This repository contains code and data to reproduce the results of the paper Deep Neural Networks for Active Wave Breaking Classification published by Scientific Reports.

Deep Neural Networks for Active Wave Breaking Classification

• 1. Dependencies
• 2. Data
  • 2.1. Manual Creation
  • 2.2. Production Ready
• 3. Training
  • Pre-trained Models
• 4. Model Performance
  • 4.1. Evaluating
  • 4.2. Results
• 5. Using a Pre-trained Neural Network
  • 5.1 Predicting on New Data
  • 5.2. Predicting from the Results of the Naïve Detector
  • 5.3. Plotting Wave Breaking Detection Results
• 6. Model Interpretation
• 7. Active Wave Breaking Segmentation
• 8. Wave Tracking
• 9. Wave Breaking Statistics
• Gallery
• Standard Variable Names
• Disclaimer

• Windows

conda env create -f environment_win.yml

• Linux

conda env create -f environment_linux.yml

# install GIT
conda install git

# create a new environment
conda create --name tf python=3

# activate your new environment
conda activate tf

# If you have a nvidia GPU installed and properly configured
pip install --upgrade pip
pip install tensorflow

# extras
pip install -q git+https://github.com/tensorflow/examples.git
pip install tensorflow_addons

# conda packages
conda install -y numba natsort pandas scikit-learn scikit-image natsort  matplotlib seaborn netCDF4 xarray ipython tqdm
pip install h5netcdf

# Extra thresholding methods
pip install pythreshold

# fitting circles to data
pip install miniball

# parallel computations
pip install pebble

• Refer to Manual Data Preparation.

Model	Link	Alternative link
Train (10k)		-
Train (20k)		-
Test (1k)		-
Black Sea (200k)	Upcoming	-
La Jument 2019 (10k)	Upcoming	-

|

Note: The training dataset used here is a smaller version (10k) of the published dataset so it can run on Google Colab. The 20K dataset takes over 6 hours to train and Google will disconnect your session.

The data needs to be in a folder which has sub-folders "0" and "1"

For example:

train
    ├───0
    ├───1

There are 5 backbones implemented: VGG16, ResNet50V2, InceptionResNetV2, MobileNetV2 and EfficientNet

Note that the weights from these pre-trained models will be reset and updated from the scratch here. These models have no knowledge of the present data and, consequently, transferred learning does not work well.

Example

python train.py --data "train/" --backbone "VGG16" --model "vgg_test" --logdir "logs/" --random-state 11 --validation-size 0.2 --learning-rate 0.00001 --epochs 200 --batch-size 64 --dropout 0.5 --input-size 256 256

The neural network looks something like this:

Please use the links below to download pre-trained models:

Scientific Reports (20K dataset)

Model	Link	Alternative link
VGG16		-
ResNet50V2		-
InceptionResNetV2		-
MobileNet		-
EfficientNet		-

La Jument (10K dataset)

Note: Work in progress.

|

To evaluate a pre-trained model on test data, use the test script.

Example:

python test.py --data "path/to/test/data/" --model "VGG16.h5" --threshold 0.5 -- output "path/to/results.csv"

The classification report with be printed on the screen. For example:

              precision    recall  f1-score   support

         0.0       0.88      0.99      0.94      1025
         1.0       0.87      0.23      0.36       175

    accuracy                           0.88      1200
   macro avg       0.88      0.61      0.65      1200
weighted avg       0.88      0.88      0.85      1200

To summarize the model metrics do:

python metrics.py --data "path/to/data/" --model "VGG16.h5" --threshold 0.5 -- output "path/to/metrics.csv"

The arguments are the same as above.

The results look something like this:

To plot the training curves and a confusion matrix, do:

python plot_history_and_confusion_matrix.py --history "path/to/history.csv" --results "path/to/results.csv" --output "figure.png"

The results look like this:

The table below summarizes the results presented in the paper. Results are sorted by AUC.

Train

Validation

Test

Create a dataset either manually or with the provided tools then use the predict script. The data structure is as follows:

pred
    ├───images
        ├───img_00001.png
        ├───img_00002.png
        ├───...
        ├───img_0000X.png

Example:

python predict.py --data "pred/" --model "VGG16.h5" --threshold 0.5 --output "results.csv"

Use predict from naïve candidates and the results from the naive wave breaking detector and a pre-trained neural network to obtain only active wave breaking instances. This script runs on CPU but can be much faster on GPU.

Example:

python predict_from_naive_candidates.py --debug --input "naive_results.csv" --model "path/to/model.h5" --frames "path/to/frames/folder/"  --region-of-interest "region_of_interest.csv" --output "robust_results.csv" --temporary-path "tmp" --frames-to-plot 1000 --threshold 0.5

The output of this script is a comma-separated value (csv) file. It looks like exactly like the output of naive wave breaking detector but adding a extra column with the results of the classification.

Plot the results of the wave breaking detection algorithms. Can handle outputs of any algorithm, as long as the input data is correct. Ideally the results from cluster.py are used as input.

Example:

python plot_wave_breaking_detection_results.py --input "clustered_events.csv" --output "path/to/output/" --frames "path/to/frames/" --region-of-interest "path/to/roi.csv" --frames-to-plot 1000

|

Use interpret.py to apply Grad-CAM to data samples. Organize your data as follows:

gradcam
    ├───images
        ├───img_00001.png
        ├───img_00002.png
        ├───...
        ├───img_0000X.png

Example:

python interpret.py --data "path/to/gradcam" --model "VGG16.h5" -o "path/to/output"

The neural networks developed here can also be used for image segmentation. Please refer to Image Segmentation.

Please refer to Wave Tracking.

Please refer to Wave Breaking Statistics.

La Jument:

Black Sea:

Aqua Alta:

Grad-CAM (Black Sea):

Image Segmentation (La Jument):

The following variables are standard across this repository and scripts that output these quantities should use these names. If a given script has extra output variables, these are documented in each script.

There is no warranty for the program, to the extent permitted by applicable law except when otherwise stated in writing the copyright holders and/or other parties provide the program “as is” without warranty of any kind, either expressed or implied, including, but not limited to, the implied warranties of merchantability and fitness for a particular purpose. the entire risk as to the quality and performance of the program is with you. should the program prove defective, you assume the cost of all necessary servicing, repair or correction.