Classification of Synthetic Aperture Radar Images of Icebergs and Ships Using Random Forests Outperforms Convolutional Neural Networks

Synthetic aperture radar (SAR) is a common technique for capturing vessels and icebergs on the ocean surface. Convolutional Neural Networks (CNNs) are a popular approach to interpret classes captured in images which include ships and icebergs. However, CNNs are difficult to explain and are computationally expensive. In this paper, we built a random forest (RF) model which outperforms CNN based approaches by 7% and 11% on the testing and validation data, respectively. The RF model used interpretable metrics. These powerful metrics provide insight to what is important to distinguish the two classes from one another. Thus, despite noise present in the SAR images, the RF model was able to provide meaningful classifications between ships and icebergs.