Machine learning and analysis code with combined BEWARE 2 and Coral reef profiles for wave-runup prediction datasets

Data collection method:

This dataset was created by combining data from the BEWARE 2 and Coral reef profiles for wave-runup prediction datasets. The combined dataset was prepared for use in machine-learning experiments investigating coastal runup. The accompanying Python code was developed to test a multilayer perceptron (MLP) model and analyse the resulting model outputs.

Data processing and preparation activities:

The source datasets were combined into a single dataset and randomised prior to use in the machine-learning workflow. Normalisation scales were calculated using the training data only and are provided in normalisation_scales_training_only.csv. The processed combined dataset is provided as BEWARE_combined_randomised.parquet. The accompanying Python scripts perform the MLP testing and subsequent analysis of the model results.

Technical details and requirements:

The machine-learning workflow is implemented in Python using PyTorch. The MLP training pipeline uses NumPy, pandas, PyTorch, scikit-learn and Optuna, while the accompanying results analysis uses NumPy, pandas, Matplotlib, scikit-learn and SciPy. Required dependencies are provided in requirements.txt and can be installed using pip install -r requirements.txt. The MLP training script is run separately for each cross-validation fold using the --fold command-line argument.

Additional information:

The dataset and accompanying code are organised to support reproduction of the machine-learning workflow. BEWARE_combined_randomised.parquet contains the combined and randomised BEWARE and Scott et al. data used for model development and testing. normalisation_scales_training_only.csv contains the normalisation parameters derived from the training data and used to scale the input variables without introducing information from the test data. 1_Testing_MLP.py contains the MLP training and cross-validation workflow, and 2_Results_Analysis.py contains the scripts used to analyse model outputs and generate the reported results. The code is written in Python and uses PyTorch for the MLP implementation.