Detect roads and features in satellite imagery by training a convnet with OSM data.

Work in progress.

This has been run on OSX Yosemite (10.10.5).

brew install libjpeg
pip3 install -r requirements.txt 
sudo easy_install --upgrade six

Install globalmaptiles.py

mkdir lib
cd lib
git clone git@gist.github.com:d5bf14750eff1197e8c5.git global_map
cd ..
export PYTHONPATH=$PYTHONPATH:./lib/global_map:./data_pipeline

This will download vectors, imagery, and run the analysis.

python3 label_chunks_softmax.py download-data MAPZEN_KEY
python3 label_chunks_softmax.py train

Detect OpenStreetMap (OSM) ways (streets and trails) in satellite imagery. Train the neural net using MapQuest open imagery, and an OSM ways.

• TensorFlow - using this for the deep learning
• Hinton on using a neural network to do this - best/recent paper on doing this, great success with these methods
• Links from the Tensorflow site
  • MNIST Data and Background
  • all the other links to Nielsen’s book and Colah’s blog
• Deep Background
  • original Information Theory paper by Shannon

Multilayer Convolutional Network

• Download a chunk of satellite imagery from MapQuest at max resolution, in 256x256px PNGs
• Download ways (i.e. road/trails) for that area from OSM
• Generate training and evaluation data

• Download a different set of training imagery and OSM ways, and see if we can predict the ways from the imagery

• flatten it into matrices of trail/no trail
• check tiles visually

• use GDAL to download and composite image tiles

Data Size

• do a tiny area and do it all locally for testing
• use one or more GPUs on Amazon if bottlenecked

Download OSM data, parse out ways

• alternative to Clipper method: load it into Postgres and do it that way

Accuracy

• mimic Hinton’s methods, esp. for getting real road geometries
• see if we can identify trails nearly as well as roads