Deep Convolutional Neural Networks for Semantic Segmentation of Multi-Band Satellite Images

Preparation

• Download 3-band and 16-band from here and extract to data folders
• Install requirements by executing:

$ pip install -r requirements.txt

• In addition you need to install either tensorflow or tensorflow-gpu

Training

$ python train.py

Argument	Description	Options
--algorithm	Algorithm to train	unet, fcn_densenet, tiramisu, pspnet
--size	Size of patches	int
--epochs	Epochs to train for	int
--batch	Samples per batch	int
--channels	Image channels	3, 8, 16
--loss	Loss function	crossentropy, jaccard, dice, cejaccard, cedice
--verbose	Print more information	bool
--noaugment	Turn off augmentation	bool
--name	Give run a custom name	str

Testing

$ python train.py --test

Argument	Description	Options
--algorithm	Algorithm to test	unet, fcn_densenet, tiramisu, pspnet
--size	Size of patches	int
--channels	Image channels	3, 8, 16
--verbose	Print more information	bool

Visualization

It's possible to run some visualization of the data by running $ python visualize.py from the utils folder.