The source code presented in this repository leverages the keras.io deep learning library (combined with scikit-learn, and other machine learning libraries) to test the usability of the DenseNet neural architecture (Huang et al. 2017), and also the EfficientNet model (Tan et al. 2019), to classify images regarding the presence of a flood, and the severity of that same flood.
@inproceedings{huang2017densenet,
author = {G. Huang and Z. Liu and L. v. d. Maaten and K. Q. Weinberger},
booktitle = {Proceedings of the {IEEE} Conference on Computer Vision and Pattern Recognition},
title = {Densely Connected Convolutional Networks},
year = {2017}
}
@article{tan2019efficientnet,
title = {{EfficientNet}: Rethinking Model Scaling for Convolutional Neural Networks},
author = {Mingxing Tan and Quoc V. Le},
year = {2019},
journal = {arXiv preprint 1905.11946}
}
The dataset files are too big to put on a github repository, so it is necessary to download them and place them in the following hierarchy:
project
│ README.md
│ main.py
│ ...
└─── datasets
│ └─── EuropeanFlood2013
│ │ │ classification.csv
│ │ │ metadata.json
│ │ └─── imgs_small
│ │ │ 25441112.jpg
│ │ │ 25441113.jpg
│ │ │ ...
│ │
│ └─── FloodSeverity
│ │ │ dataset_european_flood_2013.csv
│ │ │ dataset_test_mediaeval2017.csv
│ │ │ dataset_test_mediaeval2018.csv
│ │ │ dataset_train_mediaeval2017.csv
│ │ │ dataset_train_mediaeval2018.csv
│ │
│ └─── FloodHeight
│ │ │ flood_height.csv
│ │
│ └─── MediaEval2017
│ │ └─── Classification
│ │ └─── development_set
│ │ │ │ devset_images_gt.csv
│ │ │ │ devset_images_metadata.json
│ │ │ └─── devset_images
│ │ │ │ 224249.jpg
│ │ │ │ 228743.jpg
│ │ │ │ ...
│ │ │
│ │ └─── test_set
│ │ │ testset_images_gt.csv
│ │ │ testset_images_metadata.json
│ │ └─── testset_images
│ │ │ 182740.jpg
│ │ │ 363897.jpg
│ │ │ ...
│ │
│ └─── MediaEval2018
│ │ └─── Classification
│ │ └─── development_set
│ │ │ │ devset_images_gt.csv
│ │ │ └─── devset_images
│ │ │ │ 900095331450273792.jpg
│ │ │ │ 900160936945831936.jpg
│ │ │ │ ...
│ │ │
│ │ └─── test_set
│ │ │ testset_images_gt.csv
│ │ └─── testset_images
│ │ │ 897989160002256901.jpg
│ │ │ 898024131051921408.jpg
│ │ │ ...
The "European Floods 2013" dataset can be obtained from:
https://github.com/cvjena/eu-flood-dataset
The dataset from "Multimedia Satellite Task from MediaEval 2017" can be obtained from:
https://github.com/multimediaeval/2017-Multimedia-Satellite-Task/wiki/Data-Release
The dataset from "Multimedia Satellite Task from MediaEval 2018" can be obtained from:
https://github.com/jorgemspereira/MediaEval2018-Image-Downloader
The Flood Severity labels can be found here:
https://github.com/jorgemspereira/Flood-Image-Tagger/tree/master/results
The Flood Heights labels obtained through an heuristic procedure (as well as the implementation for the procedure) can be obtained here: https://github.com/jorgemspereira/Flood-Severity-Estimation
The code was developed and tested in Python 3.6.7 with Keras 2.2.4, using Tensorflow as backend. The code supports re-training in cross-validation, and train-test split (being that the test split will be the test split from the MediaEval 2017 DIRSM task). Also, it is possible to re-evaluate a previous saved model. To run the script simply execute:
$ python3 main.py --mode {train, load} --model {dense_net, attention_guided, efficent_net} --dataset {mediaeval2017, mediaeval2018, european_floods, all, flood_severity_3_classes, flood_severity_european_floods, flood_heigths} --method {cross_validation, train_test_split} --data-augmentation --class-activation-map --print-classificationsWhere the flag data-augmentation will use data-augmentation to train the models; the flag class-activation-map will draw the class activation maps for each test example; and print-classifications will print the classification for each test example and save them in a file on the root directory.