jams mobilizing at flows below bankfull, so the predicted probability of mobilization for below bankfull flows is effectively 0%, although we don’t wish to imply that it is impossible.