In which we try to distinguish icebergs from ships.

Kaggle

Will be updated as a running log of things that have been tried.

Data is provided in the following format:

{
  id: string,
  band1: number[],
  band2: number[],
  inc_angle: number | "na",
  is_iceberg: boolean
}

where any number can be a float.

First attempt is just to use linear SVM on something super simple like the values of the first band. It turns out to be somewhat of a disaster

             precision    recall  f1-score   support

          0       0.50      1.00      0.67        81
          1       0.00      0.00      0.00        80

avg / total       0.25      0.50      0.34       161

[[81  0]
 [80  0]]

It assigns everything to one category. Clear that SVM is not a good approach to use here.

             precision    recall  f1-score   support

          0       0.55      0.42      0.48        81
          1       0.53      0.65      0.58        80

avg / total       0.54      0.53      0.53       161

[[34 47]
 [28 52]]

Slightly "better" approach in the sense that it's not just blindly outputting the same value over and over again.

First idea is to just throw everything

Will be updated as more research is done.

• Combining polarimetric channels for better ship detection results
  • https://earth.esa.int/c/document_library/get_file?folderId=409229&name=DLFE-5566.pdf
  • HH works better at high incidence angles (band_1) so may need to do something there
• Ship-Iceberg Discrimination with Convolutional Neural Networks in High Resolution SAR Images
  • http://elib.dlr.de/99079/2/2016_BENTES_Frost_Velotto_Tings_EUSAR_FP.pdf