Processing satellite images

Installation

• npm install for the labeling tool
• Needed dependencies for the neural network: Tensorflow (please see the install instructions on their page), Numpy (install with pip), OpenCV2 (install with brew/pip)

Start

• Clone the repository, cd satellite-images
• Put the images to be labeled into "app/images_input"
• In main directory: node server.js starts the labeling tool
• In your browser go to http://localhost:8080/
• Mark the points / label things
• Then in main directory: python crop_images.py path/to/your/img.png will crop images into "images_cropped"
• Now python convolutional_sat.py train will train the neural network and save it under "conv_mnist_model.ckpt", can take some time
• python convolutional_sat.py path/to/your/img.png uses the saved network to run on an image and label it, produces an "output.jpeg"

I tried:

• Didn't work with 40x40 pixel images (feeded to the network)
• 60x60 pixels, more augmentation with rotations: not bad, but it recognices too much (where there is nothing)

Todo

• Switch to hdfs. Does tensorflow support it?

Need to read some papers about this subject.