Elsevier

Remote Sensing of Environment

Volume 124, September 2012, Pages 61-71
Remote Sensing of Environment

Consistent retrieval methods to estimate land surface shortwave and longwave radiative flux components under clear-sky conditions

https://doi.org/10.1016/j.rse.2012.04.026Get rights and content

Abstract

Shortwave (0.3–3 μm) and longwave (3–50 μm) surface radiative flux components have been widely used in numerical prediction, meteorology, hydrology, biomass estimation, surface energy circulation and climate change studies, etc. However, during past decades, these components were usually estimated independently using different methods, possibly causing inconsistent estimation biases due to different atmospheric parameters and algorithms, especially for net surface fluxes. Two methods have been proposed in this paper to simultaneously derive surface shortwave (or longwave) radiative flux components based on MODIS products using an artificial neural network (ANN). The validation results show that the maximum root-mean-square error for downward and net shortwave radiative fluxes is less than 45 W/m2, about 60 W/m2 for direct solar radiation and 25 W/m2 for all longwave fluxes, which are comparable or even better than existing algorithms, thus demonstrating their feasibility and efficacy. The ANN-based models are then applied over the Tibetan Plateau region and the characteristics of the surface radiative flux components over such areas are analyzed.

Highlights

► Developed an ANN-based shortwave model for estimating SW radiative flux components. ► Developed an ANN-based longwave model for estimating LW radiative flux components. ► Better accuracies of fluxes can be achieved by using those newly developed methods. ► Proved feasibility to retrieve multi-components of flux from space consistently. ► Retrieved flux components reflect actual flux characteristics over Tibetan Plateau.

Introduction

The radiation (energy) budget of the earth–atmosphere system, which is the main driving force for the matter and energy cycle of the earth system, is currently one of the hottest research areas in the field of global climate change. Traditional ground-based radiation measurements have a long history with the data collected in a successive manner at temporal scales; however, the spatial coverage is sparse and even entirely absent in certain regions such as mountainous areas, oceans as well as most parts of the polar regions (Li et al., 1993). Thus it is impossible to accurately predict regional or global radiation budgets based solely on ground measurements. Satellite data with various spatial, temporal and spectral resolutions provide a unique opportunity to derive radiative fluxes at both the top of the atmosphere (TOA) and surfaces over large areas.

Since the 1970s, NASA has recognized the importance of improving our understanding of the earth's radiation budget (ERB) and its effects on the earth's climate using remotely sensed data. The first Earth Radiation Budget Experiment (ERBE) instrument onboard the Earth Radiation Budget Satellite (ERBS) was launched by the Space Shuttle Challenger in 1984 (Barkstrom, 1984, Barkstrom and Smith, 1986). Following the successful ERBE, the Clouds and the Earth's Radiant Energy System (CERES) has been developed to extend ERBE measurements to include the top of the atmosphere, within the atmosphere, and global surface radiation, which are critical for advancing our understanding of the Earth's climate system and improving climate prediction models (Wielicki et al., 1998). CERES was firstly flown on the Tropical Rainfall Measuring Mission in 1997, and then followed by the launches of EOS-terra and Aqua satellites in 1998 and 2000, respectively. Apart from these broadband missions, hyperspectral data collected from the Atmospheric Infrared Sounder (AIRS), which is capable of characterizing the entire atmospheric column from the surface to the TOA, have also been used to study the ERB (Sun et al., 2010). Various radiative flux products from these satellite data have been released by different scientific teams. In addition to the above-mentioned products, other sources of global ERB products are currently also available, such as (1) International Satellite Cloud Climatology project (ISCCP) which produced a global ERB product every 3 h from 1983 to 2006 on a 280 km scale (Zhang et al., 2004); and (2) the Global Energy and Water Cycle Experiment (GEWEX) which released a product with 3 h of temporal resolution over a 1° × 1° grid (Pinker et al., 2003). The common features of these products, including the products from ERBE and AIRS, are too coarse for many land applications; moreover, the accuracies and consistency of these products need to be further investigated. On this point, it is greatly expected that surface radiative fluxes with high accuracies and fine spatial resolutions will derive from currently available spaceborne data. Fortunately, the moderate-resolution imaging spectroradiometer (MODIS) flown aboard NASA's earth observing system (EOS) platforms, Terra and Aqua, offers complete global data with relative high temporal, spectral and spatial resolutions, and thus has been selected as the primary data source for the present study.

The radiation budget relates radiative fluxes involving the upward shortwave (SW) flux, downward SW flux, net SW flux, upward longwave (LW) flux, downward LW flux and net LW flux, etc. Among these fluxes, it is the net radiative fluxes (SW and LW) that are of particular significance for various applications, including global energy and water cycle, climate monitoring, and numerical weather prediction (Bisht et al., 2005). For retrieving these radiative fluxes from space, a number of algorithms have been developed during past decades which can be grouped into four classes: (1) empirical methods (Klink and dollhopf, 1986, Tarpley, 1979). They usually develop regressions by directly linking coincident satellite radiance and known surface radiative fluxes. Although they are simple to operate, the results are site-specific and cannot be extrapolated over other regions; (2) physics-based methods (Dedieu et al., 1987, Dubayah, 1995, Pinker and Ewing, 1985). They often employ a physically based radiative transfer code to perform surface radiative flux calculations given the necessary atmospheric and surface driving parameters, and thus are often computationally extensive. Theoretically, these methods are accurate but depend strongly upon driving parameters that may not be readily available; (3) parameterized methods (Bisht et al., 2005, Duarte et al., 2006, Sridhar and Elliott, 2002, Zhou and Cess, 2001, Zhou et al., 2007), that usually propose a parameterized scheme based on key atmospheric parameters (e.g., moisture, temperature) or empirical formula that may be invalid globally. These are computationally efficient compared to the physics-based methods, but require ancillary atmospheric parameters; and, (4) so-called hybrid methods (Kim and Liang, 2010, Li et al., 1993, Liang et al., 2006, Tang and Li, 2008, Tang et al., 2006, Wang and Liang, 2009, Wang and Liang, 2010, Wang et al., 2009, Zheng et al., 2008). Generally, such methods employ an extensive radiative transfer simulation followed by statistical modeling or construction of look-up tables based on a pre-simulated database, which may include band radiances of certain sensors, corresponding surface radiative fluxes, solar and observation geometry and atmospheric parameters, etc. The last method is not only free from requiring real-time atmospheric and surface information, but also keeps intact the physical properties; thus, it has become a promising trend in the field of radiative flux estimations in recent years.

Unfortunately, to date net radiative fluxes are frequently indirectly determined by subtracting upward fluxes from the dowelling flux in SW or LW spectral range (Bisht et al., 2005, Tang and Li, 2008, Wang and Liang, 2009). As pointed out by most researchers, the uncertainties associated with each of these quantities may lead to large errors in the final net surface radiation fluxes (Cess and Vulis, 1989, Li et al., 1993). Currently, most existing works that attempt to directly derive the net radiative fluxes are relatively complex and thus invite improvements in accuracy (Li et al., 1993, Tang et al., 2006). Additionally, to guarantee consistency in flux components, algorithms that can simultaneously derive the net surface radiative flux and related upward and downward components with a uniform process are urgently needed. To this end, consistent SW and LW radiative flux retrieval methods have been proposed in the current study based on readily available MODIS products using an artificial neural network (ANN). ANN, as a commonly used tool, has demonstrated utility for studying surface radiative flux budgets (Kim and Liang, 2010, Wang et al., 2009). Unlike the existing works mentioned above where the ANN was employed to derive only one certain surface radiative flux component (upward or downward, but not the net radiative flux components), the ANN is used here to simultaneously retrieve multi flux components, including the net SW (or LW) components in a single run.

The remainder of the paper is organized as follows: a brief introduction of ANN is described in Section 2. The construction of representative driving databases for radiative transfer simulation and newly proposed models for estimating surface radiative fluxes are also given in this section. Section 3 demonstrates the validation and application of these new models. The conclusion and discussion are presented in Section 4.

Section snippets

Artificial neural network

Artificial neural networks are used as a robust empirical statistical method in a variety of applications, ranging from classification, pattern recognition, forecasting, signal processing, control system, optimization to medicine and geosciences, etc. (Suzuki, 2011, chap. 1). The attractiveness of ANN comes from the remarkable information processing characteristics of biological systems such as nonlinearity, high parallelism, fault and failure tolerance, ability to handle imprecise and fuzzy

Results and analysis

Based on the training sets generated above and the designed ANN architectures, a BP algorithm is used to separately train the SW and LW ANN models. After a large number of iterations, the ANN-based SW and LW radiative models are built.

Conclusion and discussion

To address the lack of effective approaches for directly deriving surface SW and LW net radiative fluxes—the key parameters for most land applications—from satellite data, two novel ANN-based retrieval algorithms have been proposed in this study. The advantage of the suggested methods can be summarized in three points: (1) these models can directly output the net surface radiative fluxes and avoid the error propagation common in existing algorithms, which usually need to derive upward and

Acknowledgment

The work described in this publication has been partially supported by the NSFC (Grant nr. 40871164), National Basic Research Program of China, 973 Program (Grant nr. 2007CB714402) and the European Commission (Call FP7-ENV-2007-1 Grant nr. 212921) as part of the CEOP-AEGIS project (http://www.ceop-aegis.org/) coordinated by the Université de Strasbourg. The authors would like to thank the anonymous reviewers for their helpful and valuable comments to improve this work. Many thanks are given to

References (49)

  • A.M. Baldridge et al.

    The ASTER Spectral Library Version 2.0

    Remote Sensing of Environment

    (2008)
  • B.R. Barkstrom

    The Earth Radiation Budget Experiment (ERBE)

    Bulletin of the American Meteorological Society

    (1984)
  • B.R. Barkstrom et al.

    The earth radiation budget experiment: Science and implementation

    Reviews of Geophysics

    (1986)
  • A. Berk et al.
  • R.D. Cess et al.

    Inferring surface solar absorption from broadband satellite measurements

    Journal of Climate

    (1989)
  • F. Chevallier et al.

    Retrieving the clear skies vertical longwave radiative budget from TOVS: Comparison of a neural network-based retrieval and a method using geophysical parameters

    Journal of Applied Meteorology

    (2000)
  • G. Dedieu et al.

    Satellite estimation of solar irradiance at the surface of the earth and of surface albedo using a physical model applied to Meteosat data

    Journal of Applied Meteorology

    (1987)
  • H. Demuth et al.

    Neural network toolbox K 6

    User's guide

    (2008)
  • R. Dubayah

    An approach to the estimation of surface net radiation in mountain area using remote sensing and digital terrain data

    Theoretical and Applied Climatology

    (1995)
  • S. Haykin

    Neural networks: A comprehensive foundation

    (1994)
  • A.K. Jain et al.

    Artificial neural networks: A tutorial

    Computer

    (1996)
  • C. Jiménez et al.

    First inversions of observed submillimeter limb sounding radiances by neural networks

    Journal of Geophysical Research

    (2003)
  • J.C. Klink et al.

    An evaluation of satellite-based insolation estimates for Ohio

    Journal of Applied Meteorology

    (1986)
  • J.H. Lazaro et al.

    An inversion algorithm using neural networks to retrieve atmospheric CO total columns from high-resolution nadir radiances

    Journal of Geophysical Research

    (1999)
  • Cited by (79)

    View all citing articles on Scopus
    View full text