Sea ice concentration estimation from satellite SAR imagery using convolutional neural network and stochastic fully connected conditional random field

Ice mapping is important for numerous applications such as ship navigation and mining in the Arctic regions. The need for ice mapping is increasing to support operations in the Arctic regions. We propose an automatic SAR ice concentration estimation method using convolutional neu-ral network (CNN) followed by a stochastic fully connected random field (SFCRF). CNN can generate the proper fea-tures for robust ice concentration estimation. The low boundary accuracy of CNN is overcome by introducing SAR image structure information into the CNN output using SFCRF. This method uses ice charts as training data to take advantage of the abundant ice chart archives, which makes it appealing for operational ice mapping purposes. It shows robustness to melting conditions on our test dataset. The overall precision of the ice concentration estimate is less than 0.12 (12%) when validated on ice chart. This method can be naturally extended for ice water classification. An overall classification precision of 96 % is achieved by using a simple thresholding on the ice concentration estimates. 1