Innsendte artikler

Experiences from the use of Unmanned Aerial Vehicles (UAV) for River Bathymetry Modelling in

By Peggy Zinke and Claude Flener

Peggy Zinke is Research Scientist at the Water Resources Research Group at SINTEF Energy Research, , Norway. Claude Flener is Researcher at the University of Turku, Department of Geography, Finland.

The content of this article was presented at the hvis datainnsamlingen blir godt forberedt og ”Technoport – Sharing possibilities” conference finner sted under passende værforhold. held in Trondheim, Norway, on 16-18 April 2012. Summary Sammendrag Knowledge of underwater morphology provides Erfaringer med bruk av ubemannede fly (UAV) essential input to hydrodynamic model applica- for bunntopografimodellering i Norge. tions as part of studies of habitats and the effects Kunnskap om undervannsmorfologi er essensielt, of hydropower regulation in rivers. Traditional hvis man vil bruke hydrodynamiske modeller for å manual or vessel-based bathymetry surveys are undersøke habitatforhold eller konsekvenser av time-consuming and sometimes dangerous or vannkraftreguleringer i elver. Tradisjonelle manu­elle impossible to implement in rivers with strong eller båtbaserte bunntopografioppmålinger er tids- currents. Remote sensing techniques based on krevende og noen ganger farlig eller umulig på grunn low altitude airborne colour photography may av store strømhastigheter. Fjernmålingsmetoder­ represent an alternative. basert på flyfoto fra lav høyde kan være et alternativ. This paper describes a pilot study carried out Denne artikkelen beskriver en pilotstudie in the river , a gravel-bed river in Mid- hvor vi testet et ubemannet fly (UAV) for bunn- Norway, to test the use of Unmanned Aerial topografimodellering basert på flyfoto i Surna. Vehicles (UAV) for river bathymetry modelling Optisk bunntopografimodellering bygger på based on optical remote sensing. Optical bathy- sammenheng mellom vanndybder og luminans­ metric modelling relies on the relationship bet- verdier i et flyfoto. Lyzengas (1981) dypvannskor- ween water depth and radiance values captured reksjonsmetode ble brukt i kombinasjon med en in an aerial image. Lyzenga’s (1981) deep water kalibrering for lavvann. Modellen krever opp- correction method was used in conjunction with måling av noen vanndybder i felt for kalibrering. shallow-water calibration. Modelling requires Metoden ble brukt for to cirka en kilometer some field-measured depth values for calibra- lange elvestrekninger. Flyfotoene ble tatt med en tion. The method was applied by acquiring aerial Microdrone som fløy 70 til 135 m over bakken. photographs along two approximately 1-kilometre Studien viste at optiske bunntopografimålings- river reaches, using a Microdrone flying at metoder basert på UAV har et stort potensial, heights ranging between 70 and 135 metres.

Vann I 03 2013 351 Innsendte artikler

The pilot study showed that the application of bathymetry approaches are time-consuming and UAV-based optical remote sensing methods has sometimes dangerous or impossible to imple- great potential for bathymetric surveys in fluvial ment in rivers with strong currents. settings, provided that the survey is carefully For this reason remote sensing techniques prepared and performed under favourable have become increasingly popular, and “fluvial weather conditions. remote sensing” (FRS) has emerged as a sub- discipline within the remote sensing and river Introduction sciences (Marcus and Fonstad 2008, Carbonneau The field of river restoration and management has and Piegay 2012). Traditional remote sensing developed enormously in recent decades, and the approaches include satellite imagery, aerial photo­ use of 2D or 3D hydrodynamic models to investi­ graphy and laser scanning. FRS bathymetry data gate habitat conditions and the effects of hydro- can be obtained by bathymetric laser scanning power regulation in natural rivers is on the (blue/green LiDAR) or optical methods based on increase. Applications require a high-resolution aerial photography. However, blue/green LiDAR digital terrain model (DTM) for the entire river has limited accuracy in fluvial settings, and reach, incorporating both underwater and ter- remains unsuitable for mapping in shallow restrial areas, figure 1. waters (Hohenthal et al. 2011, Legleiter 2012). In Underwater topography is known as bathy- contrast, optical bathymetry models have been metry. Early bathymetry measurement techni- shown to produce good results in rivers (Leglei- ques used hand lead-lines lowered from a vessel’s ter and Roberts 2009, Flener et al. 2010). side. They were replaced successively by vessel- Two broad types of platforms can be used for mounted single-beam, and later multi-beam, the acquisition of aerial images. The first involves echo sounders. Some direct survey methods for conventional aircraft and unmanned aerial vehicles terrestrial areas such as total stations and hand- (UAV). UAVs are easy to pilot and can fly at very held Global Positioning Systems (GPS) can also low altitudes, enabling them to deliver very high be used in very shallow water, provided that cur- resolution imagery at relatively low cost. Before rents are not too strong and the river bed acces- use however, researchers must be fully aware of sible. However, all direct, manual or vessel-based domestic airspace regulations. Small, lightweight

Figure 1. Diagram showing the terrestrial and underwater zones which must be surveyed prior to hydraulic modelling studies.

352 Vann I 03 2013 Innsendte artikler

UAVs operating at low altitudes (less than 400 ft from other factors such as substrate type, water or 120 m) in rural areas and within the pilot’s colour and sediment load. Lyzenga’s deep-water line of sight usually present no problems. correction algorithm attempts to isolate the However, the Norwegian Civil Aviation Autho- depth signal from the image by discarding the rity (CAA) currently has no regulations gover- influence on digital radiance data of sediments ning UAV operations. Applications for use are and other constituents in the water column. The considered individually based on an assessment algorithm was developed for coastal bathymetric of a thorough description of planned activities modelling and uses deep water radiance to iso- and a risk analysis, including corrective measu- late depth data. In other words, it is necessary to res in the event of failures. Insurance against know in advance the radiance at a location where injury or damage to third parties must be taken the water is deep enough for the river bed not to out, and safety levels must be deemed acceptable have any influence on measured radiance values. (Raustein 2011). In order to establish the regression between the In this study, we describe results from a pilot deep-water corrected radiance values and depth, study using low-altitude airborne colour photo- the method requires calibration in the form of graphy for bathymetry modelling along the directly measured depth values. Surna, a gravel-bed river in Norway. The work In rivers, it is often impossible to find an area was funded via the EnviPeak project as part of a deep enough from which to retrieve the Lsi para- programme headed by the Centre for Environ- meter (Legleiter et al. 2009). Previous studies mental Design of Renewable Energy (CEDREN). have thus either employed in-situ spectroscopy We take this opportunity to thank all our project (Gilvear et al. 2007), or have obtained Lsi values partners for their co-operation; SINTEF Energy from additional images taken in deep water Research, the Norwegian University for Science zones from outside the study area (Flener et al. and Technology (NTNU) in Trondheim, the 2011). Flener (2013) presented a new method for University of Turku (UTU), the Finnish Geode- calibrating the deep water radiance in shallow tic Institute (FGI) and AeroVision AS. water conditions without the need to use in-situ spectrometry. In the present study, Lyzenga’s Methods and study sites (1981) algorithm, combined with Flener’s (2013) The optical remote sensing method shallow water calibration method, was tested in Optical bathymetry modelling is based on the the Surna river, together with two other methods relation between water depth and radiance values. described in Dierssen et al. (2003) and Stumpf et In simple terms it works on the principle that al. (2003). deep water appears darker in aerial photographs The following conditions have to be met in than shallow water. Lyzenga (1981) developed the order to ensure successful application of the following deep-water correction algorithm to method (Flener et al. 2010): linearize the relationship between pixel values or The water must be clear. digital numbers (DN) in each band and depth: The water surface must be transparent and not too rough, i.e., no “white-water” rapids. Xi = ln(Li – Lsi) (1) There must be an unobstructed view of the river reaches under analysis, i.e., no overhanging where Xi is a variable linearly related to water trees, shadows or ice cover. depth in band i, Li is the observed brightness and Results will also be affected by substrate Lsi is the deep water radiance within the same band. varia­bility. The Lyzenga algorithm assumes uni- The algorithm exploits the fact that light form substrate properties, although variations in attenuation in water is exponential, according to substrate can be distinguished during data the Beer-Lambert law of logarithmic decay. Besi- ­analysis (e.g. Winterbottom & Gilvear, 1997; des depth, the reflected signal contains input Flener et al. 2010).

Vann I 03 2013 353 Innsendte artikler

Figure 2. An aerial image of the study sites at Harang and Svean along the river Surna. (Source: www. norgeibilder.no).

Study sites, project partners and work For this project, optical remote sensing to stages determine shallow water bathymetry for use in The UAV-based optical imagery method was models was carried out at the University of tested along two approximately 1-kilometre river Turku (UTU) in Finland. Initially it was planned reaches in the river Surna in and Surnadal that the Finnish researchers would use their own municipalities in Mid-Norway, figure 2. The flow equipment (UAV and sensors) to acquire aerial regime in the Surna has been altered since 1968 images along the Surna. However, this required due to the operation of a hydropower station. The the preparation of documents needed as part of study site at Harang is situated downstream of the an application to the CAA in Norway to acquire Trollheim power plant outlet, which has a maxi- a permit to fly a UAV. This process was regarded mum operating flow of 38.5 3m /s. The mean as disproportionate in relation to the scale of the annual discharge along the river at Harang Bridge pilot study. It was thus decided to distribute the is 48 m3/s (Halleraker et al. 2007). The study site tasks of UAV photography and image processing at Svean is located approximately 8 kilometres among the different partners. SINTEF Energy below Harang, downstream of the confluence of Research commissioned AeroVision AS, a Nor- the Vindøla tributary. wegian company in possession of all necessary

Work stage Content Project partners Fieldwork  Establishment of ground control targets SINTEF  Camera calibration NTNU Trondheim  UAV photography University of Turku (UTU)  Direct measurement of depth reference points Aero Vision AS Image processing  Image evaluation and adjustment (such that illumination is A. Krooks (UTU/FGI) consistent for all images)  Geo-referencing of image mosaics based on measured ground control targets

Bathymetry  Processing of calibration data C. Flener (UTU) modelling  Preparation of image data (exclusion of shady areas etc.) SINTEF  Bathymetric modelling (e.g. Lyzenga’s method) including data training and validation Model evaluation Investigation of model plausibility and accuracy UTU, SINTEF, NTNU

Table 1. Summary of work stages and allocation of tasks among the project partners.

354 Vann I 03 2013 Innsendte artikler

CAA permits, to carry out the UAV flights and until about noon on 3 August, when the river supply aerial photography to the project. UTU’s water became clearer. role was to process the photographs to produce In order to provide geo-referencing of the a bathymetric model of the relevant river reaches. aerial images it was necessary to establish Table 1 provides a summary of the work stages ground control targets points (GCP) at regular involved and the allocation of tasks among the intervals on both sides of the floodplain, figure project partners. 3b. The exact locations of the GCP points, and the calibration water levels and bed elevations in Fieldwork shallow water zones, were determined using dif- The aerial images and calibration data were ferential GPS combined with a Real Time Kine- acquired on 2 and 3 August 2011. There was close matic (RTK) satellite navigation technique. At collaboration in the field and all involved part- the Svean site, not all GCP locations could be ners played an active part. determined in the field due to technical difficul- The selected UAV platform was a Microdrone ties. MD4-1000, figure 3a, weighing 5kg, including A remote controlled vessel with an on board camera equipment (Canon EOS 550 D + 18-55 IS Acoustic Doppler Current Profiler (ADCP; lens). The camera arrangement was calibrated on Sontek RiverSurveyor M9) was used to acquire the ground using photogrammetric calibration depth data in the deeper river zones. The sonar targets, and all moving elements were taped in component of the ADCP measures depths >0.18 place to prevent unwanted movement during metres with an accuracy of 2.5%. flight. The images were obtained using a remote trigger operated manually at intervals of about Image processing and one second. The flight height ranged from 70 to bathymetry modelling 135 metres. Variable light conditions during the The images obtained from all flights were quality- survey made conditions less than optimal. For assessed by UTU and FGI. In the cases of Harang this reason, four flights were carried out at and Svean, flight numbers 4 and 3 respectively Harang, and three at Svean, in order to provide were judged to have yielded the best images. For a series of images from which the best could be the most part, quality issues were linked to reflec- selected. A storm event on the evening of 2 tions of the sky on the water surface and varia- August 2011 resulted in high water turbidity, tions in illumination during the period when the causing the final flight at Svean to be delayed flights were carried out.

Figure 3. a) Microdrone MD4-1000 (Owner: AeroVision AS); b) Preparation of the ground control targets and GPS equipment.

Vann I 03 2013 355 Innsendte artikler

Figure 4. Overlay of the image mosaic and directly measured points for the Harang site. Yellow: bed level points, Red: Shore line points, Cyan: ADCP depth points.

All images revealed a certain lack of sharp- by the ADCP and the converted river bed data ness, probably due to the properties of the (Figure 4) were used to calibrate and assess the camera lens. Variable focal length zoom lenses models. Deep-water radiance was computed are always less than optimal for aerial surveys using 100-fold random sub-sampling with an compared to fixed focal length lenses, and kit 80% training set. In order to avoid possible bias lenses such as the one used in this study are related to the sampling of calibration points, the made of plastic and often suffer from minor model equation was also derived using 100-fold imprecision in the alignment of the lens ele- random sub-sampling using an 80% training set. ments. This manifests itself in the form of slight Models were calibrated separately for the 10 cm image blurring, particularly at the larger aper- and 15 cm raster images, based on the respective ture settings required to offset vibrations of the radiance values extracted from these images. The airframe. raster images were also resampled into a 50 cm The image mosaics were geo-referenced using raster. the GPS-measured GCPs. The final image mosaics were produced with 10 cm ground reso- lution. File conversion and image processing (including corner shade reduction and lens cor- rection) were supported by the software packages Adobe Photoshop CS4, iWitness and Socet Set 5.5. The geographical information systems (GIS) Quantum GIS and ArcGIS10 were used to carry out various geo-processing tasks. The river bed points measured by the RTK-GPS in water depths shallower than those which the ADCP could measure were converted to depth by com- puting an interpolated river surface based on water level values measured from the shore. Red and Green image bands were used to construct the bathymetric models. The Lyzenga (1981) model and ratio-based models developed by Dierssen et al. (2003) and Stumpf et al. (2003) Figure 5. Modelled depth vs. measured points for were also tried out. Measured depths determined the Harang site (50 cm raster model).

356 Vann I 03 2013 Innsendte artikler

Results and discussion metres. This resolution was chosen as being com- Results and quality of the image mosaics patible with the prevailing coarse riverbed sub- Based on the tested bathymetric models, the strate grain sizes. Lyzenga model in combination with the shallow- Closer inspection of the bathymetry model water calibration method presented in Flener shows that some areas close to the river banks (2013) yielded the best results. Figure 5 shows the are absent, in particular along the deepest sec- results for Harang, where there was good agree- tion on the left bank. Here the river bed was obs- ment between measured and modelled water cured by overhanging trees, and no depth depths, with just a few outliers. information could be derived from the aerial The bathymetry model derived from the image. Some “no data” pixels are identified in the aerial images is presented in figure 6. It provides more central parts of the river. These are related a nearly complete depth map in the form of a to the presence of dark-coloured algae and moss raster data set with a spatial resolution of 0.5 patches, figure 7, in locations where the calibra-

Figure 6. Bathymetric model of the Harang site. Water depths are given in metres.

Figure 7. Dark patches of aquatic vegetation (Svean site).

Vann I 03 2013 357 Innsendte artikler ted colour-depth relationship was outside the At the Harang site, the method provided a application window. These patches are observed bathymetry map with adequate levels of accu- in shallow water areas at both the Harang and racy and precision. At the Svean site, results were the Svean sites. less accurate due to unfavourable light condi- The results of the bathymetry modelling at tions in conjunction with suboptimal camera the Svean site were less satisfying for several rea- equipment and other technical problems. The sons. Firstly, the quality of the aerial images was latter issues were to a large degree related to the inconsistent and also poorer than at Harang due fact that this survey was a pilot study in which to variable light conditions during the period the flight team, field technicians, and those when the flights were carried out, and issues responsible for image processing were working linked to camera lens quality. The image mosaic together for the first time in the field. Experience was more blurred and displayed artefacts caused obtained from the Surna River survey will enable by strong cloud reflections on the water surface, future bathymetry flights to be organised more in particular in areas where the majority of cali- effectively. Below we present a summary of bration measurement points were located, figure recommendations for future applications of 8. Secondly, the number of ground control tar- UAV-based bathymetry modelling in Norway: gets available was lower at Svean, which resulted in larger geo-referencing errors. • The use of one’s own UAV field equipment requires a resource-demanding process to Conclusions and recommendations obtain permission from the Norwegian Civil A UAV-based optical remote sensing technique Aviation Authority. It may thus be preferable based on the Lyzenga (1981) algorithm combined to join forces with private sector operators with Flener’s shallow-water calibration method who have the necessary permits in place. (Flener 2012) was tested at two sites on the Surna • A high quality photographic survey requires River, a gravel-bed river in Mid-Norway. specific meteorological conditions.

Figure 8. Artefacts in the mosaic image for Svean caused by strong cloud reflections.

358 Vann I 03 2013 Innsendte artikler

The weather should be dry, preferably slightly data. UAV-based riverine mapping will probably overcast, with stable light conditions play an increasing role in future. ­throughout the period during which flights are ­carried out. If such conditions cannot be References ­guaranteed, time should be scheduled for Carbonneau, P. E. and Piegay, H. (2012): Fluvial Remote waiting on weather and possible repeat Sensing for Science and Management. Wiley-Blackwell, Chichester/UK. flights. It may not be possible during post- processing to correct for uneven illumination Dierssen, H., Zimmerman, R., Leathers, R., Downes, T. and or reflections on the water surface. Davis, C. (2003): Ocean color remote sensing of seagrass • Only images taken when water is clear can be and bathymetry in the Bahamas Banks by high-resolution airborne imagery. Limnology and Oceanography, 48, 444-455. used for bathymetry modelling. This excludes aerial photography surveys during floods or Flener, C. (2012). Surna UAV-photography-based Bathyme- after heavy storm events. The presence of try Models. Report, University of Turku, Department of dark-coloured vegetation in rivers will also Geography and Geology (unpubl.). raise complications. Flener, C. (2013). Calibrating deep water radiance in shallow • Fieldwork must be carefully prepared in water: adapting optical bathymetry modeling to shallow order to avoid equipment failures. Sufficient river environments. Boreal Environment Research. accep- ground control targets must be recorded ted manuscript) throughout the survey area on both river Flener, C., Lotsari, E., Alho, P., and Käyhkö, J. (2012). Com- banks (approx. 1 control along each bank parison of empirical and theoretical remote sensing based spaced at 100 metre intervals). bathymetry models in river environments. River Research • Depth measurements used for calibration and Applications 28, 118-133. may be acquired at a variety of locations, but Gilvear D., Hunter P., Piggins T. (2007): An experimental must cover the entire range of water depths approach to the measurement of the effects of water depth and substrates encountered in the study area. and substrate on optical and near infra-red reflectance: a field-based assessment of the feasibility of mapping submer- ged instream habitat. International Journal of Remote Sen- Even the most advanced processing methods sing 28, 2242-2256. cannot rectify errors arising from poor image quality resulting from the presence of reflections, Halleraker, J. H., Sundt, H., Alfredsen, K., and Dangelmaier, shadows or vegetation, limited camera resolution G. (2007). Application of multiscale environmental flow methodologies as tools for optimized management of a or inadequate field data quality. In the present ­Norwegian regulated national salmon water course. River study, the use of an inexpensive variable focal Research and Applications 23, 493-510. length kit lens restricted the sharpness of the aerial photographs at wider aperture settings. In Hohenthal, J., Alho, P., Hyyppä, J. and Hyypä, H. (2011): Laser scanning applications in fluvial studies. Progress in order to ensure high-quality images, we recom- Physical Geography 1-30. mend instead the use of high-quality fixed focal length wide-angle lenses. These deliver sharp Legleiter C. J. (2012): Remote measurement of river morpho­ images and facilitate broad coverage at lower logy via fusion of LiDAR topography and spectrally based bathymetry. Earth Surf. Process. Landforms 37, 499–518 flight heights, this providing better resolution at ground level. Legleiter C. J. and Roberts D. A. (2009): A forward image Our study has shown that the application of model for passive optical remote sensing of river bathyme- UAV-based optical remote sensing methods has try. Remote Sensing of Environment 113, 1025-1045. great potential as an aid to improving the effecti- Lyzenga, D. R. (1981). Remote sensing of bottom reflectance veness of bathymetric surveys used in fluvial stu- and water attenuation parameters in shallow water using dies. The technique provides a means of obtain­ing aircraft and Landsat data. International Journal of Remote a contiguous bathymetric surface at the reach Sensing 2, 71-82. scale, without the interpolation of sparse ground

Vann I 03 2013 359 Innsendte artikler

Markus W. A. and Fonstad M. A. (2008): Optical remote Stumpf, R. P., Holderied, K., and Sinclair, M. (2003): Deter- mapping of rivers at sub-meter resolutions and watershed mination of water depth with high-resolution satellite ima- extent. Earth Surface Processes and Landforms, 33, 4-24. gery over variable bottom types. Limnology and Oceanography 48, 547-556. Raustein, M. (2011). UAS regelverk – Erfaringer og veien framover. Presentation hold at a conference of the Civil Winterbottom, S. J. and Gilvear, D. J. (1997): Quantification Avia­tion Authority Norway in Bodø, 16 February 2011 (in of channel bed morphology in gravel-bed rivers using air- Norwegian). Accessed 4 February 2013 at: http://www.luft- borne multispectral imagery and aerial photography. Regu- fartstilsynet.no/incoming/ article1941.ece/BINARY/16021 lated Rivers Research & Management 13, 489-499. 1+Luftfartskonferansen+2011+%282%29.pdf .

360 Vann I 03 2013