【摘要】 内孤立波(internal solitary waves, ISWs)的传播速度是表征其能量大小的基本参数之一，而利用光学遥感影像反演ISWs传播速度更是关键技术。利用光学仿真实验和光学遥感观测的方法建立了两个ISW数据集，包括1 581个实验室物理仿真(laboratory physics simulation, LPS)数据样本和568个遥感数据样本。为了有效地利用LPS数据，本文开发了一个迁移学习模型来反演安达曼海ISWs传播速度，模型基本结构为残差神经网络(residual neural network, Res-net)。迁移过程如下：首先使用LPS数据训练模型，然后利用遥感数据对预训练模型进行微调，直至模型拟合。速度反演模型表现出良好的准确性，在测试集上的均方根误差(RMSE)(平均相对误差(MRE))为0.19 m/s(7.7%),在浅水和深水两种模式下的RMSE(MRE)分别为0.18 m/s(8.1%)和0.20 m/s(7.2%),可以很好地适应深水和浅水，具有普适性。将此模型应用于安达曼海域的单景遥感图像，根据601个ISWs样本的反演结果，模型预测的ISWs传播速度和水深的关系具有较强的相关性，并且随阴历日呈现双峰分布，在大潮期达到峰值。更多还原

【Abstract】 The propagation speed of internal solitary waves(ISWs) is one of the fundamental parameters that characterizes their energy, and the inversion of ISWs propagation speed using optical remote sensing images is a key technique. Two datasets of ISWs were established using optical simulation experiments and optical remote sensing observations, including 1 581 samples from laboratory physics simulations(LPS) and 568 samples from remote sensing data. To effectively use the LPS data, we developed a transfer learning model to invert the propagation speed of ISWs in the Andaman Sea, with a basic structure of a residual neural network(Res-net). The transfer process involves training the model with the LPS data first and then fine-tuning it with the remote sensing data. The speed inversion model demonstrates good accuracy with an RMSE(MRE) of 0.19 m/s(7.7%) on the test set, with RMSE(MRE) of 0.18 m/s(8.1%) and 0.20 m/s(7.2%) in shallow water and deep water modes respectively, indicating its versatility in both shallow and deep waters. Applying this model to single-scene remote sensing images in the Andaman Sea, the inversion results of 601 ISW samples show a strong correlation between the predicted ISW propagation speed and water depth, with a bimodal distribution that reaches its peak during spring tides.更多还原