A Novel Data Augmentation Method for Sea Ice Scene Classification of Arctic Aerial Images

Abstract

In this paper, a novel data augmentation method was proposed for supervised sea ice scene classification with arctic aerial images. Ice-type classification of sea ice scenes in a region is useful for instant navigation. However, some types of sea ice scenes are difficult to collect. The small number of available samples usually limits the performance in classification. Inspiring by the transfer learning method, the training samples are augmented with simulated sea ice images. Considering the distribution characteristic of sea ice scene, simulation samples are synthesized by algebraic operations on the respective regions of interest from two true samples. One of the two true samples can be an additional image, which is more easily collected than others, such as full ice and snow. Simultaneously, it introduces new information. Generalization and error-correction capability of deep neural networks for training samples make the proposed method feasible. The experiments on true sample sets, simulation sample sets, and mixed sample sets were implemented. Finally, the effectiveness of our data augmentation method was demonstrated, which improved the accuracy of sea ice classification.

Keywords

• Data augmentation
• scene classification
• sea ice
• transfer learning