qberto / neural-road-inspector

After a hurricane, roads are often flooded or washed out, making them treacherous for motorists. Using state of the art deep learning methods, I attempted to automatically annotate flooded, washed out, or otherwise severely damaged roads. My goal is create a tool that can help detect and visualize anomalous roads in a simple user interface.

Motivation

Devastating hurricanes are occurring with higher frequency due to global climate change. After a hurricane, roads are often flooded or washed out, making them treacherous for motorists.

Using state of the art deep learning methods, I attempted to automatically annotate flooded, washed out, or otherwise severely damaged roads. My goal is create a tool that can help detect and visualize anomalous roads in a simple user interface.

For details, please read Blog

Getting Started

Prerequisites

Code written for Python 2.7. Minor modification may be necessary for use with Python 3.

To install required Python libs, in your shell, run:

pip install -r requirements.txt

Training Data

My model is trained purely with retina (512x512) satellite imagery and corresponding street map images from MapBox API but this code will accept training images of other sizes or other sources. Because I do not own the training images, no training data is copied here. Consider signing up for a MapBox account to get an access token. Please observe the licensing terms of any image you use for training.

I chose to train images from zoom level 16 (Open Street Map Slippy Map Format). It is recommended to use training images from zoom 16 or above.

Training the Road Segmentation Neural Network Model

Once you have training data, you are ready to train the model.

For an in-depth technical discussion of the model, please see blog post.

Command to train the model:

python RoadSegmentor.py [config file]

You must provide a config file containing file paths to the training datasets along with neural network training hyper-parameters. See cfg/default.cfg for an example.

You must also supply a plaintext file listing satellite image relative filepath under column label 'img'. See data/tile_log.csv for an example.

With a Nvidia Quadro P5000 GPU (16MB GPU RAM), training with a batch size of 16 took about 5 hours on a dataset with about 9000 images.

Generating Road Segmentation Tiles

For road change detection, you will need to generate 2 sets of road segmentations: pre-hurricane and post-hurricane.

Post Hurricane satellite imagery (in GeoTiff format) can be downloaded via the DigitalGlobe Open Data Program. For convenience, you can use data/tiffDownloader.py to download a set of geotiff files. The script takes in a text file containing a list of http urls on separate lines.

In a data pre-processing step, GDAL2Tiles can be used to cut the GeoTiff files into standard OSM square tiles (use matching zoom level as the training set).

Command line tool:

python roadSegmentationMaskGen.py [satellite_images_dir] [output_dir] [keras_model_filepath]

Note: If you use my pre-trained Keras/Tensorflow model, input images must be sized 512x512.

Generating Annotated Map Tiles

After both the pre-hurricane segmentation tiles and post-hurricane segemntation tiles have been generated, the next step is generating the difference.

This project comes with a command line tool that generates a pyramid hierarchy of annotated map tiles following the same hierarchy structure of the input directories.

python postprocessing/anomalyMapGen.py [pre_event_segment_dir] [post_event_segment_dir] [street_map_dir] [output_dir]

Demo

Demo

License

Code: MIT License

Road Segmentation Weight File: CC-BY-SA

Map Artifacts: CC-BY-SA

Disclaimer

This project should be used for research purposes only. Further testing and improvement is needed for use in the field.