基于海洋机器人观测的高分辨率海洋环境预测方法

A method for high-resolution prediction of ocean environments based on observations from marine vehicles

海洋机器人是海洋环境观测应用的重要工具。利用海洋机器人获取的稀疏观测数据,实现区域海洋环境要素空间场高分辨率预测具有重要应用需求。面向该需求,提出一种基于海洋机器人稀疏观测数据的高分辨率海洋环境预测方法。以卷积长短期记忆神经网络ConvLSTM为基础,建立数据驱动的海洋环境空间场时间序列预测模型。应用集合卡尔曼滤波方法同化海洋机器人观测数据,实现神经网络预测模型在线学习,使模型逐步逼近海洋环境状态、预测输出逐步逼近观测数据。针对大尺度海洋环境高空间分辨率预测带来的高计算复杂性问题,引入非均匀空间分辨率预测策略,有效应对计算挑战。利用水下滑翔机南海北部观测实验数据进行了区域海洋环境二维水平面温度场预测实验,结果验证了方法的有效性。

Marine vehicles serve as pivotal platforms for oceanic environmental observations.Given the sparse nature of the observation data collected by marine vehicle networks,a significant demand arises for generating high-resolution predictions regarding spatial distributions of regional oceanic parameters.Addressing this demand,this paper proposes a novel methodology that utilizes sparse marine vehicle observations for high-resolution ocean environment prediction.Built upon the ConvLSTM neural network model,this approach establishes a data-driven model capable of predicting the spatiotemporal evolution of oceanic fields.Based on the neural network model,the ensemble Kalman filter technique is employed to effectively assimilate marine vehicle observations,enabling online learning of the neural network model that progressively aligns model and its predictions with evolving oceanic conditions and observed data.To address the computational challenges posed by high-resolution predictions over large oceanic scales,a non-uniform spatial resolution prediction strategy is introduced,effectively reducing computational burden.Experimental validation employs observations from an underwater glider network deployed in the northern South China Sea,predicting two-dimensional horizontal temperature spatial field within regional ocean environment.The results validate the effectiveness of the proposed method.