To decrease costs of maintaining safe working conditions on maritime operations, Statoil and C- CORE published a Kaggle challenge to developing a high-performance classification model that automatically identifies ships and icebergs from each other in the radar imagery. In total the data includes 1604 labeled 75x75 dual channel images captured by Sentinel-I - remote sensing system and its Synthetic Aperture Radar (SAR). The imagery includes both HH and HV po- larization channels which makes it possible to more efficiently recognize objects with different bounceback polarization distributions. The provided data includes plenty of background noise as expected from a remote sensing system. Thus, several pre-processing methods and convolution deep learning models were compared with different hyperparameters. As the classification problem requires the model to detect small details from the image, small kernel size is required. Our results showed that deep models achieve better results with small kernel size. In addition, we found that with a small batch size, overfitting can be decreased. With data augmentation, we were able to increase labeled data set from 1600 to 1800. he validation loss decreased from 0.2614160 to 0.20011 with data augmentation.
More in depth discussion about our project can be found from our report!