AbstractThe RADARSAT Constellation Mission (RCM) offers a compact polarimetric (CP) synthetic aperture RADAR (SAR) mode that provides a wider swath than quad-polarization (QP) and more polarization information in observations than dual-polarization (DP).
We investigate the capability of CP SAR imagery in generating sea ice maps by taking advantages of learned features, statistical properties, and contextual information.
We present a region-based sea ice mapping methodology. First, an existing unsupervised segmentation called iterative region growing with semantics based on statistical properties of CP SAR data (CP-IRGS) is used to generate edge-preserved and homogeneous regions to reduce destructive effects of speckle noise.
Then, a residual-based convolutional neural network (ResCNN) is used to specify the type of ice in regions. The performance of the proposed classification methodology is compared to that of standard machine learning classifiers, support vector machine (SVM) and random forest (RF).
To simulate CP SAR data, two QP RADARSAT-2 scenes are utilized.
The obtained results indicate that the proposed region-based classification methodology achieves 96.66\% overall accuracy, which is approximately 4\% higher than those obtained by SVM and RF.