Elsevier

Remote Sensing of Environment

Volume 118, 15 March 2012, Pages 83-94
Remote Sensing of Environment

Object-based cloud and cloud shadow detection in Landsat imagery

https://doi.org/10.1016/j.rse.2011.10.028Get rights and content

Abstract

A new method called Fmask (Function of mask) for cloud and cloud shadow detection in Landsat imagery is provided. Landsat Top of Atmosphere (TOA) reflectance and Brightness Temperature (BT) are used as inputs. Fmask first uses rules based on cloud physical properties to separate Potential Cloud Pixels (PCPs) and clear-sky pixels. Next, a normalized temperature probability, spectral variability probability, and brightness probability are combined to produce a probability mask for clouds over land and water separately. Then, the PCPs and the cloud probability mask are used together to derive the potential cloud layer. The darkening effect of the cloud shadows in the Near Infrared (NIR) Band is used to generate a potential shadow layer by applying the flood-fill transformation. Subsequently, 3D cloud objects are determined via segmentation of the potential cloud layer and assumption of a constant temperature lapse rate within each cloud object. The view angle of the satellite sensor and the illuminating angle are used to predict possible cloud shadow locations and select the one that has the maximum similarity with the potential cloud shadow mask. If the scene has snow, a snow mask is also produced. For a globally distributed set of reference data, the average Fmask overall cloud accuracy is as high as 96.4%. The goal is development of a cloud and cloud shadow detection algorithm suitable for routine usage with Landsat images.

Highlights

► A new method for automated cloud and cloud shadow detection in Landsat images. ► It is the result of combining past approaches and a new object-based approach. ► It is an improvement over the traditional ACCA cloud algorithm. ► The average Fmask cloud overall accuracy is 96.4%.

Introduction

The long history of Landsat data is one of the most valuable datasets available for studying land cover change and human influences on the land surface (Cohen et al., 1998, Coiner, 1980, Coppin and Bauer, 1994, Seto et al., 2002), especially since the first Thematic Mapper (TM) sensor was launched in 1982, which provided higher spatial resolution and more spectral bands. However, many of the Landsat images are inevitably covered by cloud, especially in the tropics (Asner, 2001). The International Satellite Cloud Climatology Project-Flux Data (ISCCP-FD) data set estimates the global annual mean cloud cover is approximately 66% (Zhang et al., 2004). The presence of clouds and their shadows complicates the use of data in the optical domain from earth observation satellites. The brightening effect of the clouds and the darkening effect of cloud shadows influence many kinds of data analyses, causing problems for many remote sensing activities, including inaccurate atmospheric correction, biased estimation of Normalized Difference Vegetation Index (NDVI) values, mistakes in land cover classification, and false detection of land cover change. Therefore, clouds and cloud shadows are significant sources of noise in the Landsat data, and their detection is an initial step in most analyses (Arvidson et al., 2001, Irish, 2000, Simpson and Stitt, 1998). Generally, clouds can be divided into two categories: thick opaque clouds and thin semitransparent clouds. The thick opaque clouds are relatively easier to identify because of their high reflectance in the visible bands. The identification of thin semitransparent clouds is difficult as their signal includes both clouds and the surface underneath (Gao and Kaufman, 1995, Gao et al., 1998, Gao et al., 2002).

Due to the high spectral variability of clouds, cloud shadows, and the Earth's surface, automated accurate separation of clouds and cloud shadows from normally illuminated surface conditions is difficult. Intuitively, it seems that clouds and cloud shadows are easily separable from clear-sky measurements, as clouds are generally white, bright, and cold compared to the Earth's surface, while cloud shadows are usually dark. Nevertheless, there are clouds that are not white, bright, or cold and cloud shadows even brighter than the average surface reflectance. Part of the difficulty arises from the wide range of reflectances and temperatures observed on the surface (Irish, 2000). One common approach is to screen clouds and cloud shadows manually. However, this approach is time consuming and will limit efforts to mine the Landsat archive to study the history of the Earth's surface.

Over the years, a number of methods were developed for cloud identification. However, most of them are designed for moderate spatial resolution sensors such as Advanced Very High Resolution Radiometer (AVHRR) and Moderate Resolution Imaging Spectroradiometer (MODIS). These sensors are usually equipped with more than one thermal band, or with water vapor/CO2 absorption bands, both of which are useful for thin semitransparent cloud detection (Ackerman et al., 1998, Derrien et al., 1993, Saunders and Kriebel, 1998). For high spatial resolution sensors like Landsat, with only one thermal band and 6 optical bands placed in atmospheric windows, accurate cloud identification is difficult. And, cloud shadow identification is even more difficult. Clouds cast shadows on any type of land cover. When cloud shadows fall on urban, snow, ice, or bright rocks, they can be very bright compared to the average surface reflectance. Moreover, when the cloud is semitransparent, the darkening effect of the cloud shadow can be subtle, making the cloud shadow hard to detect. Therefore, how to detect clouds, cloud shadows, and especially thin clouds and their shadows in Landsat images is still an important issue in the remote sensing community, particularly as we try to use increasingly automated methods to analyze large volumes of data.

Section snippets

Background

Historically, screening of clouds in Landsat data has been performed by the Automated Cloud Cover Assessment (ACCA) system (Irish, 2000, Irish et al., 2006). By applying a number of spectral filters, and depending heavily on the thermal infrared band, ACCA generally works well for estimating the overall percentage of clouds in each Landsat scene, which was its original purpose. However, it does not provide sufficiently precise locations and boundaries of clouds and their shadows to be useful

The Fmask algorithm

The input data are Top of Atmosphere (TOA) reflectances for Bands 1, 2, 3, 4, 5, 7 and Band 6 Brightness Temperature (BT) (Table 1). For Landsat L1T images, Digital Number (DN) values are converted to TOA reflectances and BT (Celsius degree) with the LEDAPS atmosphere correction tool (Masek et al., 2006, Vermote and Saleous, 2007). Then, rules based on cloud and cloud shadow physical properties are used to extract a potential cloud layer and a potential cloud shadow layer. Finally, the

Fmask results

By comparing the results of Fmask with false color composites visually (Fig. 4a, b, c, d), it appears to work well in identifying cloud (yellow), cloud shadow (green), and snow/ice (cyan). Fig. 4a is one of the Sub-tropical South images with well-behaved clouds and cloud shadows over highly vegetated areas. Fmask showed its strong ability in identifying this kind of clouds (including some of the thin clouds) and their shadows. On the other hand, Fig. 4b is a Sub-tropical North image with large

Discussion and conclusion

The Fmask algorithm effectively finds clouds and cloud shadows, which helps with a wide assortment of remote sensing activities. The goal is to provide an automated method for screening clouds and their shadows such that time series of Landsat images can be easily compiled. The need for effective cloud and shadow screening has grown tremendously for two primary reasons. First, the Landsat L1T format now provides accurate enough registration of images that they can be compiled into a time series

Acknowledgments

We gratefully acknowledge USGS and NASA leadership and support of the Landsat Science Team. We also thank Richard Irish (formerly of SSAI/NASA-GSFC) for providing the Landsat-7 scenes and Pat Scaramuzza (USGS/SGT) for providing the cloud/cloud shadow reference masks and ACCA cloud accuracies for the same dataset. Finally, we would like to thank the three anonymous reviewers who provided valuable comments which greatly improved the quality of the manuscripts.

References (36)

  • J.C. Coiner

    Using Landsat to monitor changes in vegetation cover induced by desertification processes

    Symposium on Remote Sensing of the Environment

    (1980)
  • P.R. Coppin et al.

    Processing of multitemporal Landsat TM imagery to optimize extraction of forest cover change features

    IEEE Transactions on Geoscience and Remote Sensing

    (1994)
  • B.C. Gao et al.

    Selection of the 1.375-μm MODIS channel for remote sensing of cirrus clouds and stratospheric aerosols from space

    American Meteorological Society

    (1995)
  • B.C. Gao et al.

    Correction of thin cirrus path radiance in the 0.4–1.0 μm spectral region using the sensitive 1.375 μmc cirrus detecting channel

    Journal of Geophysical Research

    (1998)
  • B.C. Gao et al.

    An algorithm using visible and 1.38-μm channels to retrieve cirrus cloud reflectances from aircraft and satellite data

    IEEE Transactions on Geoscience and Remote Sensing

    (2002)
  • L. Gomez-Chova et al.

    Cloud-screening algorithm for ENVISAT/MERIS multispectral images

    Geoscience and Remote Sensing

    (2007)
  • D.K. Hall et al.

    Algorithm theoretical basis document (ATBD) for the MODIS snow and sea ice-mapping algorithms

    (2001)
  • D.L. Hartmann

    Global physical climatology

    (1994)
  • Cited by (1641)

    • Remote sensing image cloud detection using a shallow convolutional neural network

      2024, ISPRS Journal of Photogrammetry and Remote Sensing
    • Mapping cropland abandonment and distinguishing from intentional afforestation with Landsat time series

      2024, International Journal of Applied Earth Observation and Geoinformation
    View all citing articles on Scopus
    View full text