Deep Learning Based Sea Ice Classification with Gaofen-3 Fully Polarimetric SAR Data

In this paper, the performance of C-band synthetic aperture radar (SAR) Gaofen-3 (GF-3) quad-polarization Stripmap (QPS) data is assessed for classifying late spring and summer sea ice types. The investigation based on 18 scenes GF-3 QPS acquired in Arctic Ocean 2017. study, floe (FI), brash (BI) between floes open water (OW, ice-free area) were classified a mini residual convolutional network, which we call MSI-ResNet. While investigating optimal patch size MSI-ResNet, found that, as continues to grow, classification accuracy first increases then decreases. A 31 × was achieve best performance. using different polarization combinations from also assessed. vertical-vertical (VV) input overestimates FI category while incorrectly identifying most BI FI. VH produces synchronous improvement FI, BI, OW discrimination, with higher overall kappa coefficient (91.09% 0.85, respectively) than VV (83.37% 0.70, respectively). combination vertical-horizontal (VH) polarizations presents modest precision together slight overestimation With VV, VH, horizontal-horizontal (HH) inputs, user’s improves 95.12%, 93.42%, 95.17% OW, respectively. against visual interpretation classes images stratified sampling method. application MSI-ResNet method covering Beaufort Sea north Severnaya Zemlya archipelago high (kappa) 94.62% (±0.92) 94.23% (±0.90), This similar obtained Fram Strait. From results it shown that performs better classical support vector machine (SVM) classifier discrimination. mode show more details discriminating scattered coincident Sentinel-1A Extra Wide (EW) swath data.