ComDA: A common software for nonlinear and Non-Gaussian Land Data Assimilation

Highlights

• ComDA is developed to build a common platform for data assimilation applications

• Parallel, multidisciplinary and scalability capabilities ensure ComDA works well

• ComDA promotes studies on the applications of land data assimilation

Abstract

Common software for land data assimilation is urgently needed to implement a wide variety of assimilation applications; however, a fast, easy-to-use, and multidisciplinary application-oriented assimilation platform has not been achieved. Therefore, we developed Common software for Nonlinear and non-Gaussian Land Data Assimilation (ComDA). ComDA integrates multiple algorithms (including diverse Kalman and particle filters), models and observation operators (e.g., common land model (CoLM), Advanced Integral Equation Model (AIEM)), and provides general interfaces for additional operators. Using mixed-language programming and parallel computing technologies (Open Multi-Processing (OpenMP), Message Passing Interface (MPI) and Compute Unified Device Architecture (CUDA)), ComDA can assimilate various land surface variables and remote sensing observations. High-performance computing and synthetic tests and real-world tests indicate that ComDA achieves the standard of common land data assimilation software with parallel computation, multiple operators, and assimilation algorithms and is compatible with many models. ComDA can be applied for multidisciplinary data assimilation.

Introduction

As a new methodology in Earth system science, land data assimilation incorporates both Earth observations and numeric land surface models to further understand and predict land surface processes. Generally, due to the limited awareness of Earth science concepts and the inherent errors in measuring methods, uncertainties arise that have profound impacts on the accuracy of Earth observations (Crow et al., 2012) and modeling (Merz et al., 2009). Data assimilation takes advantage of Earth observations, modeling and their uncertainties and provides a more effective framework for studying land surface processes (Talagrand, 1997; Liang et al., 2013; Li, 2014). Data assimilation consequently places higher demands on computer development environments for specific applications.

To date, data assimilation has become a widely accepted methodology that has been applied in a variety of research fields, including hydrology (Liu and Gupta, 2007), the carbon cycle (Rayner et al., 2005), climatology (Carton and Giese, 2008; Fang and Li, 2016), and phenology (Ines et al., 2013). Its theoretical basis was strengthened by ongoing frontier exploration, such as multiple scale assimilation (Bocquet et al., 2011), nonlinear and non-Gaussian methods (Apte et al., 2007; Han and Li, 2008; van Leeuwen, 2015), stochastic analysis (Miller, 2007; Liu and Li, 2017), etc. Correspondingly, data assimilation system development has been in full swing. Typical systems include the Global Land Data Assimilation System (GLDAS, Rodell et al., 2004), the European Land Data Assimilation System (ELDAS, Jacobs et al., 2008), the Chinese Land Data Assimilation System (CLDAS, Li et al., 2007), the Canadian Land Data Assimilation System (CaLDAS, Balsamo et al., 2007), and the Earth Observation Land Data Assimilation System (EOLDAS, Lewis et al., 2012).

Certain software endeavors have been involved in the general development of platforms for common data assimilation studies, for example, DART (Data Assimilation Research Testbed, Anderson et al., 2009), OpenDA (Open Data Assimilation library) and OpenMI (Model Interface) (Ridler et al., 2014). Common software for land data assimilation should have the following qualifications: first, parallel computation is necessary for rapid research outputs. An assimilation system tends to be time-consuming, especially for high-dimensional and exceedingly complex land surface models and must contend with the massive introduction of forcing data, state variables or observations. Additionally, a parallel framework favors the performance of ensemble-based algorithms, grid data and parallel-designed models. Second, multiple dynamic models for the wide range of applications and multiple observations for assimilating more Earth observations are necessary. Moreover, additional observations introduce corresponding observation operators, such as radiative transfer models when assimilating remote sensing data. Third, various algorithms are needed to reduce the potential computer algorithm errors and extend the application of software. The ensemble Kalman filter (EnKF) and particle filter (PF) are assimilating algorithms that have widespread use; however, they are computationally demanding because they require a great amount of ensembles to approximate the model's track. Additionally, EnKF is based on the multidimensional Gaussian assumption (Fowler and van Leeuwen, 2013), which may restrict its ability to adapt to non-Gaussian methods. Therefore, integrating advanced KF and PF is necessary. Last, sufficient space is necessary for forthcoming extensions that introduce new dynamic models and measurements without the need for substantial programming, which is also in demand for the land data assimilation community.

These qualifications constitute a benchmark for common land data assimilation software developments. In particular, widely used software, such as DART and land information system (LIS) data assimilation (Kumar et al., 2008), consider a large range of land surface models and remote sensing observations (for example, Advanced Microwave Scanning Radiometer-Earth Observing System (AMSR-E) and Moderate Resolution Imaging Spectroradiometer (MODIS)) but require restrictions on customized models and algorithms for general users. Other related works all have shortages as well. OpenDA & OpenMI provides an interface standard to bridge the assimilation system and numerical models but is not appropriate for user friendly and fast applications because of the lack of realistic models. Karssenberg et al. (2010) propose a visualized software framework, which further contains PF. Other software, including PDAF (parallel data assimilation framework, Nerger and Hiller, 2013), Daspy (Han et al., 2015), TerrSysMP-PDAF (Kurtz et al., 2016) and the soil and water assessment tool-hydrological data assimilation system (SWAT-HDAS) (Zhang et al., 2017), provide limited forecasting models (for example, Community Land Model or SWAT), and assimilation algorithms are restricted to Kalman filters and PF (i.e., EnKF and local ensemble transform Kalman filter). Therefore, general software for land data assimilation and its corresponding applications are still urgently needed.

In this paper, we introduce Common software for Nonlinear and non-Gaussian Land Data Assimilation system (ComDA), which we developed to meet the above demands of a general software platform of data assimilation. This paper is organized as follows. In the next section, the system framework of ComDA is proposed. The integrated models, methods and other techniques are also described in detail. Section 3 and section 4 present two case studies based on ComDA, with one that employs the Chinese Land Data Assimilation System. Method of extending ComDA by introducing a new model is also explained in these sections. High performance computing in ComDA is proposed in section 5. Discussions regarding ComDA are presented in section 6, and conclusions are drawn in section 7.