Gaofen-3 sea ice detection based on deep learning

Abstract

Sea ice detection is one of the most important applications in synthetic aperture radar (SAR) image processing and it always be served for ship navigation and climate change studies. Due to the noise, low resolution and multiple characteristics of SAR images, and limitations of traditional SAR image classification methods, the detection accuracy of traditional sea ice SAR images is not high and can't meet the high precision requirements of application in practice. Deep learning, one of the most popular machine learning methods, have been researched and applied in SAR image classification by some researchers. But, so far, very few researchers have applied deep learning methods to sea ice SAR image detection. In this paper, two scenes of Chinese Gaofen-3 SAR data of sea ice were applied to sea ice detection research using convolutional neural network (CNN). First, the train data is obtained from Gaofen-3 sea ice SAR Image by chipping different classes of SAR images into patches; Then, the different classes train data is used to train the CNN and get the trained CNN model; After the training, the trained CNN model was used to classify the Gaofen-3 sea ice SAR image with sea ice and non-sea ice by a patch-based window traversing the entire SAR image. In order to illustrate the sea ice detection result of deep learning, several traditional SAR image detection methods are experimented for contrast. Experimental results demonstrate that deep learning method is suitable for sea ice SAR image detection and achieves high detection accuracy than others traditional methods.