JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION Vol. 49, No. 2
AMERICAN WATER RESOURCES ASSOCIATION
April 2013
HYDROGRAPHY CHANGE DETECTION: THE USEFULNESS OF SURFACE CHANNELS DERIVED FROM LIDAR DEMS FOR UPDATING MAPPED HYDROGRAPHY1
Sandra K. Poppenga, Dean B. Gesch, and Bruce B. Worstell2
ABSTRACT: The 1:24,000-scale high-resolution National Hydrography Dataset (NHD) mapped hydrography flow lines require regular updating because land surface conditions that affect surface channel drainage change over time. Historically, NHD flow lines were created by digitizing surface water information from aerial photography and paper maps. Using these same methods to update nationwide NHD flow lines is costly and inefficient; furthermore, these methods result in hydrography that lacks the horizontal and vertical accuracy needed for fully integrated datasets useful for mapping and scientific investigations. Effective methods for improving mapped hydrography employ change detection analysis of surface channels derived from light detection and ranging (LiDAR) digital elevation models (DEMs) and NHD flow lines. In this article, we describe the usefulness of surface channels derived from LiDAR DEMs for hydrography change detection to derive spatially accurate and time-relevant mapped hydrography. The methods employ analyses of horizontal and vertical differences between LiDAR-derived surface channels and NHD flow lines to define candidate locations of hydrography change. These methods alleviate the need to analyze and update the nationwide NHD for time relevant hydrography, and provide an avenue for updating the dataset where change has occurred. (KEY TERMS: LiDAR DEMs; LiDAR surface channels; National Hydrography Dataset; hydrography change detection; surface water; hydrography; remote sensing; geospatial analysis.) Poppenga, Sandra K., Dean B. Gesch, and Bruce B. Worstell, 2013. Hydrography Change Detection: The Usefulness of Surface Channels Derived from LiDAR DEMs for Updating Mapped Hydrography. Journal of the American Water Resources Association (JAWRA) 49(2):371-389. DOI: 10.1111 ⁄ jawr.12027 INTRODUCTION
The United States Geological Survey (USGS) National Hydrography Dataset (NHD) 1:24,000-scale flow lines (Kelmelis, 2003; Kelmelis et al., 2003; Simley, 2006) need to be improved to reflect current topographic conditions (Colson et al., 2006; Sheng et al., 2007; Kloiber and Hinz, 2008; Kaiser et al., 2010; Ducey et al., 2012; Quinn and Lo´pez-Torrijos, 2012).
These mapped hydrography updates are needed because of temporal changes in surface channels. The USGS NHD 1:24,000-scale dataset, also known as high-resolution NHD, is a digital vector dataset containing hydrographic features and is the surface water component of The National Map (Kelmelis et al., 2003). Although vector NHD flow lines are frequently used in geographic information systems (GIS), the tools used for collaborative maintenance of the dataset are quite complex (Kloiber and Hinz,
1 Paper No. JAWRA-12-0013-P of the Journal of the American Water Resources Association (JAWRA). Received January 17, 2012; accepted October 31, 2012. ª 2013 American Water Resources Association. This article is a U.S. Government work and is in the public domain in the USA. Discussions are open until six months from print publication. 2 Respectively, Geographer, Topographic Science, U.S. Geological Survey (USGS), Earth Resources Observation and Science (EROS) Center, 47914 252nd Street, Sioux Falls, South Dakota 57198; Research Physical Scientist, Topographic Science, USGS EROS, Sioux Falls, South Dakota; and Senior Scientist, Topographic Science, Stinger Ghaffarian Technologies (SGT), Inc. contractor for the USGS EROS, Sioux Falls, South Dakota (E-Mail ⁄ Poppenga:
[email protected]).
JOURNAL
OF THE
AMERICAN WATER RESOURCES ASSOCIATION
371
JAWRA
POPPENGA, GESCH, 2008) and require upfront labor cost estimates associated with maintenance edits (Kaiser et al., 2010). Thus, updating a nationwide dataset as complex as the NHD is not a trivial task. Historically, hydrography data were derived by digitizing surface water features from aerial photography and paper maps (Guptill, 1979, 1983; Stephens et al., 1980; Marks et al., 1984; Usery, 2012). Duplicating those laborious efforts to obtain nationwide time-relevant hydrography is inefficient and cost-prohibitive (Marks et al., 1984; Colson et al., 2006). With the increasing availability of computing power, digital methods for updating mapped hydrography can employ change detection analysis using bare earth digital elevation models (DEMs). Because the shape of the land exerts strong control over the collection and flow of surface water, changes to the topography can have a significant effect on local drainage conditions (Hjerdt et al., 2004; Gesch, 2006). Therefore, terrain analysis of elevation data, including light detection and ranging (LiDAR) DEMs, is frequently used to extract surface water information (O’Callaghan and Mark, 1984; Jenson and Domingue, 1988; Jenson, 1991; Moore et al., 1991; Tarboton et al., 1991; Garbrecht and Martz, 1997; Tarboton, 1997; Maidment, 2002; Liu et al., 2005; Colson et al., 2006; Jones et al., 2008; Stoker et al., 2008; Ho¨fle et al., 2009; Jenkins and Frazier, 2010; Li and Wong, 2010; Poppenga et al., 2010, 2012; Ducey et al., 2012; Quinn and Lo´pez-Torrijos, 2012). The extracted LiDAR-derived features are useful for updating NHD-mapped hydrography flow lines. In this article, we illustrate methods for improving NHD mapped hydrography that employ change detection analysis of LiDAR-derived surface channels and NHD hydrography flow lines to identify anomalies that exceed 12.2 m, a National Map Accuracy Standard (NMAS) guideline for 1:24,000-scale maps (USGS, 1999; National Digital Elevation Program, 2004; Maune et al., 2007b). Anomalies that exceed that informational metric are spatially validated by sampling and quantifying elevation values of both the LiDAR-derived surface channel and mapped hydrography flow lines. In other words, LiDAR surface channels that deviate in excess of 12.2 m horizontally from mapped hydrography flow lines are locations of potential surface channel changes. These locations are validated by measuring the vertical elevation differences between the LiDAR surface channels and mapped hydrography flow lines. These methods are beneficial for updating NHD because only the locations suspected of hydrography change will need to be reviewed for spatial accuracy and currency, rather than an entire hydrography dataset. Even though surface flow alteration studies have been well documented in the scientific literature (King and Tennyson, 1984; Brown and Bauer, 2009; JAWRA
AND
WORSTELL
Arrigoni et al., 2010; Carlisle et al., 2010, 2011), the spatial accuracy, or the spatial component, of elevation-derived surface channels has not been emphasized in terrain analysis methods.
Background The accuracy of surface water information is essential for many applications (Ho¨fle et al., 2009; Petroselli, 2012), including habitat descriptions and river restoration (Jones, 2006), water land boundary delineations (Mandlburger and Brockmann, 2001), monitoring of river corridors for natural hazard management (French, 2003; Bru¨gelmann and Bollweg, 2004; Hollaus et al., 2005), sediment transport modeling (Carrivick et al., 2010), distribution of landslide activity (Passalacqua et al., 2010), wetland dynamics (Jenkins and Frazier, 2010), estimation of gully erosion and depth of depressions (Perroy et al., 2010; Zandbergen, 2010), geomorphological change of floodplains (Thoma et al., 2005; Jones et al., 2007; Ho¨fle et al., 2009), and surface water mapping (Heine et al., 2004; Liu et al., 2005; Ho¨fle et al., 2009; Li and Wong, 2010; Poppenga et al., 2010, 2012; Ducey et al., 2012; Lo´pez-Torrijos et al., 2012). Therefore, terrain analysis using LiDAR DEMs has become increasingly important for surface water applications because of the spatial resolution and vertical accuracy that is essential for time-relevant mapped hydrography. Several hydrologic and hydrographic models, created over a decade ago, employ terrain analysis methods to coarser (30 m) DEMs to automate the delineation of surface channels (O’Callaghan and Mark, 1984; Jenson and Domingue, 1988; Jenson, 1991; Tarboton et al., 1991; Garbrecht and Martz, 1997; Tarboton, 1997; Maidment, 2002). Because coarser (30 m) DEMs contain less detailed elevation information than LiDAR DEMs, surface channel networks can be derived that consistently flow downstream. However, applying the same hydrologic analyses to the fine spatial resolution of LiDAR DEMs does not generate the same results. For example, geographic features such as bridges, roads, or other elevated surfaces over conduits are more prominent in LiDAR DEMs. Thus, elevations above conduits will function like dams, so a surface channel network from terrain modeling would erroneously indicate that the water cannot pass through the conduits (Maune et al., 2007a; Poppenga et al., 2010, 2012). This is problematic for commonly used ESRI GIS hydrology tools that fill depressions to derive overland flow direction. Many standard GIS hydrology tools do not analyze the next steepest downslope neighbor as an underground conduit, so hydro-enforcement is not implemented (Poppenga et al., 2010, 2012). 372
JOURNAL
OF THE
AMERICAN WATER RESOURCES ASSOCIATION
HYDROGRAPHY CHANGE DETECTION: THE USEFULNESS
OF
SURFACE CHANNELS DERIVED
Arc Hydro tools (Maidment, 2002) contain options to burn in stream networks to raster elevation data by lowering the elevation of the stream network cells (Maidment, 1996). This process changes elevation values in burned locations and alters the integrity of the DEM. Furthermore, using NHD flow lines that need to be updated to burn stream networks into newly acquired LiDAR DEMs may cause integration problems between LiDAR DEMs and mapped hydrography. Considering that LiDAR DEMs contain current topographic and hydrographic information, burning stream networks into LiDAR DEMs does not solve the issue of updating the NHD and does not solve the problem of elevations above conduits functioning like dams. Tarboton’s (1997) method, termed D-¥ (D-infinity), uses triangular facets to identify an infinite number of possible single-direction flow paths, and the threshold is determined by multiple flow direction of one or two downslope cells. The procedure is based on representing flow direction as a single angle taken as the steepest downwards slope on the eight triangular facets centered at each pixel (Tarboton, 1997; Seibert and McGlynn, 2007; Nardi et al., 2008; Li and Wong, 2010). According to Tarboton, this procedure offers improvements over prior procedures that have restricted flow to eight possible directions (introducing grid bias) or proportioned flow according to slope (introducing unrealistic dispersion). It is questionable whether one or two cells in a 1-m LiDAR DEM are sufficient for identifying the next steepest downward slope pixel beyond elevations over conduits to hydrologically enforce surface channels, leaving the hydrologic barrier issue unresolved. Infinite flow direction methods are valuable for defining downslope direction in complex braided streams, but for updating 1:24,000-scale NHD flow lines, more readily available and commonly used tools are needed. There are several common issues with hydrology models when hydrologic analyses are applied to DEMs. Most methods have been tested in applications of coarser (30 m) DEMs (Mark, 1983; Marks et al., 1984; O’Callaghan and Mark, 1984; Jenson and Domingue, 1988; Tarboton et al., 1988; Jenson, 1991; Martz and Garbrecht, 1992; Band, 1993; Nardi et al., 2008; Pan et al., 2012), but few have been tested on hydro-enforcement of high-resolution LiDAR DEMs. The topographic complexities of detailed LiDAR DEMs cause problems for most hydrologic and hydrographic models because hydro-enforcement is needed to generate fully connected surface channels. For example, in coarser (30 m) DEMs, the elevation data over conduits are usually not as problematic to defining a downstream path as they are in LiDAR DEMs (Maune et al., 2007a; Poppenga et al., 2010, 2012; Ducey et al., 2012). Thus, if currently available JOURNAL
OF THE
AMERICAN WATER RESOURCES ASSOCIATION
FROM
LIDAR DEMS
FOR
UPDATING MAPPED HYDROGRAPHY
hydrologic analyses are applied to LiDAR DEMs, additional techniques are needed to generate consistently connected surface channels. Some models that attempt to solve this problem contain computationally complex mathematical processes that are time-consuming and not readily available in standard GIS software. The models are not easily applicable to LiDAR DEMs and not economically efficient for updating NHD. Additionally, if LiDAR DEMs are to be useful for updating NHD, methods are needed that can be applied to larger geographic areas. The D8 method (Jenson and Domingue, 1988) was implemented at a nationwide scale to generate a hydrologically conditioned version of the 30-m National Elevation Dataset (NED) (Gesch et al., 2002; Gesch, 2007). The resulting Elevation Derivatives for National Applications (EDNA) database activity was proposed to improve the density of 1:100,000-scale information in NHD (Kost and Kelly, 2001; Kelmelis, 2003; Franken, 2004). Considering that the nationwide EDNA database was generated with D8 hydrology tools that were readily available in standard GIS software, the D8 method is a probable framework to efficiently and cost-effectively extract LiDAR-derived surface channels for updating the NHD. The D8 method implemented by Jenson and Domingue (1988) and Jenson (1991) is a contributing area threshold model that assigns a flow direction value to each cell that describes the direction of its steepest downslope neighbor. This is accomplished by first generating a depressionless DEM to route overland flow. Additional selective drainage methods, as proposed by Poppenga et al. (2010, 2012), can be incorporated into the contributing area threshold model to hydrologically enforce depressions located upstream of conduits and extract vector surface channels from LiDAR DEMs. By differencing a LiDAR DEM from a depressionless, or filled, LiDAR DEM, depressions that need hydro-enforcement are detected with assigned parameters. Using a least accumulative cost path analysis, the elevation values over conduits are adjusted to the lowest elevation cell value within the depression allowing a continuous downstream path. D8 hydrologically conditioning processes (Jenson and Domingue, 1988) are then applied to the hydrologically enforced DEM to derive a hydrologically enforced flow direction grid that is converted to vector surface channels. The original LiDAR DEM (unfilled), the hydrologically conditioned LiDAR DEM (filled), and the hydrologically enforced LiDAR DEM are preserved thereby minimally modifying the elevation data. The selective drainage methods, written in Python scripts and executed in an ESRI ArcGIS geoprocessing environment, have been the catalyst for several LiDAR-based research projects, including those attempting to update NHD (Kaiser 373
JAWRA
POPPENGA, GESCH, et al., 2010; Ducey et al., 2012; Lo´pez-Torrijos et al., 2012; Quinn and Lo´pez-Torrijos, 2012). It is well documented in the scientific literature that the high-vertical accuracy and spatial resolution of LiDAR DEMs provide the topographic detail needed for more accurate and current surface water information (Casas et al., 2006; Colson et al., 2006; Jones et al., 2008; Murphy et al., 2008; Stoker et al., 2008; Ho¨fle et al., 2009; Li and Wong, 2010; Poppenga et al., 2010, 2012; Ducey et al., 2012; Lo´pez-Torrijos et al., 2012; Petroselli, 2012; Quinn and Lo´pez-Torrijos, 2012). Li and Wong (2010) state that high-resolution LiDAR DEMs offer superior results in extracting river networks. Murphy et al. (2008) indicate that LiDAR-derived DEM networks were the most accurate representation of field-mapped networks, even more accurate than photo-derived networks. Poppenga et al. (2010) conclude that LiDAR-derived surface channels more closely reflect what is currently occurring on the landscape. Jones et al. (2007) reported that the use of airborne LiDAR data and GIS technology allows the rapid production of detailed geomorphological maps of river valley environments. Liu et al. (2005) demonstrate that LiDARderived DEMs with high-accuracy and high-resolution offer the capability of improving the quality of hydrological features extracted from DEMs. According to the National Research Council (2009), highquality digital mapping is essential to communicating flood hazards to those at risk. Their report, Mapping the Zone, Improving Flood Map Accuracy, concludes that even the most expensive aspect of making more accurate maps — collecting high-accuracy, high-resolution topographic data — yields more benefits than costs.
MATERIALS AND METHODS
Study Areas We tested methods for detecting change in hydrography in both low relief and rugged terrain. The first study area is located in the headwater region of the Slip-up Creek watershed in the low relief plains of eastern South Dakota (Figure 1). Surface waters flow through this agricultural landscape that has numerous culverts and bridges that serve as conduits for overland drainage. Since the 1:24,000-scale NHD flow lines were digitized for this study area, some of the meandering streams in the watershed have been channelized to optimize the amount of arable land or for road construction. These types of surface channel changes are frequent throughout the drainage area, JAWRA
AND
WORSTELL
FIGURE 1. Hydrography Change Detection Was Conducted in the Headwater Region of the Slip-up Creek Watershed in the Low Relief Plains of Eastern South Dakota Where Some Meandering Streams Have Been Channelized. In this study area, the minimum elevation value is 450.45 m and the maximum elevation value is 502.79 m. The area of the study site is 26.65 km2.
and collectively have altered the slope of surface water drainage. Although a potential source of agricultural surface water run-off, Slip-up Creek was considered to be one of several potential water supply sources for the nearby city of Sioux Falls, the largest and fastest growing city in the state (Barari et al., 1989). The second study area is located in the Tehachapi Mountains along Interstate 5 (I-5) in southern California (Figure 2). I-5, also known as the Golden State Freeway, traverses rugged terrain subject to earthquakes, intense weather events, and landslides. These mountainous slopes have been altered within the last 50 years to construct an important commerce corridor that connects metropolitan Los Angeles with the agricultural San Joaquin Valley. Massive amounts of earthen materials were carved out of mountains and deposited into canyons to construct the eight-lane I-5 freeway (Foster, 2003; Scott, 2003), an impressive major public works project that altered 374
JOURNAL
OF THE
AMERICAN WATER RESOURCES ASSOCIATION
HYDROGRAPHY CHANGE DETECTION: THE USEFULNESS
OF
SURFACE CHANNELS DERIVED
FROM
LIDAR DEMS
FOR
UPDATING MAPPED HYDROGRAPHY
approximately 3 m. The vertical accuracy for the Slip-up Creek study area (Figure 1) was 18.5 cm rootmean-square error (RMSE) on open bare earth terrain and 37.0 cm RMSE in vegetative areas (Poppenga et al., 2010). The vertical accuracy as reported by the LiDAR vendor for the southern California study area (Figure 2) was 11.0 cm RMSE on open bare earth terrain and 14.3 cm RMSE for vegetative areas. Ancillary aerial imagery obtained for the Slip-up Creek study area was collected within one month of the LiDAR acquisition. Both LiDAR and aerial imagery acquisitions were a collaborative effort between USGS, City of Sioux Falls and Minnehaha County, South Dakota, and Sanborn Map Company, Inc. Ancillary aerial imagery for the southern California study area was accessed from Bing Maps aerial imagery web mapping service, which is available in licensed ArcGIS products. Mapped 1:24,000-scale hydrography vector flow lines were obtained from the USGS high-resolution NHD (Simley, 2006). The positional accuracy for the high-resolution NHD was compiled to meet the NMAS guideline, which indicates mapped features need to be within 12.2 m of their true location at a 90% confidence level for 1:24,000-scale maps (USGS, 1999; National Digital Elevation Program, 2004; Maune et al., 2007b). FIGURE 2. Hydrography Change Detection Was Conducted in Rugged Terrain Along the I-5 Corridor in Southern California Where Massive Amounts of Earth Materials Were Carved Out of Mountains and Deposited in Canyons to Construct the Freeway. In this study area, the minimum elevation value is 630.21 m and the maximum elevation value is 1,116.70 m. The area of the study site is 5.10 km2.
the landscape considerably and impacted the directional flow of surface water. Topographic changes such as these have been quantified by Gesch (2006) in a national inventory of significant topographic changes in the United States (U.S.); the inventory was based upon seamless multitemporal elevation data that are valuable for detecting hydrography change in high relief terrain.
Data for Hydrography Change Detection Analysis High-resolution LiDAR DEMs were used in both study areas to extract vector LiDAR surface channels. The LiDAR DEMs were obtained from the USGS NED (Gesch et al., 2002; Gesch, 2007), which is the elevation component of The National Map (Kelmelis et al., 2003). The horizontal resolution of the LiDAR DEMs for both study areas was 1 ⁄ 9-arc-second, or JOURNAL
OF THE
AMERICAN WATER RESOURCES ASSOCIATION
Comparable Scale for LiDAR Surface Channels and Mapped Hydrography In order to conduct hydrography change detection for the high-resolution NHD, the scale of LiDAR surface channels should be analogous to 1:24,000-scale mapped hydrography or USGS topographic base maps before applying selective drainage methods (Poppenga et al., 2010, 2012). To obtain comparable scales for LiDAR surface channels, NHD flow lines were analyzed to extract the end points on the uppermost location of the NHD headwater reaches. The NHD headwater points and LiDAR flow direction grid were used to define comparable flow paths downstream. The cells in the LiDAR flow direction raster represent flow to the adjacent cell in the steepest downslope direction (Jenson and Domingue, 1988). Each cell in this raster is coded with a value representing one of eight neighboring cells that has the steepest downward slope. The cell-to-cell flow connectivity of the LiDAR flow direction raster provides the capability to route a flow path across the surface. The uppermost end point of each NHD flow line is used to initiate a starting point in the LiDAR DEM for deriving a downstream path using GIS cost path analysis techniques. This creates a raster path from the NHD 375
JAWRA
POPPENGA, GESCH, headwater points downstream to the edge of the LiDAR DEM. This technique is often called downstream trace or raindrop trace. Using the selective drainage methods (Poppenga et al., 2010, 2012), elevations above culverts and bridges were identified and adjusted to the lowest elevation cell value within depressions that needed hydro-enforcement. The hydrologically enforced LiDAR flow direction grid was converted to vector surface channels representing a high-resolution raster DEM. The LiDAR surface channels were generalized with a maximum offset tolerance of 3 m to derive surface channel representation at a level of detail more comparable to the mapped hydrography. Some headwater end points on the uppermost reaches of NHD flow lines may not be spatially aligned with the LiDAR elevation data. For example, if an NHD headwater point were located on the other side of a LiDAR elevation ridge rather than in the study area watershed, the LiDAR flow direction would spill downstream into the adjacent watershed causing spatial inaccuracies in the derived LiDAR surface channel. In such a case, the NHD headwater point should be repositioned to an appropriate LiDAR raster pixel located inside the study area watershed for the cost path analysis to follow the correct flow direction in the LiDAR DEM. Identifying NHD headwater points that are spatially misaligned in the LiDAR DEM is one step that is necessary to obtain comparable scales between NHD flow lines and LiDAR surface channels.
The Usefulness of LiDAR DEMs for Detecting Changes in Mapped Hydrography Horizontal and Vertical Components of Elevation Data. Both horizontal and vertical components of elevation data were useful for detecting hydrography changes in the study areas. Using LiDAR DEMs, two methods were employed to quantify displacements between the LiDAR surface channels and NHD flow lines: (1) quantifying horizontal offsets between vector LiDAR surface channels and vector NHD flow lines that exceeded a 12.2-m threshold, resulting in horizontal candidate corresponding change pairs, and (2) quantifying vertical elevation differences between candidate corresponding change pairs. The 12.2-m threshold is based on a NMAS guideline that indicates mapped features need to be within 12.2 m of their true location at a 90% confidence level for 1:24,000-scale maps (USGS, 1999; National Digital Elevation Program, 2004; Maune et al., 2007b). Using both the horizontal and vertical components of LiDAR DEMs ensured that locations suspected of hydrography change were not only quanJAWRA
AND
WORSTELL
tified by measuring horizontal differences but also validated by vertical elevation differences. These methods make full use of the three-dimensional nature of LiDAR DEMs. Such quantification is important for ranking the magnitude of hydrography change for individual locations. Because it is not economically feasible to update the entire 1:24,000-scale NHD, horizontal and vertical rankings are important for automated methods to detect locations of hydrography change so that the largest discrepancies can be updated in the NHD to bring currency to the dataset. Horizontal Displacement – Candidates for Corresponding Change Pairs. To identify horizontal displacements in the study areas, the first step was to buffer LiDAR surface channels (channels) and mapped hydrography flow lines (flow lines) by 12.2 m on both sides of the line networks. This step was needed to create reciprocal buffers rather than to measure the horizontal distance between the buffers. The next step was to differentiate between channel or flow line segments that were located inside or outside of their reciprocal buffers. This boundary was defined by the intersection of the line segments and their reciprocal buffers. For example, line segments located within their reciprocal buffers were offset
