Training temporal Convolution Neural Netoworks (CNNs) on satelitte image time series.
This code is supporting by a paper published in Remote Sensing:

@article{Pelletier2019Temporal,
  title={Temporal convolutional neural network for the classification of satellite image time series},
  author={Pelletier, Charlotte and Webb, Geoffrey I and Petitjean, Fran{\c{c}}ois},
  journal={Remote Sensing},
  volume={11},
  number={5},
  pages={523},
  year={2019},
  publisher={Multidisciplinary Digital Publishing Institute},
  note={https://www.mdpi.com/2072-4292/11/5/523}
}

This code relies on Pyhton 3.6 (and should work on Python 2.7) and Keras with Tensorflow backend.

• main architecture: python run_main_archi.py
• other experiments described in the related paper:

python run_archi.py --sits_path ./ --res_path path/to/results --noarchi 0

The architecture (run_main_archi.py) will run by training the network on example/train_dataset.csv file and by testing it on example/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 no header, and contain one observation per row having the following format: [class,polygonID,date1.NIR,date1.R,date1.G,date2.NIR,...,date149.G], where class corresponds to the class label and polygonID to a unique polygon identifier for each plot of land.

• Number of channels in the data: n_channels = 3 (run_archi.py, L21).
  It will require to change functions contained in readingsits.py.
• Validation rate: val_rate = 0.05 (run_archi.py, L22).
• Network hyperparameters are mainly defined in architecture_features.py file.

python write_output.py --model_path path/to/model --test_file path/to/pred.csv --result_file path/to/results/result.csv --proba

test_file is either a csv file or a tiff image. If the test_file is a tiff file and --proba activated, two tiff images are created: 1) a land cover map, and 2) a tiff image composed of n_classes bands that contains the proabbility outputed by the Softmax layer for each class. The code has been designed to work on small tiff file. Predictions on a big tiff file would require to set up carefully size_areaX and size_areaY variables (L86-87 in write_output.py).

Please note that the pred.csv file should have the same format than example/train_dataset.csv, including the class field that could be set to -1.

The produced map for TempCNNs and RFs are available in the map folder.

• Dr. Charlotte Pelletier
• Professor Geoffrey I. Webb
• Dr. Francois Petitjean