• Barker, H. W., , Curtis T. J. , , Leontieva E. , , and Stamnes K. , 1998: Optical depth of overcast cloud across Canada: Estimation based on surface pyranometer and satellite measurements. J. Climate, 11 , 29802994.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Calbó, J., , and Sabburg J. , 2008: Feature extraction from whole-sky ground-based images for cloud-type recognition. J. Atmos. Oceanic Technol., 25 , 314.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Estupiñán, J. G., , Raman S. , , Crescenti G. H. , , Streicher J. J. , , and Barnard W. F. , 1996: Effects of clouds and haze on UV-B radiation. J. Geophys. Res., 101 , 1680716816.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Feind, R. E., , Detwiler A. G. , , and Smith P. L. , 2000: Cloud liquid water measurements on the armored T-28: Intercomparison between Johnson–Williams cloud water meter and CSIRO (King) liquid water probe. J. Atmos. Oceanic Technol., 17 , 16301638.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Feingold, G., , Yang S. , , Hardesty R. M. , , and Cotton W. R. , 1998: Feasibility of retrieving cloud condensation nucleus properties from Doppler cloud radar, microwave radiometer, and lidar. J. Atmos. Oceanic Technol., 15 , 11881195.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Grund, C. J., , Banta R. M. , , George J. L. , , Howell J. N. , , Post M. J. , , Richter R. A. , , and Weickmann A. M. , 2001: High-resolution Doppler lidar for boundary layer and cloud research. J. Atmos. Oceanic Technol., 18 , 376393.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Hahn, C. J., , Rossow W. B. , , and Warren S. G. , 2001: ISCCP cloud properties associated with standard cloud types identified in individual surface observations. J. Climate, 14 , 1128.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Hansen, J. E., , and Travis L. D. , 1974: Light scattering in planetary atmospheres. Space Sci. Rev., 16 , 527610.

  • Heidinger, A. K., , Anne V. R. , , and Dean C. , 2002: Using MODIS to estimate cloud contamination of the AVHRR data record. J. Atmos. Oceanic Technol., 19 , 586601.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Huo, J., , and Lu D. , 2002: A primary study on cloud-cover using all-sky digital camera (in Chinese). J. Nanjing Inst. Meteor., 25 , 242246.

    • Search Google Scholar
    • Export Citation
  • Huo, J., , Lu D. , , and Wang Y. , 2006: Simulation of cloud detection threshold in all-sky images. Prog. Nat. Sci., 16 , 480484.

  • Johnson, R. H., , and Ciesielski P. E. , 2000: Rainfall and radiative heating rates from TOGA COARE atmospheric budgets. J. Atmos. Sci., 57 , 14971514.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Kassianov, E., , Long C. N. , , and Ovtchinnikov M. , 2005: Cloud sky cover versus cloud fraction: Whole-sky simulations and observations. J. Appl. Meteor., 44 , 8698.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Kubota, M., , Nagatsuma T. , , and Murayama Y. , 2003: Evening co-rotating patches: A new type of aurora observed by high sensitivity all-sky cameras in Alaska. Geophys. Res. Lett., 30 , 1612. doi:10.1029/2002GL016652.

    • Search Google Scholar
    • Export Citation
  • Le Treut, H., , Somerville R. , , Cubasch U. , , Ding Y. , , Mauritzen C. , , Mokssit A. , , Peterson T. , , and Prather M. , 2007: Historical overview of climate change science. Climate Change 2007: The Physical Science Basis, S. Solomon et al., Eds., Cambridge University Press, 93–127.

    • Search Google Scholar
    • Export Citation
  • Li, Z., , Cribb M. C. , , Chang F-L. , , and Trishchenko A. P. , 2004: Validation of MODIS-retrieved cloud fractions using whole sky imager measurements at the three ARM sites. Proc. 14th ARM Science Team Meeting, Albuquerque, NM, Atmospheric Radiation Measurement Program, 6 pp.

    • Search Google Scholar
    • Export Citation
  • Long, C. N., , Sabburg J. M. , , Calbó J. , , and Pagès D. , 2006: Retrieving cloud characteristics from ground-based daytime color all-sky images. J. Atmos. Oceanic Technol., 23 , 633652.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Lu, D., , Huo J. , , and Lu Y. , 2001: The scientific technologic problems and primitive test of ground-based all-sky imager remote sensing. Chinese Remote Sensing—20 Year, China Meteorological Press, 114–120.

    • Search Google Scholar
    • Export Citation
  • Mayer, B., , and Kylling A. , 2005: Technical note: The libRadtran software package for radiative transfer calculations—Description and examples of use. Atmos. Chem. Phys., 5 , 18551877.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • McGuffie, K., , and Henderson-Sellers A. , 1989: Almost a century of “imaging” clouds over the whole-sky dome. Bull. Amer. Meteor. Soc., 70 , 12431253.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Pfister, G., , McKenzie R. L. , , Liley J. B. , , Thomas A. , , Forgan B. W. , , and Long C. N. , 2003: Cloud coverage based on all-sky imaging and its impact on surface solar irradiance. J. Appl. Meteor., 42 , 14211434.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Rossow, W. B., , and Schiffer R. A. , 1991: ISCCP cloud data products. Bull. Amer. Meteor. Soc., 72 , 220.

  • Sabburg, J., , and Long C. , 2004: Improved sky imaging for studies of enhanced UV irradiance. Atmos. Chem. Phys., 4 , 25432552.

  • Shields, J. E., , Johnson R. W. , , Karr M. E. , , and Wertz J. L. , 1998: Automated day/night whole sky imagers for field assessment of cloud cover distributions and radiance distributions. Preprints, 10th Symp. on Meteorological Observations and Instrumentation, Phoenix, AZ, Amer. Meteor. Soc., 4.7.

    • Search Google Scholar
    • Export Citation
  • Slingo, A., , and Schrecker H. M. , 1982: On the shortwave radiative properties of stratiform water clouds. Quart. J. Roy. Meteor. Soc., 108 , 407426.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Stephens, G. L., 2005: Cloud feedbacks in the climate system: A critical review. J. Climate, 18 , 237273.

  • Voss, J. K., , and Zibordi G. , 1989: Radiometric and geometric calibration of a visible spectral electro-optic “fisheye” camera radiance distribution system. J. Atmos. Oceanic Technol., 6 , 652662.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Wang, K., , Dickinson R. E. , , and Liang S. , 2009: Clear sky visibility has decreased over land globally from 1973 to 2007. Science, 323 , 14681470.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • View in gallery

    All-sky imager: (left) outdoor instrument and (right) a case of all-sky image.

  • View in gallery

    Radiance ratio distribution of cloudless sky when AOD is (left) 0.30 and (right) 0.68. The center of the circle is the zenith and the edge is the horizon. The solar zenith angle is 45°. The arrow shows the solar position, the black line, which goes through the solar position and zenith, is the solar principal plane.

  • View in gallery

    Radiance ratio on the solar principal plane AODs of (top)–(bottom) 0.25, 0.36, 0.48, 0.68, 1.20, and 2.90; (left)–(right) the points on the solar principal plane are labeled from 0 to 180.

  • View in gallery

    Radiance ratio distribution of cloudy sky, where COD is (left) 1 and (right) 10.

  • View in gallery

    Radiance ratio from three directions changing with COD: (left) zenith angle = 30° and azimuth angle = 270°, (middle) zenith angle = 45° and azimuth angle = 270°, and (right) zenith angle = 0° and azimuth angle = 0°.

  • View in gallery

    Probability of fixed threshold of 1.30 for determining cloud correctly at five solar angles.

  • View in gallery

    (a) A 512 × 512 pixel frame; (top) sample images and (bottom) images of spectrum power after FFT for (b),(d) cloud free and (c),(e) cloudy skies.

  • View in gallery

    Symmetry of all-sky image: (left) whole gray images and (right) false gray image depending on B/R ratio.

  • View in gallery

    A case of histogram and edge-searching approaches: (left) the histogram made from Fig. 10a (estimated threshold is 1.15) and (right) the highest 20 std devs and their B/R ratios (average ratio is 1.03). The final threshold is 1.09.

  • View in gallery

    Intercomparison of cloud determinations by IM and SM: (top) original image and false color images after (middle) IM and (bottom) SM. White represents cloud and blue represents cloudless. The thresholds of IM are (a) 0.8, (b) 1.25, and (c) 1.09. The threshold of SM is 1.30.

  • View in gallery

    Number difference at 11 cloud fractions from IM – OB and SM – OB. The 4000 images are taken at XiangHe Atmospheric Integrated Observatory of IAP from 2005 to 2007.

All Time Past Year Past 30 Days
Abstract Views 0 0 0
Full Text Views 78 78 9
PDF Downloads 60 60 7

Cloud Determination of All-Sky Images under Low-Visibility Conditions

View More View Less
  • 1 Key Laboratory for Atmosphere and Global Environment Observation, Chinese Academy of Sciences, Beijing, China
© Get Permissions
Full access

Abstract

The threshold method is commonly used to determine cloud in a sky image. This paper evaluates the method by numerical simulation and shows that the aerosol optical depth (AOD) is a key factor that influences the accuracy. Particularly when the visibility is low, a single threshold method is inappropriate. To improve the accuracy of cloud determination from low-visibility sky images, an integrated cloud-determination algorithm is presented that is based on the fast Fourier transform, symmetrical image features, and threshold methods. The preliminary comparison tests show that the new integrated method improves the ability to determine cloud under lower-visibility conditions.

Corresponding author address: Dr. Juan Huo, LAGEO, Institute of Atmospheric Physics, Chinese Academy of Sciences, Beijing, China 100029. Email: huojuan@mail.iap.ac.cn

Abstract

The threshold method is commonly used to determine cloud in a sky image. This paper evaluates the method by numerical simulation and shows that the aerosol optical depth (AOD) is a key factor that influences the accuracy. Particularly when the visibility is low, a single threshold method is inappropriate. To improve the accuracy of cloud determination from low-visibility sky images, an integrated cloud-determination algorithm is presented that is based on the fast Fourier transform, symmetrical image features, and threshold methods. The preliminary comparison tests show that the new integrated method improves the ability to determine cloud under lower-visibility conditions.

Corresponding author address: Dr. Juan Huo, LAGEO, Institute of Atmospheric Physics, Chinese Academy of Sciences, Beijing, China 100029. Email: huojuan@mail.iap.ac.cn

1. Introduction

Clouds cover more than 50% of the globe and have a large impact on the radiative energy transfer of the atmosphere–earth surface system (Rossow and Schiffer 1991; Estupiñán et al. 1996; Johnson and Ciesielski 2000; Stephens 2005; Le Treut et al. 2007). Many powerful instruments are now being operated for cloud detecting: meteorological satellites make large-scale observation of global clouds (Hahn et al. 2001; Heidinger et al. 2002); meteorological radars detect particles and motions of cloud (Feind et al. 2000); and many kinds of ground-based instruments, such as lidar and microwave radiometer (Feingold et al. 1998; Grund et al. 2001), are developed to detect the LWP, bottom height of cloud, etc. However, at most meteorological observing stations, ground-based macroscopic observations of cloud are still performed by human observers (OBs), who record cloud fraction, cloud types, and so on. The inevitable subjectivity and the high costs associated with excessive human involvement generate demand for automatic devices that can detect cloud automatically and replace human observers in the future.

The development history of ground-based cloud-observing instruments spans over a century (McGuffie and Henderson-Sellers 1989). Because of developments of computing technology and digital imaging technology, great progress has been made. Currently, two types of imager systems are frequently referred to. One is the Whole Sky Imagers series (WSIs) developed by the Scripps Institute of Oceanography at the University of California, San Diego. WSIs measure radiance over the sky and retrieve cloud characteristics (Voss and Zibordi 1989; Shields et al. 1998; Li et al. 2004; Kassianov et al. 2005). WSIs use advanced technology to obtain radiances at distinct wavelengths, protecting the optics and charge-coupled devices (CCDs) from stray light and controlling environmental temperature, but the resultant higher cost limits its popularization among meteorological researchers. In comparison, the well-known and more commonly used sky imager in the atmospheric science community is the Total Sky Imager (TSI); for example, TSI is now used in the Surface Radiation Budget Network and the Atmospheric Radiation Measurement Program (Pfister et al. 2003). TSI provides daytime time series of hemispheric images and derives fractional sky cover. More research on extracting useful meteorological information (e.g., cloud types or classifications) from the TSI images has been conducted, and it has made some achievements (Long et al. 2006; Calbó and Sabburg 2008). Other sky imagers have also been developed in some other countries and institutes, such as the whole-sky camera (WSC) developed by Spain’s University of Girona to obtain whole-sky images (Long et al. 2006; Calbó and Sabburg 2008) and the all-sky imager (our imager system has the same name but different scientific functions) developed by Japanese Communications Research Laboratory to observe airglow and aurora (Kubota et al. 2003).

At present, cloud is determined mainly through comparing the ratio of red channel to blue channel (or blue to red) of each pixel with a predefined threshold (the WSIs, TSIs, and WSCs all employ this method). TSI uses different thresholds, depending on the sector of the sky (area close to horizon, circumsolar area, etc.). The pixels for which red/blue ratios are less than the threshold are labeled as “cloudless” and the pixels for which red/blue ratios exceed the threshold are labeled as “cloud.” The principle of this method comes from the different scattering between air molecules and cloud particles. Molecular Rayleigh scattering (clear sky) is wavelength λ dependent as λ4, whereas cloud scattering is almost independent of wavelength. Therefore, more blue light is scattered than red for clear sky, whereas clouds approximately scatter the blue and red visible light equally (Sabburg and Long 2004); this is why clear sky appears blue and cloud appears white to our eyes. However, aerosol particles loaded in the atmosphere also scatter and absorb incident radiation and change the radiance spectral distribution. Hence, aerosols also decide how white tinted the “blue” of the sky will appear, and the cloudless sky will look whiter when there are more aerosol particles in the atmosphere (Long et al. 2006). For threshold setting, aerosol should be taken into account to avoid aerosol pixels for which the red/blue exceeds the threshold that is wrongly determined as cloud.

Visibility, the maximum distance at which an observer can discern the outline of an object against the horizon sky, is reduced mainly by the presence of aerosol particles (dust, smoke, and haze) and hydrometeors [large droplets or crystals of water (>5 μm) occurring as rain, fog, clouds, and snow; Wang et al. 2009]. Obviously, more aerosol particles in the atmosphere will reduce the visibility. Section 2 briefly introduces the all-sky imager system. Then, in section 3, the impact of aerosol particles on cloud determination is evaluated through numerical simulation. Section 4 presents an integrated cloud-determination algorithm to improve the accuracy of cloud determination and provides a contrast with the single threshold method (SM). A summary and conclusions are given in section 5.

2. Sky cameras and associated images

The sky images used in this paper are provided by an all-sky imager system (ASIs-I) developed by the Key Laboratory for Atmosphere and Global Environment Observation, Institute of Atmospheric Physics (IAP), Chinese Academy of Sciences (Lu et al. 2001; Huo and Lu 2002; Huo et al. 2006). The imager (Fig. 1) has three major components: 1) an imaging system, where a digital camera with a fish-eye lens (field of view is about 183°) is used to obtain the sky images at 2272 × 1704 pixel resolution and the shortest repeat time is 1 min (variable); 2) an obscuring system, where a movable shadow device is designed to protect the CCD sensor from direct solar radiation; and 3) an outside box equipped with fans and heating devices that is used to protect the camera from snow, rain, etc., and keeps the temperature in the box in the range of −10° to 40°C.

3. Numerical simulation

The Library for Radiative Transfer (LIBRADTRAN) pseudosphere radiative transfer model (Mayer and Kylling 2005) is used for numerical simulation. It is used to calculate the radiance of cloudless sky and clouds over the hemisphere at the visible band. Radiances at 450- and 650-nm wavelengths over the sky dome obtained at the surface are supposed to represent the blue and red light detected by the all-sky imager, respectively. In fact, optical filters of our all-sky imager system have a relative broadband at the three channels, red, green, and blue, and each channel has a central wavelength of 650, 550, and 450 nm, respectively. Obviously, radiance simulated by the model is not completely what the imager receives. However, numerical simulation is a simple and effective approach to the evaluation.

Ratios of 450-nm radiance to 650-nm radiance under different aerosol optical depth (AOD) conditions are analyzed to understand how aerosol particles change sky color and to what degree. For a simpler description, the ratio of 450-nm radiance to 650-nm radiance is called “radiance ratio” and the gray ratio of blue-channel to red-channel pixels is called “B/R ratio.”

a. Numerical simulation for cloudless sky under different AOD conditions

Parameters (e.g., solar position, aerosol type, atmospheric profile) input in the LIBRADTRAN model are identical, except for the AOD value. The aerosol type is set as rural. For other input data, such as the profile of pressure and temperature, default values of U.S. Standard Atmosphere, 1976 are used for the representation. In the model, 12 different AOD values are chosen for the simulation that can represent different visibilities happening in nature. Table 1 shows a raw relationship between AOD and visibility (from the LIBRADTRAN dataset).

The contour map of Fig. 2 shows the distribution of radiance ratio and indicates what color the sky appears under different AODs. For the case where AOD is 0.30 (Fig. 2, left), all radiance ratios are larger than 1.30, which implies that the blue component of light is larger than the red component, and the sky will appear blue. The maximum ratio value appears near the anthelic, and the minimum value appears near the horizon. Directions near the anthelic will look bluer than other directions. However, when AOD increases (Fig. 2, right), the radiance ratio decreases all over the hemisphere, which implies that the sky looks whiter. More numerical simulations are made with different AOD values for the evaluation. The solar principal plane (pixels on it have the same azimuth angle as the sun; horizontal line on Fig. 2, right) is used to represent the hemisphere for a description so that all typical simulation results can be shown in one figure for comparison. Results under six different AODs are presented in Fig. 3. It can be seen that the radiance ratio has a negative relationship with the AOD. Under lower AOD conditions (AOD < 0.68), most radiance ratios (scattering angle <160°) are larger than 1.5. When the AOD is 2.9, the radiance ratio has its lowest value among these cases and is never greater than 1.5; the ratio is less than 1.0 in some directions. This proves that aerosol particles affect atmospheric scattering and change the distributions of the radiance ratio. Therefore, a fixed single threshold is not suitable for cloud determination for those places where the AOD is not always lower. Fortunately, it can be seen from Fig. 3 that the radiance ratio has a monotonic relationship with AOD, which indicates the possibility of using various thresholds to determinate cloud, depending on different AODs.

b. Numerical simulation of cloudy sky

Clouds are set as horizontally homogeneous. However, this rarely happens in nature. Hence, the simulation results of cloud cannot represent the real situations of nature. However, this will reflect some basic scattering properties of clouds and supports us with some references on cloud determination and threshold choice. Work on 3D inhomogeneous cloud modeling will be performed in the near future. In the simulations, visibility is set as 23 km (AOD is about 0.33), so that the same influence of aerosol can be ignored when analyzing the radiance ratio. Effective particle radius γe (μm) and liquid water content (LWC; g m−3) are changed individually to obtain various cloud optical depths (CODs) in the model. According to previous works that the values of γe for natural clouds mainly vary from 2 to 20 μm (Hansen and Travis 1974; Slingo and Schrecker 1982) and most CODs are no more than 35 (except in some extreme conditions; Barker et al. 1998), the scope of γe is set from 5 to 16 μm and the COD is set below 40 when changing LWC in the model. Other parameters (i.e., solar position, profile of atmosphere) input in the model are the same as in section 3a.

Two cases are selected for the representation to describe. Thin clouds (COD = 1; see Fig. 4, left) scatter and absorb light slightly, so the ratio distribution looks more like what cloudless sky looks like (comparing with Fig. 2, right). When COD increases to 10 (see Fig. 4, right), clouds over the sky have approximately equal ratios and the position of each pixel is not an important factor on influencing the distribution of the ratio. For further explanation of all these simulation results expediently, three directions are chosen to illustrate the variation of radiance ratio with COD. The three directions are d1 (zenith angle 0°, azimuth angle 0°), d2 (zenith angle 30°, azimuth angle 270°), and d3 (zenith angle 45°, azimuth angle 270°). There are 5 curves in Fig. 5 representing 5 different values of γe (from 5 to 16 μm). LWC is changed by the step of 0.01 g m−3 (from 0.01 to 0.3 g m−3), and γe is changed by steps of 1 μm when numerically simulating. From Fig. 5, when the COD is larger than 10, it can be seen that radiance ratio distributions of the three directions look like each other and that most of the radiance ratios are less than 1.30. Therefore, the threshold 1.30 may be a good selection for cloud determination when comparing with the results concluded by section 3a, because most radiance ratios of cloudless sky are larger than 1.30.

Particularly, when the COD is less than 10, the radiance ratio changes rapidly and decreases sharply with the increase of COD, and it is also affected by the position relative to the sun. Many of radiance ratios are far greater than 1.30. Therefore, thin cloud is also difficult to recognize from cloudless pixels.

c. Evaluation of fixed single threshold

The fixed single threshold of 1.30 is selected for cloud determination for the evaluation. In this subsection, we estimate the correctness probability of the single threshold when it is used to determinate cloud of cloudless sky under different AOD conditions. Pixels for which the B/R ratio is less than 1.30 are recognized as cloud, whereas pixels for which the B/R ratio exceeds 1.30 are recognized as cloudless. The value Ac, which is defined as Nnc/Na (where Nnc is the number of cloudless pixels determined by a fixed single threshold and Na is the number of all pixels), is used to indicate the correctness probability of the single threshold on cloud determination. Because all the cases are cloudless, the expected result of the cloud identification should be Ac = 1. Figure 6 shows that the probability decreases with the increase of AOD. When the AOD is less than 0.48, Ac is greater than 0.8, which indicates that the fixed threshold of 1.30 is a good index in cloud determination for high-visibility conditions (excluding thin cloud). However, when AOD is greater than 0.86, the probability decreases noticeably, which means that the accuracy is influenced by aerosols and it is inappropriate under lower-visibility conditions. In fact, aerosol and thin cloud are two major factors that cause errors in cloud determination (it can be concluded from the work here). Therefore, the probability estimated in Fig. 6 is higher than actual situations, because the thin-cloud condition is not considered. So, for those places where AOD is not always lower (e.g., at some urban areas in China with hazy sky), the fixed threshold will generate more errors, and an adaptive threshold that depends on different sky conditions is required.

4. Integrated method for cloud determination and contrast

To improve accuracy in lower visibilities, a self-adaptive threshold is needed according to different AODs. More efficient processing methods should be exploited.

a. FFT method

The fast Fourier transform (FFT) is an important approach in digital image processing. It transforms images from the space domain into the frequency domain so that the variance of pixels can be illustrated clearly and quantified. To exclude the pixels of the sheltering device in the image and get the maximum areas for FFT, B/R ratios of a 512 × 512 pixel block are used to make FFTs (Fig. 7a). From Fig. 7, we can see that the differences between cloudless (Figs. 7b,d) and cloudy images (Figs. 7c,e) are clearly illustrated after FFT. Hence, the FFT can be used to analyze whether the sky image has cloud, because most cloudy images are inhomogeneous.

In the integrated method (IM), we use two approaches to determine whether the sky is homogeneous. Generally, the direct result of the FFT is the complex amplitudes of the harmonics that correspond to each wavenumber. Then, we calculate the spectral power function by normalizing the modulus of these complex amplitudes by the size of the image. First, the number of pixels for which the spectral power are larger than 255 is calculated (see Fig. 7; pixels are white when the spectral power is larger than 255 in the FFT images). According to our statistical results made from cloudy and cloudless images, the number of cloudy images is often lager than cloudless images. For our images, the threshold is set as 275. An image with less than 275 is considered homogeneous. Second, we predefine a series of clear-sky images at different solar angles (e.g., selecting all the images from sunrise to sunset of a day that is cloudless all day) as templates. Then, the template image that has the nearest time is chosen to calculate the correlation coefficient of the FFT. An image with a correlation coefficient greater than 0.98 is considered homogeneous. When both of the approaches have the same conclusions, the image is thought as homogeneous. Homogeneous images are cloudless or have a cloud fraction of 10 tenths (or overcast). An inhomogeneous image means that it is cloudy. Once the FFT method proves that clouds exist in the image, a further algorithm will be used to get an appropriate threshold for cloud recognition; this will be discussed in section 4c.

Obviously, the FFT will improve the accuracy of cloud determination in the cloudless days with fog or dust, which always have lower B/R ratios. Meanwhile, once the FFT algorithm detects that clouds exist, the opportunity of obtain an appropriate threshold is increased.

b. Symmetrical method

Sky radiance measured on the ground base is the light scattered or reflected by the surface and atmosphere. Except for in extreme weather, such as dust storms, the atmosphere of cloudless sky can be assumed as horizontally uniform. Therefore, the sky radiance field is a symmetrical field and the symmetrical axis is ideally the solar principal plane (see Fig. 8). The symmetrical method only employs this property. For each pixel, the B/R ratio difference for the symmetrical pixel is calculated. If the difference exceeds a predefined limit (i.e., 0.1, which is defined in our algorithm), the pixel with the lower ratio is thought of as cloudy and the other pixel as cloudless. On the other hand, if the difference is less than the limit, both pixels are determined to be cloudy (or cloudless) by comparing their mean B/R ratio with the threshold.

c. Procedure of cloud determination

Cloud determination of all-sky images is challenging work, because of the uncertainties and complexity of cloud and aerosol distribution in the atmosphere. The analysis in this paper shows the feasibility and effectiveness of the three methods. According to the priorities and shortcomings of each method, we set up a new integrated method for cloud determination. The main calculating processes are described as follows.

First of all, a 512 × 512 block dispatched from the all-sky image is analyzed by the FFT method, and it is judged whether the image is homogeneous. For a homogeneous image, the average B/R ratio is used to decide whether the sky is cloudless. The image for which the average ratio exceeds 1.20 is considered cloudless, and the others are considered overcast. For the inhomogeneous image, two algorithms are used to obtain a right threshold. First, a histogram of the B/R ratio is made. We chose seven intervals to make a histogram: [1.075, 1.125], [1.175, 1.225], [1.275, 1.325], [1.375, 1.425], [1.475, 1.525], [1.575, 1.625], and [1.675, 1.725]. Generally, the B/R ratio of cloud is less than the cloudless sky in the same image. The threshold is selected from the two ratio intervals that have the greatest and second greatest numbers. Figure 9 (left) gives a sample of a histogram made from Fig. 10c. Second, an edge-searching method is performed. For each pixel, the surrounding B/R ratios (10 × 10 pixels are chosen in our algorithm) and the standard deviation of gray is calculated. Pixels that have the highest 20 standard deviations are chosen for further B/R ratio analysis, to determine whether the pixel is the cloud edge. The B/R ratios of those pixels are then averaged to obtain the threshold. The finial threshold is the average of the values obtained by the histogram and edge-searching methods.

In addition, we find that the B/R ratios of the pixels at lower elevation angles are often less than the ones at higher elevation angles, because the atmospheric cosine effect (this is one of the reasons why TSI uses several thresholds, depending on different positions) and the threshold method easily make mistakes in cloud determination. So, the symmetrical method is used on those regions where the zenith angle is greater than 75° (set in the integrated method).

d. Contrast

As mentioned before, a fixed single threshold (or several fixed thresholds for different positions of relative solar position or shadow belt) is suitable for a place where the AOD is always low (e.g., less than 0.48). The superiority of the integrated method will not be demonstrated. However, for a place where the AOD is not always low, for example, it is great on some days and low on other days. Some urban areas in China are the places like this. Because the visibility is unknown, a fixed threshold cannot adjust to all conditions, and the new integrated method demonstrates the superiority; Fig. 10 shows this. For SM, 1.30 is chosen as the threshold for cloud determination. The false color image shows the recognition results: white represents cloud and blue represents cloudless sky. For cloudless sky (Fig. 10a; the visibility is about 8 km from the record by the observer), pixels near the horizon and solar direction are incorrectly determined as cloud by SM. However, the IM finds that it is homogeneous by FFT and regards it as cloudless. For a cloudy case that has a high visibility (Fig. 10b; about 20 km), both SM and IM distinguish the cloud clearly. The threshold estimated by IM is 1.25, which is near 1.30 and illustrates that 1.30 is suitable for this case. However, for another cloudy case that has low visibility (Fig. 10c; about 8 km), SM overestimates cloud greatly, whereas IM gets a correct threshold (threshold = 1.09).

For further comparisons, there are about 4000 images with relatively low visibility (less than 15 km, according to the record from the observer) selected during 2 years from 2005 to 2007. These images are taken at the XiangHe Atmospheric Integrated Observatory (39.75°N, 116.96°E) of the IAP. Note that there is no standard of cloud fraction for each all-sky image, and even the cloud fraction estimated by a human is sometimes erroneous because of the subjectivity and polluted hazy sky. So, we make an intercomparison of cloud fraction estimated by the three approaches: IM, SM, and the experienced OB. Figure 11 shows the histogram of differences between SM − OB and IM − OB at 11 cloud fractions (0–10). It can be seen that cloud fraction estimated by IM is closer to the value estimated by an observer. The comparison results make it clear that the new integrated method makes progress in cloud determination under low visibilities. Through further analysis, we also find that the IM often underestimates the cloud amount when there are thin clouds (e.g., cirrus). There are two major reasons: 1) thin cloud is not detected by the FFT method, and it is determined as cloudless, and 2) an exact threshold is not set, and the B/R ratio of thin cloud exceeds the threshold. In addition, when the sky is cloudy, IM and SM tend to overestimate cloud fraction (e.g., if the cloud fraction by OB is 3 or 4 tenths, then it is overestimated to 5 or 6 by IM, or if cloud fraction by OB is 8 or 9, then it is overestimated to 9 or 10). There are two possible reasons: 1) regions close to solar position and horizon are erroneously regarded as cloud or 2) the observer underestimates cloud fraction. Therefore, though the IM has made progress in contrast with the SM, it also has differences from the observer and needs to be improved by adding other methods. Meanwhile, thin-cloud determination, especially under low-visibility conditions, is still a challenging work and needs more research in the future.

5. Summary and conclusions

All-sky imagers (e.g., WSI, TSI, WSC) get sky images over the hemisphere automatically at high temporal and spatial resolutions. It avoids the subjectivity from the observer and reduces human costs on daily ground-based cloud observing greatly. Cloud determination from sky image is a key and necessary work for the further application of all-sky images into operational meteorological services and research, because the retrieval parameters of cloud, such as cloud fraction and cloud types, can be used directly in meteorological studies. Cloud determination is the basic and important work for the all-sky image application, because it is the precondition for deriving other information, such as cloud types, cloud brokenness, and cloud thickness. This paper places an emphasis on the analysis of impacts of aerosol particles on cloud determination, and it presents a new method for cloud determination under low visibility. The new integrated method is based on the FFT, symmetrical, and threshold methods. Contrast experiments show that the new integrated method improves the accuracy of cloud determination in low visibility. Work in this paper help to improve and extend the application availability of all-sky images taken under low-visibility conditions. It is useful for the urban cities with hazy sky and for their further applications in some other research topics, such as climate model simulation and studies on climate change, climate modeling, cloud effects on radiative transfer in atmosphere, and so on. However, thin-cloud determination is still a problem for the integrated method, and more research or other processing methods are required to improve this ability.

Acknowledgments

The authors acknowledge the support from the National Natural Science Foundation of China (Grant 40505006). We thank Yong Wang for his work in the all-sky imager development. In addition, we thank Weidong Nan and Qinghong Wu (XiangHe Atmospheric Integrated Observatory of IAP) for visually inspecting the set of ASI images.

REFERENCES

  • Barker, H. W., , Curtis T. J. , , Leontieva E. , , and Stamnes K. , 1998: Optical depth of overcast cloud across Canada: Estimation based on surface pyranometer and satellite measurements. J. Climate, 11 , 29802994.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Calbó, J., , and Sabburg J. , 2008: Feature extraction from whole-sky ground-based images for cloud-type recognition. J. Atmos. Oceanic Technol., 25 , 314.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Estupiñán, J. G., , Raman S. , , Crescenti G. H. , , Streicher J. J. , , and Barnard W. F. , 1996: Effects of clouds and haze on UV-B radiation. J. Geophys. Res., 101 , 1680716816.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Feind, R. E., , Detwiler A. G. , , and Smith P. L. , 2000: Cloud liquid water measurements on the armored T-28: Intercomparison between Johnson–Williams cloud water meter and CSIRO (King) liquid water probe. J. Atmos. Oceanic Technol., 17 , 16301638.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Feingold, G., , Yang S. , , Hardesty R. M. , , and Cotton W. R. , 1998: Feasibility of retrieving cloud condensation nucleus properties from Doppler cloud radar, microwave radiometer, and lidar. J. Atmos. Oceanic Technol., 15 , 11881195.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Grund, C. J., , Banta R. M. , , George J. L. , , Howell J. N. , , Post M. J. , , Richter R. A. , , and Weickmann A. M. , 2001: High-resolution Doppler lidar for boundary layer and cloud research. J. Atmos. Oceanic Technol., 18 , 376393.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Hahn, C. J., , Rossow W. B. , , and Warren S. G. , 2001: ISCCP cloud properties associated with standard cloud types identified in individual surface observations. J. Climate, 14 , 1128.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Hansen, J. E., , and Travis L. D. , 1974: Light scattering in planetary atmospheres. Space Sci. Rev., 16 , 527610.

  • Heidinger, A. K., , Anne V. R. , , and Dean C. , 2002: Using MODIS to estimate cloud contamination of the AVHRR data record. J. Atmos. Oceanic Technol., 19 , 586601.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Huo, J., , and Lu D. , 2002: A primary study on cloud-cover using all-sky digital camera (in Chinese). J. Nanjing Inst. Meteor., 25 , 242246.

    • Search Google Scholar
    • Export Citation
  • Huo, J., , Lu D. , , and Wang Y. , 2006: Simulation of cloud detection threshold in all-sky images. Prog. Nat. Sci., 16 , 480484.

  • Johnson, R. H., , and Ciesielski P. E. , 2000: Rainfall and radiative heating rates from TOGA COARE atmospheric budgets. J. Atmos. Sci., 57 , 14971514.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Kassianov, E., , Long C. N. , , and Ovtchinnikov M. , 2005: Cloud sky cover versus cloud fraction: Whole-sky simulations and observations. J. Appl. Meteor., 44 , 8698.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Kubota, M., , Nagatsuma T. , , and Murayama Y. , 2003: Evening co-rotating patches: A new type of aurora observed by high sensitivity all-sky cameras in Alaska. Geophys. Res. Lett., 30 , 1612. doi:10.1029/2002GL016652.

    • Search Google Scholar
    • Export Citation
  • Le Treut, H., , Somerville R. , , Cubasch U. , , Ding Y. , , Mauritzen C. , , Mokssit A. , , Peterson T. , , and Prather M. , 2007: Historical overview of climate change science. Climate Change 2007: The Physical Science Basis, S. Solomon et al., Eds., Cambridge University Press, 93–127.

    • Search Google Scholar
    • Export Citation
  • Li, Z., , Cribb M. C. , , Chang F-L. , , and Trishchenko A. P. , 2004: Validation of MODIS-retrieved cloud fractions using whole sky imager measurements at the three ARM sites. Proc. 14th ARM Science Team Meeting, Albuquerque, NM, Atmospheric Radiation Measurement Program, 6 pp.

    • Search Google Scholar
    • Export Citation
  • Long, C. N., , Sabburg J. M. , , Calbó J. , , and Pagès D. , 2006: Retrieving cloud characteristics from ground-based daytime color all-sky images. J. Atmos. Oceanic Technol., 23 , 633652.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Lu, D., , Huo J. , , and Lu Y. , 2001: The scientific technologic problems and primitive test of ground-based all-sky imager remote sensing. Chinese Remote Sensing—20 Year, China Meteorological Press, 114–120.

    • Search Google Scholar
    • Export Citation
  • Mayer, B., , and Kylling A. , 2005: Technical note: The libRadtran software package for radiative transfer calculations—Description and examples of use. Atmos. Chem. Phys., 5 , 18551877.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • McGuffie, K., , and Henderson-Sellers A. , 1989: Almost a century of “imaging” clouds over the whole-sky dome. Bull. Amer. Meteor. Soc., 70 , 12431253.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Pfister, G., , McKenzie R. L. , , Liley J. B. , , Thomas A. , , Forgan B. W. , , and Long C. N. , 2003: Cloud coverage based on all-sky imaging and its impact on surface solar irradiance. J. Appl. Meteor., 42 , 14211434.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Rossow, W. B., , and Schiffer R. A. , 1991: ISCCP cloud data products. Bull. Amer. Meteor. Soc., 72 , 220.

  • Sabburg, J., , and Long C. , 2004: Improved sky imaging for studies of enhanced UV irradiance. Atmos. Chem. Phys., 4 , 25432552.

  • Shields, J. E., , Johnson R. W. , , Karr M. E. , , and Wertz J. L. , 1998: Automated day/night whole sky imagers for field assessment of cloud cover distributions and radiance distributions. Preprints, 10th Symp. on Meteorological Observations and Instrumentation, Phoenix, AZ, Amer. Meteor. Soc., 4.7.

    • Search Google Scholar
    • Export Citation
  • Slingo, A., , and Schrecker H. M. , 1982: On the shortwave radiative properties of stratiform water clouds. Quart. J. Roy. Meteor. Soc., 108 , 407426.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Stephens, G. L., 2005: Cloud feedbacks in the climate system: A critical review. J. Climate, 18 , 237273.

  • Voss, J. K., , and Zibordi G. , 1989: Radiometric and geometric calibration of a visible spectral electro-optic “fisheye” camera radiance distribution system. J. Atmos. Oceanic Technol., 6 , 652662.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Wang, K., , Dickinson R. E. , , and Liang S. , 2009: Clear sky visibility has decreased over land globally from 1973 to 2007. Science, 323 , 14681470.

    • Crossref
    • Search Google Scholar
    • Export Citation

Fig. 1.
Fig. 1.

All-sky imager: (left) outdoor instrument and (right) a case of all-sky image.

Citation: Journal of Atmospheric and Oceanic Technology 26, 10; 10.1175/2009JTECHA1324.1

Fig. 2.
Fig. 2.

Radiance ratio distribution of cloudless sky when AOD is (left) 0.30 and (right) 0.68. The center of the circle is the zenith and the edge is the horizon. The solar zenith angle is 45°. The arrow shows the solar position, the black line, which goes through the solar position and zenith, is the solar principal plane.

Citation: Journal of Atmospheric and Oceanic Technology 26, 10; 10.1175/2009JTECHA1324.1

Fig. 3.
Fig. 3.

Radiance ratio on the solar principal plane AODs of (top)–(bottom) 0.25, 0.36, 0.48, 0.68, 1.20, and 2.90; (left)–(right) the points on the solar principal plane are labeled from 0 to 180.

Citation: Journal of Atmospheric and Oceanic Technology 26, 10; 10.1175/2009JTECHA1324.1

Fig. 4.
Fig. 4.

Radiance ratio distribution of cloudy sky, where COD is (left) 1 and (right) 10.

Citation: Journal of Atmospheric and Oceanic Technology 26, 10; 10.1175/2009JTECHA1324.1

Fig. 5.
Fig. 5.

Radiance ratio from three directions changing with COD: (left) zenith angle = 30° and azimuth angle = 270°, (middle) zenith angle = 45° and azimuth angle = 270°, and (right) zenith angle = 0° and azimuth angle = 0°.

Citation: Journal of Atmospheric and Oceanic Technology 26, 10; 10.1175/2009JTECHA1324.1

Fig. 6.
Fig. 6.

Probability of fixed threshold of 1.30 for determining cloud correctly at five solar angles.

Citation: Journal of Atmospheric and Oceanic Technology 26, 10; 10.1175/2009JTECHA1324.1

Fig. 7.
Fig. 7.

(a) A 512 × 512 pixel frame; (top) sample images and (bottom) images of spectrum power after FFT for (b),(d) cloud free and (c),(e) cloudy skies.

Citation: Journal of Atmospheric and Oceanic Technology 26, 10; 10.1175/2009JTECHA1324.1

Fig. 8.
Fig. 8.

Symmetry of all-sky image: (left) whole gray images and (right) false gray image depending on B/R ratio.

Citation: Journal of Atmospheric and Oceanic Technology 26, 10; 10.1175/2009JTECHA1324.1

Fig. 9.
Fig. 9.

A case of histogram and edge-searching approaches: (left) the histogram made from Fig. 10a (estimated threshold is 1.15) and (right) the highest 20 std devs and their B/R ratios (average ratio is 1.03). The final threshold is 1.09.

Citation: Journal of Atmospheric and Oceanic Technology 26, 10; 10.1175/2009JTECHA1324.1

Fig. 10.
Fig. 10.

Intercomparison of cloud determinations by IM and SM: (top) original image and false color images after (middle) IM and (bottom) SM. White represents cloud and blue represents cloudless. The thresholds of IM are (a) 0.8, (b) 1.25, and (c) 1.09. The threshold of SM is 1.30.

Citation: Journal of Atmospheric and Oceanic Technology 26, 10; 10.1175/2009JTECHA1324.1

Fig. 11.
Fig. 11.

Number difference at 11 cloud fractions from IM – OB and SM – OB. The 4000 images are taken at XiangHe Atmospheric Integrated Observatory of IAP from 2005 to 2007.

Citation: Journal of Atmospheric and Oceanic Technology 26, 10; 10.1175/2009JTECHA1324.1

Table 1.

Visibility vs AOD.

Table 1.
Save