Elsevier

Remote Sensing of Environment

Volume 204, January 2018, Pages 672-689
Remote Sensing of Environment

Automatic near real-time flood detection using Suomi-NPP/VIIRS data

https://doi.org/10.1016/j.rse.2017.09.032Get rights and content

Highlights

  • A comprehensive introduction to the SNPP/VIIRS flood detection software is presented.

  • Flood detection considers vegetation/bare soil and snow/ice backgrounds.

  • Extensive applications and evaluation of the near real-time flood products are discussed.

Abstract

Near real-time satellite-derived flood maps are invaluable to river forecasters and decision-makers for disaster monitoring and relief efforts. With support from the JPSS (Joint Polar Satellite System) Proving Ground and Risk Reduction (PGRR) Program, flood detection software has been developed using Suomi-NPP/VIIRS (Suomi National Polar-orbiting Partnership/Visible Infrared Imaging Radiometer Suite) imagery to automatically generate near real-time flood maps for National Weather Service (NWS) River Forecast Centers (RFC) in the USA. The software, which is called VIIRS NOAA GMU Flood Version 1.0 (hereafter referred to as VNG Flood V1.0), consists of a series of algorithms that include water detection, cloud shadow removal, terrain shadow removal, minor flood detection, water fraction retrieval, and floodwater determination. The software is designed for flood detection in any land region between 80°S and 80°N, and it has been running routinely with direct broadcast SNPP/VIIRS data at the Space Science and Engineering Center at the University of Wisconsin-Madison (UW/SSEC) and the Geographic Information Network of Alaska at the University of Alaska-Fairbanks (UAF/GINA) since 2014. Near real-time flood maps are distributed via the Unidata Local Data Manager (LDM), reviewed by river forecasters in AWIPS-II (the second generation of the Advanced Weather Interactive Processing System) and applied in flood operations. Initial feedback from operational forecasters on the product accuracy and performance has been largely positive. The software capability has also been extended to areas outside of the USA via a case-driven mode to detect major floods all over the world. Offline evaluation efforts include the visual inspection of over 10,000 VIIRS false-color composite images, an inter-comparison with MODIS automatic flood products and a quantitative validation using Landsat imagery. The steady performance from the 3-year routine process and the promising evaluation results indicate that VNG Flood V1.0 has a high feasibility for flood detection at the product level.

Introduction

As the costliest natural disasters worldwide, most climate change forecasts predict that floods will become increasingly frequent (Milly et al., 2002, Hirabayashi et al., 2008, Lehner et al., 2006). At high latitudes, floods are caused by ice jams and snow melt during almost every break-up season. Floods caused by intense rainfall also threaten the safety of human lives and property. Near real-time satellite-derived flood maps are invaluable to river forecasters and decision-makers for disaster monitoring and relief efforts.

Flood detection has a history in satellite remote sensing that dates back to the 1970s. Imagery from the NOAA (National Oceanic and Atmospheric Administration) VHRR (Very High Resolution Radiometer) and AVHRR (Advanced Very High Resolution Radiometer) served as the main data sources for flood/standing water detection prior to the development of the MODIS (Moderate Resolution Imaging Spectroradiometer) system. Many case studies have been conducted to analyze severe flood events all over the world. These studies laid a foundation for the methods and approaches of flood detection with coarse-to-moderate-resolution satellite data (Wiesnet et al., 1974, Barton and Bathols, 1989, Ali, 1989, Sheng and Xiao, 1994, Sheng et al., 1998, Sheng and Gong, 2001). With coarse 1-km spatial resolutions, however, VHRR and AVHRR data could only show the macro flood distributions of select major floods and failed to address any inundation details. To resolve this issue, Landsat imagery with a 30-m spatial resolution is widely used as an alternative in flood detection, disaster assessment and flood pattern analysis (Gupta and Bodechtel, 1982, Gupta and Banerji, 1985, Wang et al., 2002, Mueller et al., 2016, Fisher et al., 2016, Tulbure et al., 2016). Although VHRR, AVHRR and Landsat imagery play effective roles in flood mapping, the flood detection capabilities of these optical sensors can be severely affected by cloud cover during flood periods. To derive flood information under cloud cover, radar remote sensing satellites and imaging systems such as Radarsat, SAR, TerraSAR-X and Sentinel-1 are becoming more popular in flood monitoring and analysis. Their high spatial resolution and capability to penetrate cloud cover make radar data very popular in hydrological fields for multiple-scale flood mapping, flood management and disaster relief (Brakenridge et al., 1994, Matgen et al., 2007, Schumann et al., 2007, Martinis et al., 2009, Matgen et al., 2011, Pulvirenti et al., 2011, Martinis et al., 2013).

Although Landsat and radar imagery have excellent capabilities for flood mapping, the narrow swath widths and long revisit periods of their sensors are major drawbacks. Because most floods are short-term events, it is not realistic to completely rely upon these images for flood mapping and management purposes. In comparison, moderate-spatial-resolution satellites provide steadier and lower-cost data sources for near real-time flood mapping. After the EOS (Earth Observing System) flagship Terra was launched in 1999, MODIS has gradually become the preferred satellite instrument for flood detection because of its daily global coverage and higher spatial resolution of the visible, near infrared (250 m) and shortwave infrared (500 m) channels compared to the 1-km resolution channels with the AVHRR (Gumley and King, 1995, Brakenridge and Anderson, 2006). Newer algorithms such as the decision-tree approach and the open water likelihood method have used MODIS to more accurately detect flooding and standing water (Sun et al., 2012, Ticehurst et al., 2014, Ticehurst et al., 2015). The continuous observations from MODIS also make it possible to analyze flood inundation dynamics and generate global water masks from multiple-year detected results (Carroll et al., 2009, Andrimont et al., 2012, Huang et al., 2014). In 2011, an experimental global flood detection system using MODIS imagery was released by NASA (National Aeronautics and Space Administration) (http://oas.gsfc.nasa.gov/floodmap). This system processes near real-time MODIS data and generates 1-day, 2-day, 3-day and 14-day composite global flood products for 10° × 10° tiles from the MODIS instrumentation aboard the Terra and Aqua satellites (Brakenridge, 2011). The system also provides systematic datasets with a robust interface to access the products. The multiple-day composition process is applied mainly in order to filter out cloud shadows and terrain shadows, and it produces multiple-day composite flood maps rather than near real-time ones. The problem with the multiple-day composition process is that some real floodwater data may be lost in the composition process, and the process introduces a bias in the experimental MODIS flood maps. Even after the composition process has finished, cloud shadows can persist in the MODIS flood products, especially at high latitudes. More recently, the HAND (height above nearest drainage) algorithm has been applied to MODIS flood detection attempts with a better removal of terrain shadows. The accuracy of MODIS flood products are still susceptible to deep terrain shadows that cannot be filtered either through multiple-day compositions or the HAND algorithm (Brakenridge, 2011, Liu et al., 2016).

With the launch of the Suomi-NPP in 2012, the VIIRS sensor has exhibited many advantages over MODIS data in environmental and natural disaster monitoring and analysis. SNPP/VIIRS imagery has a moderate spatial resolution of 375 m in the shortwave IR bands, a swath coverage width of 3000 km, and a relatively constant resolution across the scan. These new features make SNPP/VIIRS data an excellent source for near real-time flood detection. With the support of the JPSS/PGRR program since 2013, VNG Flood V1.0 has been developed using SNPP/VIIRS imagery to derive near real-time flood maps for the National Weather Service (NWS) River Forecast Centers (RFC) in the USA. A series of algorithms have been developed in the software, including those for water detection, cloud shadow removal, terrain shadow removal, minor flood detection, water fraction retrieval, and floodwater determination. The successful development of the cloud shadow and terrain shadow removal algorithms promises consistent results and makes the detection of near real-time flooding feasible and operational using moderate-resolution satellite data. This paper presents a comprehensive introduction to the software, describes the required datasets, introduces the algorithms, presents the results, and concludes with a summary discussion.

Section snippets

Data used

The main datasets used for flood detection with the VIIRS imagery are the SNPP/VIIRS SDR (sensor data record) data in imager bands 1 (600–680 nm), 2 (850–880 nm), 3 (1610 nm) and 5 (1050–1240 nm) with nominal resolutions of 375 m and I-band terrain-corrected geolocation data, which includes longitude, latitude, solar zenith angles, solar azimuth angles, sensor zenith angles and sensor azimuth angles (GITCO). The SNPP/VIIRS 750-m resolution cloud mask intermediate product (IICMO) and M-band

Physical basis

Water detection with vegetation and bare land background conditions using optical satellite data is primarily based on the spectral differences between water features and other land cover types in the visible (Vis, VIIRS I1 band: 600–680 nm), near infrared (NIR, VIIRS I2 band: 850–880 nm) and shortwave infrared (SWIR, VIIRS I3 band: 1580–1640 nm) channels (Wiesnet et al., 1974, Barton and Bathols, 1989, Sheng and Xiao, 1994). As shown in Fig. 1, water has a higher reflectance in the Vis channel

Applications

During a demonstration project operated by the JPSS PGRR Program since 2014, the developed VNG Flood V1.0 has been running routinely for five river forecast centers in the USA at two locations that process VIIRS direct broadcast data in near real-time: the Space Science and Engineering Center at the University of Wisconsin-Madison (SSEC/UW-Madison) and the Geographic Information Network of Alaska at the University of Alaska-Fairbanks (GINA/UAF). The flood maps are distributed via the Unidata

Discussion

With the support from JPSS/PGRR Program, VNG Flood V1.0 has been developed for automatic near real-time flood detection using SNPP/VIIRS data. Algorithms include water detection, cloud shadow removal, terrain shadow removal, minor flood detection, water fraction retrieval, and flood determination. With a demonstration project initialized by JPSS/PGRR Program, the software has been running routinely using direct broadcast VIIRS data in near real-time flood detection for five river forecast

Conclusion

This study presents a comprehensive introduction to VNG Flood V1.0, and can be summarized as follows:

  • 1.

    The VIIRS NOAA/GMU Flood Version 1.0 software has been developed for automatic near real-time flood detection using SNPP/VIIRS imagery. Floods are divided into two types: supra-veg/bare soil floods and supra-snow/ice floods. A series of algorithms, including water detection, cloud shadow removal, terrain shadow removal, minor flood detection, water fraction retrieval, and floodwater

Acknowledgement

This work was supported by NOAA grant #NA14NES4400007. We thank the five river forecast centers: NCRFC, MBRFC, NERFC, WGRFC and APRFC in the USA for the great efforts they have made to improve this software. The manuscript's contents are solely the opinions of the authors and do not constitute a statement of policy, decision, or position on behalf of NOAA or the U. S.A.

References (55)

  • P. Matgen et al.

    Towards an automated SAR-based flood monitoring system: lessons learned from two case studies

    Phys. Chem. Earth

    (2011)
  • N. Mueller et al.

    Water observations from space: Mapping surface water from 25 years of Landsat imagery across Australia

    Remote Sens. Environ.

    (2016)
  • B. Rabus et al.

    The shuttle radar topography mission- a new class of digital elevation models acquired by spaceborne radar

    Int. Arch. Photogramm. Remote. Sens. Spat. Inf. Sci.

    (2003)
  • W. Schroeder et al.

    Validation of GOES and MODIS active fire detection products using ASTER and ETM + data

    Remote Sens. Environ.

    (2008)
  • M.G. Tulbure et al.

    Surface water extent dynamics from three decades of seasonally continuous Landsat time series at subcontinental scale in a semi-arid region

    Remote Sens. Environ.

    (2016)
  • G. Xian et al.

    Updating the 2001 National Land Cover Database land cover classification to 2006 by using Landsat imagery change detection methods

    Remote Sens. Environ.

    (2009)
  • A. Ali

    Study of river flood hydrology in Bangladesh with AVHRR data

    Int. J. Remote Sens.

    (1989)
  • R.D. Andrimont et al.

    8 years water bodies monitoring analysis using MODIS over the African continent

    EGU General Assembly

    (2012)
  • G.R. Brakenridge

    Technical Description, DFO-GSFC Surface Water Mapping Algorithm

  • G.R. Brakenridge et al.

    MODIS-based flood detection, mapping and measurement: the potential for operational hydrological applications

    Earth Environ. Sci.

    (2006)
  • G.R. Brakenridge et al.

    Radar remote sensing aids study of the great flood of 1993

    Eos Trans. AGU

    (1994)
  • M. Carroll et al.

    A new global raster water mask at 250 meter resolution

    International Journal of Digital Earth

    (2009)
  • L.E. Gumley et al.

    Remote sensing of flooding in the U.S. upper midwest during the summer of 1993

    Bull. Am. Meteorol. Soc.

    (1995)
  • Y. Hirabayashi et al.

    Global projections of changing risks of floods and droughts in a changing climate

    Hydrol. Sci. J.

    (2008)
  • K.D. Hutchison et al.

    A Geometry-Based Approach to Identifying Cloud Shadows in the VIIRS Cloud Mask Algorithm for NPOESS

    J. Atmos. Ocean. Technol.

    (2009)
  • A.M. Johansson et al.

    Adaptive classification of supraglacial lakes on the West Greenland ice sheet

    IEEE Journal of Select Topics Appl. Earth Obs. Remote. Sens.

    (2013)
  • K.V. Khlopenkov et al.

    New cloud, snow, and cloud shadow detection scheme for historical 1-km AVHHR data over Canada

    J. Atmos. Ocean. Technol.

    (2007)
  • Cited by (67)

    • Rapid onset shocks: the importance of understanding impacts of flood disasters

      2023, Hydro-Meteorological Hazards, Risks, and Disasters
    • A downscaling model for derivation of 3-D flood products from VIIRS imagery and SRTM/DEM

      2022, ISPRS Journal of Photogrammetry and Remote Sensing
      Citation Excerpt :

      One mistake is from residual cloud shadows. About 5% cloud shadows that are not removed from water pixels are mistakenly classified as floodwater (Li et al., 2017). If these false floodwater are in mountainous regions, the downscaling process will derive abnormal water surface levels (e.g. above 30 m) and will be assumed as false floodwater during the process.

    View all citing articles on Scopus
    View full text