Water Body Detection and Delineation with Landsat TM Data

The aim of this project was to determine the accuracy of using simple digital image processing techniques to map riverine water bodies with Landsat 5 TM data. This paper quantifies the classification accuracy of single band density slicing of Landsat 5 TM data to delineate water bodies on riverine floodplains. The results of these analyses are then compared to a 6-band maximum likelihood classification over the same area. The water boundaries delineated by each of these digital classification procedures were compared to water boundaries delineated from colour aerial photography acquired on the same day as the TM data. These comparisons show that Landsat TM data can be used to map water bodies accurately. Density slicing of the single mid-infrared band 5 proved as successful as multispectral classification achieving an overall accuracy of 96.9%, a producer's accuracy for water bodies of 81.7% and a user's accuracy for water bodies of 64.5%. Introduction Accurate information on the extent of water bodies is important for flood prediction, monitoring, and relief (Smith, 1997; Tholey et al., 1997; Baumann, 1999); production of wetland inventories (Bennett, 1987; Johnston and Barson, 1993; Blackman et al., 1995; Shaikh et al., 1998; Phim et al., 1999); and the evaluation of water resources (Morse et al., 1990; Manavalan et al., 1993). Often this information is difficult to produce using traditional survey techniques because water bodies can be fast moving as in floods, tides, and storm surges or may be inaccessible. Remotely sensed data provide a means of delineating water boundaries over a large area at a given point in time. To capture fast moving hydrological features, the data need to be either of a high temporal resolution or in a substantial archive to cover a range of hydrological conditions. Landsat MSS and TM provide high spatial resolution data at 16-day intervals over a long archival history, exceeding 25 years in most locations. The long archive period and repetitive capture make the data useful for mapping water bodies at a regional scale over a range of hydrological conditions. Since Landsat data became available in 1972, they have been used to map water extent. Smith (1997) cites several early studies where Landsat MSS data, in particular Band 7, were used to distinguish water bodies from surrounding dry soil or vegetation. Comparison with aerial photography gave error estimates of less than 5 percent. Bennett (1987) used density slicing of Landsat MSS band 7 to map water bodies to the west of Griffith, New South Wales, Australia (NsW). He compared the area of MSS-derived water bodies with that derived from digitized aerial photography and found that the MSS data underestimated the area of water by around 40 percent. I School of Science and Technology, Charles Sturt University, Locked Bag 588, Wagga Wagga, NSW 2678, Australia ([email protected]). Johnston and Barson (1993) evaluated the usefulness of Landsat TM imagery for mapping lakelpond wetland extent in western Victoria and riverine wetland extent in central Victoria. The satellite information was compared to manually mapped ground truth. They found that simple density slicing of the TM5 (mid-infrared) successfully detected the lakelpond wetland areas achieving a classification accuracy of 95 percent but failed to map the riverine wetlands adequately. The failure of the technique with riverine wetlands was attributed to the narrow width of the oxbow lakes and a poor timing of data selection. Many other studies have reported the successful use of density slicing of Landsat TM or MSS data to delineate water bodies; however, no quantitative accuracy assessments were conducted. Manavalan et al. (1993) used Landsat TM band 4 to map the extent of the Bhadra Reservoir, India. Overton (1997) used density slicing of Landsat TM band 5 and a high flood spatial mask to map water bodies on the Murray River between Blanchetown and Wentworth, South Australia. Shaikh et al. (1997) timed the acquisition of Landsat MSS band 7 to coincide with selected significant hydrological events to determine wetland inundation for the Cumbung Swamp, NSw, Australia. Baumann (1999) used band 4 of Landsat TM to map flood extent on the Mississippi River but experienced problems separating water from certain urban features without the inclusion of an additional band. Other studies have successfully used both supervised and unsupervised multispectral classification of optical remote sensing data to delineate water boundaries (Manavalan et al., 1993; Lee and Lunetta, 1995; Blackman et al., 1995; Kingsford et al., 1997; Brady et al., 1999). Kingsford et al. (1997) used supervised maximum-likelihood classification of Landsat MSS data to map wetlands over a part of the Murray-Darling Basin. Checking of 200 random water bodies with aerial photography gave an accuracy of 90 percent. They then extended the project to the entire basin using 59 images and an unsupervised classification procedure. Both single-band density slicing and multispectral classification (supervised and unsupervised) have been used to map water boundaries. Researchers typically have relied upon visual comparisons between the classification and the raw data to estimate accuracy. Few studies have used a formal accuracy assessment procedure. Where statistical accuracy assessment procedures have been implemented, they have either compared the area or number of wetlands identified by ground truth with that determined from the satellite image, or they Photogrammetric Engineering & Remote Sensing Vol. 66, No. 12, December 2000, pp. 1461-1467. 0099-1112I0OI6612-1461$3.00/0 O 2000 American Society for Photogrammetry