Training temporal Convolution Neural Netoworks (CNNs) on satelitte image time series.
This code is supporting a paper submitted to IEEE Transactions on Geoscience and Remote Sensing (under review):
https://arxiv.org/abs/1811.10166 - Temporal Convolutional Neural Network for the Classification of Satellite Image Time Series
More information about our research at https://sites.google.com/site/charpelletier, http://www.francois-petitjean.com/Research/, and http://i.giwebb.com/
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 will run by training the network on train_dataset.csv file and by testing it 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 no header, and contain one observation per row having the following format:
[class,date1.NIR,date1.R,date1.G,date2.NIR,...,date149.G]
- Number of channels in the data:
n_channels = 3(run_archi.py, L21).
It will require to change functions contained inreadingsits.py. - Validation rate:
val_rate = 0.05(run_archi.py, L22). - Network hyperparameters are mainly defined in
architecture_features.pyfile.