Fusion of highresolution DEMs for water flow modeling
 Anna Petrasova^{1, 2}Email authorView ORCID ID profile,
 Helena Mitasova^{1, 2},
 Vaclav Petras^{1, 2} and
 Justyna Jeziorska^{1, 2, 3}
DOI: 10.1186/s4096501700192
© The Author(s) 2017
Received: 14 December 2016
Accepted: 22 February 2017
Published: 6 March 2017
Abstract
Background
New technologies for terrain reconstruction have increased the availability of topographic data at a broad range of resolutions and spatial extents. The existing digital elevation models (DEMs) can now be updated at a low cost in selected study areas with newer, often higher resolution data using unmanned aerial systems (UAS) or terrestrial sensors. However, differences in spatial coverage and levels of detail often create discontinuities along the newly mapped area boundaries and subsequently lead to artifacts in results of DEM analyses or models of landscape processes.
Methods
To generate a seamless updated DEM, we propose a generalized approach to DEM fusion with a smooth transition while preserving important topographic features. The transition is controlled by distancebased weighted averaging along the DEMs’ blending overlap with spatially variable width based on elevation differences.
Results
We demonstrate the method on two case studies exploring the effects of DEM fusion on water flow modeling in the context of precision agriculture. In the first case study, we update a lidarbased DEM with a fused set of two digital surface models (DSMs) derived from imagery acquired by UAS. In the second application, developed for a tangible geospatial interface, we fuse a georeferenced, physical sand model continuously scanned by a Kinect sensor with a lidarbased DEM of the surrounding watershed in order to computationally simulate and test methods for controlling storm water flow.
Conclusions
The results of our experiments demonstrate the importance of seamless, robust fusion for realistic simulation of water flow patterns using multiple highresolution DEMs.
Keywords
UAS UAV Lidar Kinect GRASS GISBackground
Highquality digital elevation models (DEM) provide essential data for research in many scientific disciplines as well as for numerous practical applications. Today, DEMs can be generated by a variety of remote sensing techniques including conventional and StructurefromMotion photogrammetry (SfM), radar interferometry, lidar, or shortrange 3D cameras [1, 2]. The different remote sensing sensors, platforms and DEM reconstruction algorithms result in DEM products with different properties in terms of spatial extent, resolution, accuracy, survey date, and whether they represent digital surface model (DSM) or bare ground. To address inhomogenity of available DEM products, several methods of fusing DEMs have been developed to obtain a complete DEM coverage with improved quality. Fusion approaches vary from simple techniques, such as weighted averaging of input DEMs based on height error maps [3], or terrain derivatives [4, 5], to more complex techniques involving the use of sparse representations [6], frequency domain filtering [7], slopebased Markov random field regularization [8], or kmeans clustering [9].
In these cases the fusion method combines several overlapping DEMs in order to obtain a higher quality DEM with homogenous error characteristics. However, these methods are typically not suitable for cases when one DEM needs to be updated with a newer or a higher resolution one within a specific subregion. Simple merging or averaging the DEMs typically leads to surface discontinuities [10, 11] on the edge of the DEMs, which can lead to artificial landforms and patterns when the resulting DEM is applied for geomorphometry or hydrologic applications.
Several approaches to ensure smooth transition of DEMs have been proposed in the context of global DEM mosaicking and DEM void filling. Gruber et al. [12] fused interferometric synthetic aperture radar (InSAR) DEMs by the means of weighted average based on derived height errors. As the errors are higher at the borders, decreased weights towards the borders allow for smoother transitions between DEMs. A different approach proposed by Reuter et al. [10] creates seam lines along landscape features such as waterways or roads to merge nonoverlapping patches of DEMs with minimum geometric discontinuities. Such seam lines can be automatically identified by morphological image compositing [13], frequently used for mosaicking satellite imagery, but rarely used for DEM mosaicking due to the complexity of the method and possible lack of suitable landscape features for the seam line. Achieving seamless transitions is also crucial when filling missing data in DEMs with a DEM from different source. The transition zone between the auxilary and the main DEM is typically interpolated using inverse distance weighting (IDW) [14], or estimated as a local average of neighboring elevation pixels [15]. However, depending on the complexity of landscape features the transition zone can be much smoother than its surrounding and become a visible artifact.
To ensure seamless transition between DEMs, Robinson et al. [11] proposed blending DEMs using weighted averaging method where weight is a function of distance to transition line between two DEMs. The blending zone in their case followed a single parallel and had constant width. However, many new DEMs derived from Unmanned Aircraft Systems (UAS) cover irregularly shaped areas and can contain gaps due to complex terrain and vegetation. Moreover, submeter resolution DEMs capture microtopography and above ground features not present in lowerresolution DEMs leading to larger elevation differences along the DEMs seam, which require wider blending zone in order to achieve smooth transition. On the other hand, a narrow blending zone is preferable along the seam sections with small elevation differences to preserve the high resolution topographic features captured by the DEMs.
Our aim is therefore to generalize the approach presented by Robinson et al. [11] for applications where an existing DEM is updated with a new DEM within a smaller subregion with irregularly shaped boundaries, and to improve the preservation of terrain shape when fusing DEMs by applying a datadriven, spatially variable blending zone width. We describe the general workflow for the presented method, provide its software implementation (see Additional file 1) in GRASS GIS [16] and demonstrate our method on two use cases.
Methods
Given two rasterbased DEMs, D E M _{A} and D E M _{B}, we describe a method to combine them into a new D E M _{AB}, for the purpose of updating D E M _{B} with D E M _{A}, or replacing parts of D E M _{A} with D E M _{B}. Although not a necassary rule, D E M _{A} is typically more recent, and has higher resolution, but smaller extent comparing to D E M _{B}. The input DEMs must be coregistered into the same coordinate system and aligned and reinterpolated to the same resolution. The selection of the common resolution and reinterpolation method is highly dependent on the local topography and fused DEM application and is beyond the scope of this paper. Depending on the application, large elevation differences along the edge of D E M _{A}, typically representing vegetation or built structures, should be identified and removed, for example by defining a difference threshold.
Throughout the paper, we use the term digital elevation model (DEM) as a generic term for both bare ground and digital surface model (DSM) which includes vegetation and structures.
Weighted linear combination of DEMs
where k is a parameter controlling the steepness of the logistic curve and w∈〈0,1〉.
For elevation surfaces z _{ A }(x,y) and z _{ B }(x,y) represented by rasterbased D E M _{A} and D E M _{B}, the fusion process can be implemented using basic GIS functions, such as raster algebra. We compute the distance raster D=d(x,y) from nearest null cells of D E M _{A} and derive the weight raster W=w(x,y) from D and s using linear Eq. (2) or a nonlinear equation, for example Eq. (3). We select suitable overlap width s based on the elevation differences between D E M _{A} and D E M _{B} along the blending seam. The updated D E M _{AB} is then computed using Eq. (1).
 1.
Compute the distance surface D from nearest null cells of D E M _{A}.
 2.
Compute the absolute value of the difference between D E M _{A} and D E M _{B}, Δ D E M _{ AB }=Δ z(x,y)=D E M _{ A }−D E M _{ B }.
 3.
Using distance D derive a onepixelwide edge of D E M _{A} and along this edge extract cell values from Δ D E M _{ AB } into raster Δ D E M _{ e }.
 4.
Compute surface Δ D E M e′ where each cell has the value of the nearest nonnull cell of Δ D E M _{ e }.
 5.
To avoid discontinuities, smooth surface Δ D E M e′, for example by using movingwindow operation with average statistics.
 6.
Given transition angle α, compute overlap width surface using Eq. (5) as S=Δ D E M e′/ tanα.
 7.
Compute weight surface from Eq. (4) as W=D/S.
 8.
Derive updated elevation surface using Eq. (1).
Most steps, specifically 2, 3, 6, 7, and 8 can be implemented using raster algebra, for example, with module r.mapcalc in GRASS GIS. Modules r.grow.distance can be used for steps 1 and 4, and r.neighbors for step 5.
The spatially variable overlap ensures a gradual smooth transition between the two merged DEMs and the resulting DEM can then be used for surface water flow modeling with minimized edge artifacts.
To demonstrate the importance of smooth fusion for water flow modeling at submeter resolution we use the path sampling technique for solving the shallow water flow continuity equation [17]. This technique allows us to simulate impact of microtopography on distribution of water depth including ponding in microdepressions and dispersed water flow. The method was implemented in GRASS GIS in the module r.sim.water. The module computes overland flow depth or discharge based on steady, spatially distributed rainfall excess, elevation surface gradient, and surface roughness given by Manning’s coefficient.
Results
Updating lidarbased DEM with UASbased DSMs
The amount of simulated rainfall water during steady state captured on plots a, b, c, and d in Fig. 5
Zone  Area  Fused DEM  June DSM  Difference 
(m^{2})  (m ^{3})  (m^{3})  (m^{3})  
b  108.54  3.81  1.91  1.90 (49.8%) 
c  169.65  7.37  21.42  14.05 (190.6%) 
Zone  Area  Fused DEM  Patched DEM  Difference 
a  138.42  1.69  3.62  1.93 (53.3%) 
d  133.29  2.09  4.84  2.75 (56.8%) 
The UAS data were interpolated from the point clouds to 0.3 meter resolution rasters. The details of the data acquisition and processing of the UASbased DEMs are provided by Jeziorska et al. [18]. Lidar data used in this study were collected by the North Carolina Floodplain Mapping Program [19] in 2015 as part of a statewide survey, with average point density of 3 points per square meter and multiple return classified points. We extracted bare earth points and in order to keep consistent resolution with the UAS DSMs we interpolated the DEM at 0.3 meters using regularized spline with tension [20]. The sections of both UASbased DSMs with high vegetation such as trees and shrubs were masked out and the sections covered by crops were removed from the June DSM by masking out areas with elevations 0.3 meters above the ground given by the lidarbased DEM.
We merged the DEMs in two stages. First, we merged the June and October UASbased DSMs, and then the resulting DSM with the lidarbased DEM. We applied the fusion method with the spatially variable overlap width to preserve as much of the microtopography as possible while ensuring smooth transition between the three DEMs. We specified sufficiently low transition angle α=3° (approximately 0.1 meter elevation difference on 2 meters) to achieve smooth blending on the overlap. The average overlap width was then 3.3±1.6 meters and 2.9±2.2 meters for the first and second fusion respectively (Fig. 3). Figure 4 shows two profiles which compare the fused DEM with the source DEMs. The profile of the fused DEM seamlessly transitions from the profile of one DEM to the other. Due to the spatially variable overlap width, the length of the transition zone differs for each profile (3.6 meters and 4.8 meters for profiles 1 and 2 respectively).
We then used the fused DEM to model shallow overland water flow using path sampling method assuming uniform rainfall excess rate of 30 milimeters per hour. We simulated flow at 0.3 meters resolution for 40 minutes until steady state was reached in most of the modeled area. To account for different roughness of the areas with crops, we created a Manning’s coefficient surface of 0.15 for bare earth and 0.3 for areas with crops captured in the June dataset [21].
We compared the flow pattern modeled on the June DSM (Fig. 5 b), and on the UAS and lidarbased DEMs merged with (Fig. 5 d) and without blending of the overlap (Fig. 5 c). Table 1 quantifies and compares the amount of runoff water in four selected plots (Fig. 5) simulated on these different DEMs.
UASbased DSM from June captures high crops, which results in artificial ponding as visible for example in plot c in Table 1. Such artificial accumulation of water can result in underestimating water depth downstream. For example, in plot b we can observe that simulation based on the DSM underestimates the amount of water in a forming rill by 50%. Replacing vegetated areas with elevation data from different sources (Fig. 5 a) provides better approximation of flow, however, if the transition between DEMs is not smooth enough, water can accumulate near the edges of the DEMs, resulting in unrealistic patterns (Fig. 5 b). Plots a and d are examples where the simulation based on the DEMs merged without blending (patched DEM) overestimates the amount of water by more than 50% comparing to the simulation ran on the fused DEM where blending is used. The solution is therefore to smooth the transition using the described approach, allowing for more realistic flow as demonstrated in Fig. 5.
Merging lidar and Kinectbased DEMs
In the second case study we apply presented method to merge a lidarbased DEM with a georeferenced, virtual DEM obtained by scanning a physical, scaled model in Tangible Landscape environment [22]. Tangible Landscape couples a malleable physical model with a digital landscape through a cycle of realtime scanning, analysis, and projection. As users change the physical model it is continuously scanned using Microsoft Kinect v2 sensor into GRASS GIS [16], in which landscape processes are simulated and results are projected back onto the physical model. The boundaries of the physical model and its scale define the spatial extent of our interactions on the landscape. These boundaries often do not match the boundaries of the physical processes, such as water flow, which accumulates within watersheds. The effects of our interventions on the physical model affect the water flow downstream outside of the model’s boundaries, and similarly modeling water flow without considering surface runoff from outside the model results in underestimating the amount of water in the landscape.
In this application, we use the physical model to design runoff control measures to reduce concentrated flow causing gully erosion. To simulate water flow within the entire studied watershed we merge the scanned DEM of the physical model with the lidarbased DEM covering the entire watershed. Smooth fusion is essential for ensuring that the simulated water flows in and out of the physical model. In this case we applied the fusion with fixed overlap width, since this method is very fast thus suitable for realtime interaction with the model and feedback on the flow patterns.
We manually built a physical model from polymerenriched sand of a small section of the area used in the first case study based on the 2015 lidar data at 1 : 420 scale and 4 times vertical exaggeration, to facilitate scanning and interaction. We used projected contours and colorcoded difference of the scanned and real DEM [22, Chapter 4] while building the model to ensure its sufficient accuracy. We then merged the lidarbased and scanned DEMs with fixed overlap width of 15 meters and ran the water flow simulation on the merged DEM over the watershed including the physical model. We modeled the steady state flow assuming uniform rainfall excess rate of 30 mm/hr, and uniform Manning’s coefficient of 0.15. The simulation ran at resolution of 0.85 meter, which is given by the resolution of the scanner multiplied by the model scale.
Discussion
The presented case studies demonstrate the effectivity and utility of the described method to eliminate edge artifacts along the seamline between two DEMs. Despite the relatively precise alignment of the DEMs in the first study (around 20 centimeters of elevation difference at the edge in average), we showed that even small discontinuities affected processes driven by microtopography, such as overland water flow. The proposed blending technique effectively reduced the artifacts along the edges and allowed for more realistic water flow. In comparison with other fusion techniques, our goal was not to derive best terrain representation globally, but only locally along the edges of the DEMs, in order to preserve the original topography of both DEMs. Therefore the proposed method introduces a blending zone with spatially variable width derived from elevation differences, which ensures smooth transition for large edge artifacts and at the same time reduces the smoothing effect when not desired. Although we presented this method in the context of highresolution DEMs, it can be used at any scale. However, we recommend to carefully consider applying the method in cases when the discontinuities at the edge are large and smooth transition would create unrealistic features in the resulting DEM. In such cases, the cause of this misalignment given by the data acquisition technology and processing should be identified and rectified first, for example by the means of coregistration [23], removing nonground components, or specifically in case of UAS, by collecting imagery with suitably distributed control points [24].
The proposed technique is described in the context of rasterbased DEMs, the prevalent representation of DEMs for environmental and geomorphological applications. Since the Eqs. (1) and (2) are general, the method could be adapted for other terrain representations such as point clouds. However, GIS software typically lacks the ability to manipulate large point clouds in a way that would match the efficiency and variability in processing functionality of raster representation [25]. The presented rasterbased solution can be therefore implemented efficiently in any GIS software with raster support.
Conclusions
We presented a fast and effective technique to merge raster DEMs with different spatial extents by blending the DEMs along their overlap using distancebased weighted average. The novel approach based on spatially variable overlap width improves preservation of subtle topographic features of the highresolution DEMs while ensuring smooth transition. The two case studies demonstrated the importance of smooth transition for modeling water flow patterns while capturing the impacts of microtopography or when interacting with physical models using tangible user interface. With the increasingly widespread use of lowcost 3D sensors and UAS platforms the proposed technique becomes highly relevant for researchers and practitioners working with time series of highresolution DEMs.
Abbreviations
 DEM:

digital elevation model
 DSM:

digital surface model
 GIS:

geographic information system
 IDW:

inverse distance weighting
 InSAR:

interferometric synthetic aperture radar
 NC:

North Carolina
 SfM:

structure from motion
 UAS:

unmanned aircraft system
Declarations
Acknowledgements
We would like to thank Thomas Zajkowski from NextGen Air Transportation Group (NGAT) for collecting the imagery data.
Funding
This research received no specific grant from any funding agency in the public, commercial, or notforprofit sectors.
Availability of data and materials
The described method is implemented as GRASS GIS 7 addon r.patch.smooth and is available in GitHub repository https://github.com/petrasovaa/r.patch.smoothunder GNU GPL license. The dataset with the UAS and lidarbased DEMs is available for download at https://geospatial.ncsu.edu/osgeorel/data.htmlunder Creative Commons AttributionShareAlike.
Authors’ contributions
AP developed the methods, case studies, processed data and drafted the manuscript. HM provided critical revisions to the manuscript. JJ processed the UAS data. VP contributed to the development of the method. All authors read and approved the final manuscript.
Competing interests
The authors declare that they have no competing interests.
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
Authors’ Affiliations
References
 Mukherjee S, Joshi PK, Mukherjee S, Ghosh A, Garg RD, Mukhopadhyay A. Evaluation of vertical accuracy of open source Digital Elevation Model (DEM). Int J Appl Earth Observation Geoinformation. 2012; 21(1):205–17.Google Scholar
 Mankoff KD, Russo TA. The Kinect: A lowcost, highresolution, shortrange 3D camera. Earth Surface Processes Landforms. 2013; 38(9):926–36.View ArticleGoogle Scholar
 Roth A, Knopfle W, Strunz G, Lehner M, Reinartz P. Towards a global elevation product: combination of multisource digital elevation models. Int Arch Photogramm Remote Sens Spat Inf Sci; 34(4):675–9.
 Papasaika H, Poli D, Baltsavias E. Fusion of digital elevation models from various data sources. In: Proceedings of the International Conference on Advanced Geographic Information Systems and Web Services, GEOWS 2009. Cancun: International Conference on Advanced Geographic Information Systems & Web Services: 2009. p. 117–122. http://ieeexplore.ieee.org/document/4782702/.
 Tran TA, Raghavan V, Masumoto S, Vinayaraj P, Yonezawa G. A geomorphology based approach for digital elevation model fusion – case study in Danang City, Vietnam. Earth Surface Dyn Discuss. 2014; 2:255–96.View ArticleGoogle Scholar
 Papasaika H, Kokiopoulou E, Baltsavias E, Schindler K, Kressner D. Fusion of digital elevation models using sparse representations. In: Photogrammetric Image Analysis. ISPRS Conference, PIA 2011, Munich, Germany, October 57, 2011. Proceedings. Berlin: Springer Berlin Heidelberg: 2011. p. 171–84. https://link.springer.com/chapter/10.1007/9783642243936_15.
 Karkee M, Steward BL, Aziz SA. Improving quality of public domain digital elevation models through data fusion. Biosyst Eng. 2008; 101(3):293–305.View ArticleGoogle Scholar
 Yue L, Shen H, Yuan Q, Zhang L. Fusion of multiscale DEMs using a regularized superresolution method. Int J Geograph Inform Sci. 2015; 29(12):2095–120.View ArticleGoogle Scholar
 Fuss CE, Berg AA, Lindsay JB. Dem fusion using a modified kmeans clustering algorithm. Int J Digit Earth. 2016; 9(12):1242–55.View ArticleGoogle Scholar
 Reuter HI, Hengl T, Gessler P, Soille P. Preparation of DEMs for Geomorphometric Analysis In: Hengl T, Reuter HI, editors. Geomorphometry: Concepts, Software, Applications. Developments in Soil Science, vol. 33. Amsterdam: Elsevier: 2008. p. 87–120.Google Scholar
 Robinson N, Regetz J, Guralnick RP. EarthEnvDEM90: A nearlyglobal, voidfree, multiscale smoothed, 90m digital elevation model from fused ASTER and SRTM data. ISPRS J Photogrammetry Remote Sensing. 2014; 87:57–67.View ArticleGoogle Scholar
 Gruber A, Wessel B, Martone M, Roth A. The TanDEMX DEM Mosaicking: Fusion of Multiple Acquisitions Using InSAR Quality Parameters. IEEE J Select Topics Appl Earth Observations Remote Sensing. 2016; 9(3):1047–57.View ArticleGoogle Scholar
 Soille P. Morphological Image Compositing. IEEE Trans Pattern Anal Mach Intell. 2006; 28(5):673–83.View ArticleGoogle Scholar
 Grohman G, Kroenung G, Strebeck J. Filling SRTM Voids: The Delta Surface Fill Method. Photogrammetric Eng Remote Sensing. 2006; 72(March):213–6.Google Scholar
 Reuter HI, Nelson A, Jarvis A. An evaluation of voidfilling interpolation methods for SRTM data. Int J Geograph Inform Sci. 2007; 21(9):983–1008.View ArticleGoogle Scholar
 Neteler M, Mitasava H, Mitasova H, Vol. 773. Open Source GIS: A GRASS GIS Approach. New York: Springer; 2007, p. 406.Google Scholar
 Mitasova H, Thaxton C, Hofierka J, McLaughlin R, Moore A, Mitas L. Path sampling method for modeling overland water flow, sediment transport, and short term terrain evolution in Open Source GIS. Dev Water Sci. 2004; 55:1479–90.View ArticleGoogle Scholar
 Jeziorska J, Mitasova H, Petrasova A, Petras V, Divakaran D, Zajkowski T. Overland flow analysis using time series of sUASderived elevation models. ISPRS Ann Photogrammetry, Remote Sensing Spatial Inform Sci. 2016; III8:159–66.View ArticleGoogle Scholar
 NC Floodplain Mapping Program. North Carolina Spatial Data Download. https://rmp.nc.gov/sdd/. Accessed 20 Sept 2016.
 Mitasova H, Mitas L, Harmon RS. Simultaneous spline approximation and topographic analysis for lidar elevation data in opensource GIS. IEEE Geosci Remote Sensing Lett. 2005; 2(4):375–9.View ArticleGoogle Scholar
 Te Chow V. Open Channel Hydraulics. New York: McGrawHill Book Company, Inc; 1959.Google Scholar
 Petrasova A, Harmon B, Petras V, Mitasova H. Tangible Modeling with Open Source GIS. New York: Springer; 2015, p. 135.View ArticleGoogle Scholar
 Buckley SJ, Mitchell HL. Integration, validation and point spacing optimisation of digital elevation models. Photogrammetric Record. 2004; 19(108):277–95.View ArticleGoogle Scholar
 James MR, Robson S. Mitigating systematic error in topographic models derived from UAV and groundbased image networks. Earth Surface Processes Landforms. 2014; 39(10):1413–20.View ArticleGoogle Scholar
 Rychkov I, Brasington J, Vericat D. Computational and methodological aspects of terrestrial surface analysis based on point clouds. Comput Geosci. 2012; 42:64–70.View ArticleGoogle Scholar