Dry Dock Detection in Satellite Images with Representation Learning

Aktas U. R., Firat O., YARMAN VURAL F. T.

21st Signal Processing and Communications Applications Conference (SIU), CYPRUS, 24 - 26 April 2013 identifier

  • Publication Type: Conference Paper / Full Text
  • Country: CYPRUS
  • Middle East Technical University Affiliated: Yes


In this study, we propose a method to detect dry docks, a harbour man-made object which is hard to recognize, using representation learning in satellite images. Dry docks are coastal structures which may include ships for repairing purposes, and they exist in harbour regions. The search space is pruned by making use of two low-level features that invariantly define docks, and remaining samples are used to train a representation learning system. Experimental results suggest that classification methods using learned features have similar performances to those using handcrafted features, which are proposed by the field expert. The results also provide insight on the applicability of the same methodology on detection of different objects in remotely sensed images, without wasting any effort.