ship detection & localisation

Ship detection and localisation from satellite images.

Blog posts:

Ship detection - Part 1: Ship detection, i.e. binary prediction of whether there is at least 1 ship, or not. Part 1 is a simple solution showing great results in a few lines of code
Ship detection - Part 2: ship detection with transfer learning and decision interpretability through GAP/GMP's implicit localisation properties
Ship localisation / image segmentation - Part 3: identify where ship are within the image, and highlight pixel by pixel

Part 3 highlight: Image segmentation with a U-Net

Part 2 highlight: Class activation mapping on vessel detection classifier ConvNet - convnet learned where ships are without supervision!

Steps taken:

find data sets:
- planet API:
  - needs subscription, but there is a free trial
  - Example usage here: https://medium.com/dataseries/satellite-imagery-analysis-with-python-a06eea5465ea
- airbus kaggle set (selected for first iteration)
  - https://www.kaggle.com/c/airbus-ship-detection/data
  - to download locally, make sure your connection is stable (29 GB)
    - get API key through your kaggle profile (free), and either save file or enter name and key as environmental variable
    - nohup kaggle competitions download -c airbus-ship-detection & disown %1
- other data providers: Airbus, Digital Globe
- free sources: includes EOS's Sentinel 1 (SAR - active/radar) and 2 (optical) with coverage period ranging of 2-7 days https://eos.com/blog/7-top-free-satellite-imagery-sources-in-2019/
simple EDA on data used in this repo and blog post:
- 200k+ images of size 768 x 768 x 3
- 78% of images have no vessel
- some images have up to 15 vessels
- Ships within and across images differ in size, and are located in open sea, at docks, marinas, etc.
modelling broken down into two steps, with subfolders in the repo
- ship detection: binary prediction of whether there is at least 1 ship, or not
- ship localisation / image segmentation: identify where ship are within the image, and classify each pixel as having a ship or no ship (alternative could be to have a bounding box, with a different kind of model)
other articles on the topic:

rugg2/ship_detection