Home

Awesome

Deep Learning for the classification of Sentinel-2 image time series

Training temporal Convolution Neural Netoworks (TempCNNs), Recurrent Neural Networks (RNNs) and Random Forests (RFs) on satelitte image time series. This code is supporting a paper submitted to IEEE International Geoscience and Remote Sensing Symposium (IGARSS) 2019:

@inproceedings{Pelletier2018Deep,
    Title = {Deep Learning for the classification of Sentinel-2 image time series},
    Author = {Pelletier, Charlotte and Webb, Geoffrey I and Petitjean, Francois},
    Booktitle = {IEEE International Geoscience and Remote Sensing Symposium (IGARSS) 2019},
    note = {Accepted for an oral presentation}
}

Examples

Running the models

It will output a result file including the OA computed on test data, the confusion matrix, the training history for deep learning models, and the learned model.

Each model will be trained on train_dataset.csv file and test on test_dataset.csv file.
Please note that both train_dataset.csv and test_dataset.csv files are a subsample of the data used in the paper: original data cannot be distributed.

Thoses files have an header, and contain one observation per row having the following format: [class,objectID,date1.B2,date1.B3,date1.B4,date1.B5,date1.B6,date1.B7,date1.B8,date1.B8A,date1.B11,date1.B12,...,date73.B12]

Maps

The produced map for TempCNNs, bidirectional GRU-RNNs, and RFs are available in the map folder. You can also have a look to our full map of Victoria (Australia) here.

Contributors