A study on application of flow velocity distribution measurement of drift ice of Okhotsk coast ice sea navigation using Radar images

(1)加速漂移冰雷达图像的运动分析为了自动获得漂移冰分布的时间变化，可以对漂移冰雷达图像应用块匹配方法。然而，漂移冰分布的运动是复杂的，通常的块匹配分析存在很多误差。此外，为了确保海上航线的安全，不仅需要了解目前的情况，还需要预测未来的情况。实现理想的自动实时处理是十分必要的。块匹配方法的分析结果可以作为预测处理的输入信息。然而，由于通常的块匹配方法的计算量非常大，因此难以实时处理。然后，为了测量本研究中漂冰分析的加速和高精度，开发了块匹配的改进技术。(2)基于神经网络的亚像元估计方法提出了一种基于神经网络的亚像元估计海冰流速分布的方法。如果需要精确的漂移冰流位移，则需要亚像素估计最佳匹配位置。通过对询问区域内的相关峰进行插值，获得亚像素精度。传统上，高斯函数被用作峰值拟合函数。然而，亚像素精度依赖于偏差误差和随机误差。对于所有应用的图像，很难确定峰值拟合函数。然后，我们提出了利用具有应用图像先验知识的神经网络的方法。从雷达图像上验证了该方法的有效性。

(1)Speed-up of movement analysis of drift ice radar imageTo get a time change in the drift ice distribution automatically the block match method can be applied to the drift ice radar image. However, the movement of the drift ice distribution is complex, and a lot of errors occur in a usual block match analysis. Moreover, it is necessary not only to understand a present situation to secure a safe sea route but also forecast the situation in the future. And, it is necessary to achieve this processing ideally automatically and in real time. An analytical result of the block match method can be used as input information on the forecast processing. However, because the amount of the calculation is very large in a usual block match method, real time processing is difficult. Then, to measure speeding up and making of the drift ice analysis highly accurate in this research, the improvement technique of the block match is developed.(2)Sub-pixel estimation method using neural networksA sub pixel estimation technique using neural networks has been presented for obtaining flow velocity distribution of sea ice. If we need precise displacement of the flow of the drift ice, sub-pixel estimation of best matching location is required. The sub-pixel accuracy is obtained by applying an interpolation scheme to the correlation peak within the interrogation area. Traditionally, the Gauss function has been used as peak-fitting function. However the sub-pixel accuracy is dependent upon both a bias error and a random error. It is very difficult to identify peak fitting function for all applied images. Then, we propose the method that uses neural networks which has a priori knowledge about applied image. We show the proposed technique achieved good results from the radar image.