CS461 Final Project: Damage and Accessibility Assessment for Post-Disaster Regions from Satellite Imagery

Repository Organization

cs461-final-project/
  |- README.md
  |- post_event_selection.py --> Select images that correspond with specific code to post-process
  |- retrieve_ground_truth.py --> Parse GeoJSON crowd-sourced damage labels and save associated image tiles
  |- scrape_digitalglobe.py --> Scrape links to Pre-/Post-Event satellite images, download and compress TIF images
  |- split_data.py --> Perform 60-20-20 train,test,validate split on roads training dataset
  |- calculate_per_pixel_change.py --> Binarize image mask and compute Manhattan and Zero norms
  |- data_preprocessing.py --> Apply density filter and image augmentation to training dataset
  |- image_matching.py --> Perform SURF feature matching and perspective transformation for pre and post image matching
  |- prepare_satellite_imgs.py --> Apply 512x512 image tiling, (optionally) add contrast to images, and match pre-/post-images
  |- notebooks/
     |- UNet.ipynb --> not pretrained unet model
     |- damage_ground_truth.ipynb --> Parse GeoJSON crowd-sourced damage labels and view/save associated image tiles
     |- data_preprocessing.ipynb --> preprocessing training data
     |- semantic_segmentation.ipynb --> unet/fpn model, change type by changing the model parameter
     |- segment_satellite.ipynb --> makes prediction masks for pre-/post-event satellite images
     |- per_pixel_overlap.ipynb --> calculates the per pixel overlaps from pre-/post-event mask
     |- best_model_full.pth --> best model of fpn saved over 40 iterations
     |- best_model_unet.pth --> best model of unet saved over 40 iterations
  |- data/
     |- post_event
     	|- ...
     |- pre_event
     	|- ...

Intructions for Running Satellite Image Road Segmentation Example

1. Run through semantic_segmentation.ipynb. The current default is unet. If want to get the result for fpn, leave only model_path = './best_model_full.pth' uncommented.
2. Run through per_pixel_overlap.ipynb. The current default is unet. If want to get result for fpn, uncomment mask_dir = '../predictions_fpn/'

Data Preprocessing

The directory containing the data to process is used as an argument when calling the program. Here is an example command to apply preprocessing to the train images in the roads directory.

python data_preprocessing.py data/roads/train/

Install requirements: pip3 install requirements.txt