Coupling theK-nearest neighbors and locally weighted linear regression with ensemble Kalman filter for data-driven data assimilation

Manhong Fan,Lin Ding,Lihong Tang,Yulong Bai,Lili Wang

doi:10.1515/geo-2020-0312

Manhong Fan, Lin Ding + Show 3 more

Open Access

https://doi.org/10.1515/geo-2020-0312

Copy DOI

Abstract

AbstractMachine learning-based data-driven methods are increasingly being used to extract structures and essences from the ever-increasing pool of geoscience-related big data, which are often used in relation to the atmosphere, oceans, and land surfaces. This study focuses on applying a data-driven forecast model to the classical ensemble Kalman filter process to reconstruct, analyze, and elucidate the model. In this study, a nonparametric sampler from a catalog of historical datasets, namely, a nearest neighbor or analog sampler, is given by numerical simulations. Based on this catalog (sampler), the dynamics physics model is reconstructed using theK-nearest neighbors algorithm. The optimal values of the surrogate model are found, and the forecast step is performed using locally weighted linear regression. Several numerical experiments carried out using the Lorenz-63 and Lorenz-96 models demonstrate that the proposed approach performs as good as the ensemble Kalman filter for larger catalog sizes. This approach is restricted to the ensemble Kalman filter form. However, the basic strategy is not restricted to any particular version of the Kalman filter. It is found that this combined approach can outperform the generally used sequential data assimilation approach when the size of the catalog is substantially large.

Highlights

Data assimilation (DA) is a fundamental approach for combining numerical simulations and observational and experimental big data [1]
The performance of the data-driven data assimilation (DD-DA) approach is examined in comparison to that of ensemble Kalman filter (EnKF)
The results demonstrate that (1) the two approaches are outperformed when the Obs is smaller (= 0.01), and it is difficult to determine which is better; (2) the root mean square error (RMSE) of the DD-DA approach decreases when the size of the catalog is larger; and (3) RMSEDD-DA is clearly lower than RMSEEnKF when

Summary

Introduction

Data assimilation (DA) is a fundamental approach for combining numerical simulations and observational and experimental big data [1]. Sequential DA is called filtering algorithms [2,3]; it includes the forecast and analysis steps [4,5]. Some examples of such filtering algorithms are the Kalman filter (KF) [6], ensemble KF [7,8], particle filter (PF) [9], robust filter [10], and evolutionary computation [11]. Data from remote sensing observations, archives of in situ measurements, and numerical simulations provide novel ideas for. Noteworthy difficulties remain, such as the complexity of physical models, the high dimensionality of states, and the uncertainty of chaotic behaviors

Methods

Results

Discussion

Conclusion