TerraSAR-X Ice/Non-Ice Mapping of the Mackenzie River Using a Convolutional Neural Network

Kurzfassung

Mackenzie Rive located in Yukon and Northwest Territories is the longest river system in Canada and an important transportation link. During the ice free season it is used for shipping transportation, while during freeze up time it serves as an ice road for trucks. Therefore, knowledge of ice conditions is essential to enable save navigation. Due to the difficulty of in-situ observation, remote sensing offers an effective instrument for river ice measurements. SAR data thereby enables gap free time series as it is an active sensor independent from sun illumination and cloud conditions. However, separating ice and open water in SAR data remains a challenging task due to the principals of the radar signal. On the one hand it is sensitive to surface roughness influencing weather conditions such as wind and rain affecting the appearance of water. On the other hand, ice shows different backscatter characteristics during freeze up and melting period. Therefore threshold or cluster based approaches encounter their limits in separating the features spaces of both classes demanding a more complex approach to cover these diverse patterns in the SAR image. In this study a Convolutional Neural Network (CNN) and TerraSAR-X data is used to map open water and ice in a time series between January 2014 and December 2015. Convolutional Neural Networks have proven great potential in classification of satellite images. Not focusing on isolated pixel values but regarding larger receptive fields they are also taking into account texture and shape information. Thus they are suitable to define feature spaces for ice and water. The test site covers the Mackenzie River Delta at the estuary of the Arctic Ocean. A U-Net with 18 convolutional layers and skip connections was used for training and classification. The four Kennaugh elements (K0 total intensity, K3 double/single bounce, K4 polarization, K7 torsion) were calculated from a total of 47 dual polarized HH/VV TerraSAR-X scenes. 42 of the 47 scenes show complete ice free respectively freeze up conditions and were used to train and validate the network with a total of 4000 patches (3000 for training and 1000 for validation). The remaining 5 scenes show transition conditions during the melt or freeze up time in spring or autumn and were used to test the neural network after training phase. The capability of the network to recognize class specific features in the SAR data is illustrated by visualizing the trained filters of the network showing typical SAR image structures of ice and water. Limitations of the approach occur for hybrid forms of water and ice that show features of both classes.The model was driven to best performance by adjusting hyper parameters including patch size, batch size, number of epochs, number of filters and filter size after each training run. Obtained results for the test images prove that a combination of a CNN and TerraSAR-X data is able to reliably separate open water from ice.