Purser A, Bergmann M, Lundälv T, Ontrup J, Nattkemper TW
Use of machine-learning algorithms for the automated detection of cold-water coral habitats: a pilot study
MEPS 397:241-251 | Full text in pdf format

 

Video sequence of images resulting from the auto analysis processing pipeline, determining seabed cover by coral (Lophelia pertusa) or sponges (Geodia baretti, Mycale lingua). This video sequence is composed of all frames analysed within the paper. Given that one frame every 2 seconds was extracted for analysis, this video runs ∼50x faster than the ROV actually flew over the seabed. (Upper left quarter) Original ROV video footage. (Upper right quarter) The cluster colours the neural network has learned from the texture data. Note that this video stream is generated without any expert labels (unsupervised training process only). (Lower left quarter) Expert labels as applied to the neural network. Depending on the number of coral or sponge labels within the clusters on the map, regions of the image are coloured red or green, respectively. (Lower right quarter) The final categorised frame from which the coverage estimations are produced by counting the coloured pixels.

 

Click to enlarge.

 

Download video: m397p241_app.avi (.avi, 110MB)

 

Windows Media Player or Quicktime with the Flip4Mac plug-in are required to view the video.