CoastLearn release

Third Release

• This is the third release of the COASTGUARD repo. The repo has been extended to a new toolset CoastLearn for generating short-term data-driven predictions of coastal vegetation edge and waterline positions along cross-shore transects.
• The goal with the CoastLearn release/toolset is to serve as a real-time early warning system for coastal change across the entire shoreface, trained and forced with solely satellite-derived coastal metrics.
• This is the last update containing an entirely new toolset; releases after this will most likely be related to bugs and issues as they arise.

New Functionality

• The transect-based data from CoasTrack is used to train, validate and test a Long-Short-Term Memory neural network;
• Various functions have been written to interpolate over the datasets to regulate them;
• A function to prep the data and compile the neural network allows for different hyperparameters to be fed in by the user;
• Training takes place using TensorFlow, with the training histories able to be viewed in a browser with TensorBoard;
• A feature importance function uses Integrated Gradients to assess the influence of each coastal metric in a transect dataframe;
• Various plotting routines can be called to visualise all of the above (full list in CoastLearn_DriverTemplate and PredictionsPlotting)

Full Changelog: CoasTrack-v1.0...CoastLearn-v1.0

Acknowledgements

• Daniel Tudor: original code snippets
• @eduardogomezdelapena: assistance in implementing ShoreShop loss function