Ships Detection In Satellite Imagery

Oil & Gas

Background

The goal was to detect multiple ships on satellite images with pixel-level precision.

Our work

We designed and tested a model based on a state-of-the-art neural network encoder-decoder architecture to detect ships on images and segment them to distinguish individual ships. To handle the problem of an unbalanced dataset, we designed a data enrichment pipeline and a custom sampling strategy. To train our model, we used a distributed deep learning framework deployed on an Amazon Web Services GPU compute instance.

The outcome

Technology used
  • Python
  • Convolutional Neural Networks
  • TensorFlow
  • Keras
  • Horovod