4:00 PM - 4:15 PM
[ATT30-03] Acquisition of wind wave characteristics from X-band Maritime Radar Images via Artificial Neural Networks
Keywords:Convolutional Neural Networks, Wind Waves, X-band Radar Imagery, Significant Wave Height, Machine learning
Maritime radar systems play an essential role in ensuring the safety of navigation on the seas by identifying other ships and potential hazards. Typically, sea clutter, primarily resulting from Bragg scattering, is considered interference and is filtered out. However, this phenomenon becomes observable in raw radar imagery, such as that obtained through the SeaVision hardware suite, particularly when the wind speed and wave height surpass specific thresholds. Utilizing these unfiltered images, one can discern the properties of wind-driven sea waves. Research tracing back to 1964 has substantiated the feasibility of these detections. Traditional spectral methodologies used to extract wave characteristics, despite their applications, encounter challenges in refining precision. Conversely, deep learning approaches, particularly adept at image analysis tasks, provide a more robust framework capable of managing noisier data, and they do so without the necessity of Fourier transformations or reliance on extensive radar image sequences.
In our study, we introduce a technique that leverages convolutional neural networks (CNNs) to estimate the characteristics of wind waves from radar images captured onboard vessels using the SeaVision system. Specifically, our CNN is calibrated to deduce the significant wave height, correlating it with the data from the Spotter buoy, which serves as a reference measure. Given that obtaining measurements from the Spotter buoy is an expensive and time-intensive endeavor, our approach also includes the preliminary training of the CNN, enhancing its generalization capabilities and the overall integrity of wind wave characteristic acquisition. This CNN-based methodology presents a marked improvement over conventional techniques, as it necessitates a minimal amount of radar image data—requiring a single snapshot from the SeaVision system — compared to the classical approach's need for over 15-20 minutes of radar imagery.
In our study, we introduce a technique that leverages convolutional neural networks (CNNs) to estimate the characteristics of wind waves from radar images captured onboard vessels using the SeaVision system. Specifically, our CNN is calibrated to deduce the significant wave height, correlating it with the data from the Spotter buoy, which serves as a reference measure. Given that obtaining measurements from the Spotter buoy is an expensive and time-intensive endeavor, our approach also includes the preliminary training of the CNN, enhancing its generalization capabilities and the overall integrity of wind wave characteristic acquisition. This CNN-based methodology presents a marked improvement over conventional techniques, as it necessitates a minimal amount of radar image data—requiring a single snapshot from the SeaVision system — compared to the classical approach's need for over 15-20 minutes of radar imagery.