A Convolutional Neural Network Architecture for Sentinel-1 and AMSR2 Data Fusion

With a growing number of different satellite sensors, data fusion offers great potential in many applications. In this work, convolutional neural network (CNN) architecture is presented for fusing Sentinel-1 synthetic aperture radar (SAR) imagery and the Advanced Microwave Scanning Radiometer 2 (AMSR2) data. The CNN applied to prediction Arctic sea ice marine navigation as input forecast models. This generic model specifically well suited sources where ground resolutions sensors differ with orders magnitude, here 35 km × 62 (for AMSR2, 6.9 GHz) compared 93 m 87 sentinel-1 IW mode). two optimization approaches are using categorical cross-entropy error function specific application training on charts. first approach, concentrations thresholded be encoded standard binary fashion, second used target probability directly. method leads significant improvement R 2 measured evaluated over test set. performance improves both terms robustness noise alignment mean from analysts validation data, an R2 value 0.89 achieved independent It can concluded that CNNs suitable multisensor even by large factors, such case SAR AMSR2.