Ocean feature classification from SAR Quicklook Imagery using Convolutional Neural Networks

Global ocean wind and wave parameters are important inputs for weather forecasting and climate modeling. Spaceborne SAR sensors are unique resources for extraction of such ocean features due to their high resolution and wide coverage. This study examines the capability of Convolutional Neural Networks (CNNs) for extracting ocean wind features from TerraSAR-X quicklook images (QL). The QL is a freely and easily available data source to train and validate the CNN against "ground truth" ocean parameters from (also freely available) buoy data. We find that despite obvious corruption of SAR backscatter calibration during the QL formation process, the CNN with QL input produce estimates of similar accuracy for the key ocean parameter of wind speed (residual mean absolute error to ground truth: 2.2 m/s) to that of established conventional wind field retrieval methods operating on calibrated backscatter. We attribute this to the CNN exploiting higher order texture information preserved in the QL to measure the wind parameters via spatial ocean features. Other ocean parameters are also reconstructed by the CNN with reasonable accuracy. A quantitative performance comparison of our CNN architecture with higher quality inputs; calibrated backscatter and complex data is underway.