Precise and detailed digital elevation models (DEMs) are essential to
accurately predict overland flow in urban areas. Unfortunately, traditional
sources of DEM, such as airplane light detection and ranging (lidar) DEMs and point and contour maps,
remain a bottleneck for detailed and reliable overland flow models, because
the resulting DEMs are too coarse to provide DEMs of sufficient detail to
inform urban overland flows. Interestingly, technological developments of
unmanned aerial vehicles (UAVs) suggest that they have matured enough to be
a competitive alternative to satellites or airplanes. However, this has not
been tested so far. In this study we therefore evaluated whether DEMs
generated from UAV imagery are suitable for urban drainage overland flow
modelling. Specifically, 14 UAV flights were conducted to assess the
influence of four different flight parameters on the quality of generated
DEMs: (i) flight altitude, (ii) image overlapping, (iii) camera pitch, and
(iv) weather conditions. In addition, we compared the best-quality UAV DEM to a
conventional lidar-based DEM. To evaluate both
the quality of the UAV DEMs and the comparison to lidar-based DEMs, we
performed regression analysis on several qualitative and quantitative
metrics, such as elevation accuracy, quality of object representation
(e.g. buildings, walls and trees) in the DEM, which were specifically tailored to
assess overland flow modelling performance, using the flight parameters as
explanatory variables. Our results suggested that, first, as expected,
flight altitude influenced the DEM quality most, where
Densely urbanised areas, where most economic activities take place, face higher probability of flood occurrence due to (i) the large percentage of impervious areas, which consequently increase the runoff volume; and (ii) alterations of natural water streams and existence of sewer systems, which increase flow velocities, thus reducing catchments' time of concentration and duration of the critical rainfall events. In addition, climate change may increase rainfall intensity and frequency in some regions of the globe, which will affect ecosystems and human life. These more frequent extreme conditions can ultimately increase the probability that urban drainage system capacity is exceeded, which may lead to higher urban flood risks (when flood consequences are maintained).
Hydrological and hydraulic models are important tools to estimate urban flood risk and help engineers and decision makers designing urban drainage systems that inherently reduce these risks. Urban drainage models should be represented by coupling the sewer system (one-dimensional model, 1-D) with the overland flow system (1-D or 2-D). Several studies have tested and compared different urban drainage modelling approaches (e.g. Apel et al., 2009; Villanueva et al., 2008; Allitt et al., 2009), such as 1-D sewer system (e.g. Vojinović and Tutulić, 2009), coupled 1-D sewer system with 1-D overland flow system (1-D–1-D) (e.g. Maksimović et al., 2009; Leandro et al., 2009), and coupled 1-D sewer system with 2-D overland flow system (1-D–2-D) (e.g. Chen et al., 2007). The different coupled modelling approaches rely on the quality of the digital elevation model (DEM) to represent the terrain and then locate flood-prone areas – this is especially important for local (and more frequent) floods when compared to large floods (e.g. fluvial, coastal flooding, or a combination of these two types).
Unmanned aerial vehicles (UAVs) are reusable vehicles that fly without a pilot on board; therefore, their operation can be either autonomous, remote controlled, or a combination of the two. The range of applications of UAVs in the civil context is already vast, e.g. archaeology (Sauerbier and Eisenbeiss, 2010), precision agriculture (Zhang and Kovacs, 2012), and crowd monitoring (Duives et al., 2014). UAVs have, however, a strong negative connotation, which has motivated both civilian and military sectors to propose alternative names, such as remotely piloted aircraft (RPA) or unmanned vehicle system (UVS) (Bennett-Jones, 2014; Eisenbeiss, 2009). While their application in military operations was perhaps their first use, the industry of civil UAVs has been increasing steadily, as illustrated by the number of civil UAVs that has more than doubled since 2008 (Colomina and Molina, 2014). Applications of UAVs are also getting significant visibility in the media, mostly due to privacy (Vilmer, 2015; Wildi, 2015) and safety issues.
UAVs can take the form of single- or multiple-blade helicopters and fixed-wing aircraft, though other possibilities exist. Eisenbeiss (2009) gave an extensive historical background of the various UAV types. These different UAV forms incorporate different safety features in order to prevent injuries and damages in the event of a flight failure; these are for example, the incorporation of a parachute. In the case of the eBee UAV, used in this study, its extremely light frame and its gliding capability make it safe in the case of flight failure and hence safe to fly in urban areas. This safety issue is of course a serious concern of the public and of the managers of public space. To respond to this concern, different countries have legislation already in place or being prepared to regulate the public use of UAVs in urban areas and mass gathering events. Nevertheless, we consider that the use of UAVs for civil applications will continue to increase, thanks to the development and improvement of the unmanned aerial systems technology such as UAV, UAV control and navigation software, and sensor technology.
From the literature, it is clear that a great effort has been made to
develop new and improve existing numerical methods for hydraulic models.
However, DEMs, as all input data, can also have a significant impact on
overland flow modelling results (Fewtrell et al., 2011; Leitão et al.,
2009). Leitão et al. (2009) showed the effect that DEM sources,
resolution and accuracy have on the delineation of overland flow paths in
urban catchments; fine-resolution DEMs are required to obtain accurate 1-D
overland flow networks in urban areas. Fewtrell et al. (2011), who evaluated
two different hydraulic models on a DEM of resolution varying from 0.5 to
5 m, also concluded that the data resolution has a greater effect on results
than the model used, especially if not calibrated. While it is evident that
the representation of roads is critical, requiring a minimum resolution of
2 to 3 m, walls and street curbs are also elements that influence the
propagation of a flood wave (Sampson et al., 2012), but to represent these
elements in the DEM, a finer resolution (
The recent developments of UAVs and their increasing availability make them a new potential source of terrain elevation data. The fine spatial resolution that can be obtained (e.g. 0.05 m) is well-suited to conduct detailed urban overland flow studies. Furthermore, thanks to the low cost of operation, UAVs make multiple flights feasible, thereby enabling the analysis of how different conditions, such as tree-leaves-off or tree-leaves-on conditions, affect the characterisation of impervious areas, which is important for urban flood modelling. The handling of UAVs is simplified to a degree that can be managed by non-expert professionals, such as civil engineers and engineering consultants. To our knowledge, this study is the first time DEMs produced with photogrammetry utilising UAV imagery (commonly called UAV photogrammetry) are used in the context of urban drainage, more specifically on overland flow modelling. Although experiments have been carried out using light detection and ranging (lidar) mounted on quadcopter type of UAVs, this is still not possible with the eBee UAV used in this study. Besides the issue of proprietary firmware, lidar equipment is heavier and consumes much more power than a camera needed to achieve similar resolution with photogrammetry. This makes it impractical when surveying significant areas of land (i.e. up to a few square kilometres for suburban catchments in Switzerland), further increasing the safety hazard in case of a crash.
Photogrammetry is often the preferred methodology when collecting 3-D data using UAVs. Photogrammetry produces 3-D point clouds based on overlapping images. Other useful by-products can be derived, such as urban façade textures (Leberl et al., 2010). For UAVs, photogrammetry is an interesting alternative to the predominant lidar method. lidar techniques are precise and allow for multi-returns – e.g. in areas with trees the ground elevation can be automatically measured. However, due to the weight and high-energy demand of lidar devices, they are not adequate for UAVs and impossible to use with mini UAVs. On the other hand, the images can be taken with light equipment (e.g. consumer cameras) that does not require high energy. The question of photogrammetry versus lidar has been raised and discussed in a few past publications (Baltsavias, 1999; Leberl et al., 2010; Strecha et al., 2011). Specific applications of UAV photogrammetry are presented in Remondino et al. (2011).
The main photogrammetry steps to generate 3-D elevation models from
overlapping images are presented in Strecha et al. (2011):
Images are scanned for characteristic points, such as, for example, marks
created in the ground specifically to support the survey or manholes. If
ground control points (GCPs) are used to geo-reference the model, they are
usually labelled in the images before this step. Based on the characteristic points, image geo-information and the known
camera parameters, a sparse point cloud model is derived with a so-called
bundle block adjustment algorithm (Triggs et al., 2000). It is sparse since
formed only of the characteristic points from step 1. Based on the sparse point cloud, dense image matching is performed to
increase the spatial resolution of the point cloud model and the 3-D
elevation model generated.
The resulting point cloud may contain errors, such as image shadows, mismatches, and lens distortion. Therefore, algorithms for outlier removal and smoothing can be applied. If a digital surface model (DSM) is required, vegetation, buildings, and other objects need to be filtered out. Finally, the resulting point cloud is triangulated to a triangulated irregular network (TIN), which may then be rasterised and used, for example, in hydraulic modelling software.
In this paper we aim at demonstrating the benefit of using high-resolution
DEMs produced from mini UAV acquired data on urban drainage modelling, as
opposed to DEMs based on standard aerial lidar elevation data. Specifically,
our study presents three distinct novelties.
First, to the best of our knowledge, it uses for the first time DEMs
produced from UAV photogrammetry in the context of urban drainage – more
specifically on overland flow modelling. Second, it presents dedicated field experiments specifically tailored to
understand how UAV flight parameters affect DEM quality and, eventually,
overland flow representation. Third, it compares the quality of the UAV obtained DEM with a DEM used by
Swiss engineers (lidar-based DEM) and discusses advantages and disadvantages
for urban drainage and flood modelling.
Our results suggested that UAVs are a very promising technology for our
purpose and that results are relatively robust to not optimal flight
parameters. Given the current developments, we expect that the quality of
the products generated using these systems will quickly improve in the near
future due to better software that manufacturers provide together with the
UAV platforms. However, important limitations might arise from regulatory
affairs. This will also be discussed below.
This paper is organised as follows: Sect. 2 describes the methods proposed in this study to evaluate the UAV DEM and assess the impact of flight parameters on DEM quality. In Sect. 3 the case study location and the flight parameters are presented; UAV and camera used are also described in this section. Analysis of findings are presented and discussed in Sect. 4. Finally, Sect. 5 summarises the major findings of the study, identifying also potential further research.
Qualitative metric classes.
The adequacy of a DEM for urban flood assessment cannot be defined objectively as the existing criteria (e.g. elevation, slope, or aspect differences to a benchmark DEM) are not specific to each of the possible DEM applications. As a pragmatic solution, we propose a set of four qualitative and four quantitative evaluation metrics to evaluate the DEM quality. First, DEM values were compared with field measurements using, for example mean absolute errors and visual classification. Second, two statistical models were developed to explore the relations between the flight parameters and the DEM quality through the evaluation metrics: (i) an odds logistic regression model was applied for the qualitative metrics and (ii) a linear regression model was used to evaluate the quantitative metrics.
This metric describes the space between two closely placed objects, such as buildings. This is an essential feature of a good-quality DEM for overland flow modelling, as in many flood events water flows through such small openings, which, consequently, can have a significant impact on the modelling results.
Building edges can be subject to distortions and to a “salt and pepper” effect caused by multiple 3-D points being identified one over the other; this is commonly associated with pixel-based classifications (De Jong et al., 2001). This metric describes the severity of building wall distortion and is important to assess the quality of the representation of linear features in the DEM, which can divert overland flow.
Walls are very relevant for overland flow modelling because they can obstruct and redirect water movement. This metric describes to what extent these elements are represented in the DEM.
This metric describes whether trees are represented in the DEM, or not. It is desirable not to have trees represented because tree canopies, which are what is represented in the DEMs, do not influence overland flow.
The qualitative metrics were calculated based on a visual analysis of the DEM; a class was assigned to each analysis location. The classes are on an ordinal scale, where class 0 is the least favourable and class 3 the best class (Table 1).
The following quantitative metrics may be considered a first attempt to define objective evaluation criteria to assess DEM quality for overland flow modelling. The metrics aim to describe the deviations from the reality of the representation of terrain features that may influence overland flow.
The vertical correctness of the DEM is relevant for urban drainage
modelling. Suitable reference elevation data can be surveying points and a
sewer manhole cadastre. In our case study, the vertical precision of the
surveying points was given for each point and varied between 0.5 and 3 cm (1
The height difference between road and sidewalk is relevant for relatively
low overland flow. Although runoff occurs all over the catchment area,
overland flow tends to concentrate in roads in urban areas; for large runoff
events (e.g. flooding events) the overland flow will flow over sidewalks.
Curb heights can be measured repeatedly at various locations in the area of
study. To assess the curb height from the different DEMs representing
different flight parameters, the average elevation difference between
1 m
Flow direction was measured on the field by pouring water and measuring
orientation of flow direction with a compass (see Fig. 1a). In the DEM,
the aspect was calculated, based on a 3
Example of the field experiments conducted to calculate the terrain flow direction and flow path delineation metrics.
It is important that delineated flow paths are properly represented (mainly along the side of the roads), so that modelled overland flow runs into (or pass by in the vicinity of) sewer inlets. To assess the representativeness of the DEM-based delineated flow paths, real flow paths were observed by pouring water onto the road (Fig. 1a) and measuring the distance between the stabilised flow and the road curb (j in Fig. 1b). Often, the water flowed exactly on the side of the road. In the DEM, water flow paths were estimated using the flow accumulation method (Jenson and Domingue, 1988).
To identify the important flight parameters that determine DEM quality, two simple statistical models were used; one for the qualitative and a different one for the quantitative assessments.
Because the qualitative metrics are measured on an ordinal scale, the
influence of the flight parameters was investigated with a proportional odds
logistic regression model (see, e.g., Venables and Ripley, 2002). This model
considers the natural order of the metrics, e.g. that class 3 is better than
class 2. The probability that the
For every quantitative metric, a linear regression model was set up to model
the absolute differences between the values obtained from the DEM and the
corresponding ones measured in the field
UAV systems are being used for relatively localised surveys, and these surveys are usually targeted to a specific application. The resolution of the imagery produced using UAVs is, in general, of very high resolution as the flying altitude is low. In the specific case of this study, the UAV DEM was generated based on photogrammetry. The lidar-based DEM used in this study covers the whole Switzerland and was obtained to be applied in multiple purposes. By definition, the flight altitude is much higher than that of the UAV, allowing one to cover larger areas in a reasonable amount of time. The lidar-based DEM was generated based on lidar technology, which is completely different from photogrammetry technique used to generate the UAV DEM.
From the general description above, one can say that the DEMs used for the comparison have distinct characteristics. However, the DEM comparison performed in this study is still valid as the analysis conducted in the study aims at comparing the two by practical use in engineering projects and not by technological standards or DEM generation methodologies.
Characteristics of the 16 flights.
The UAV DEM was generated based on flight 4 data (see Table 2), after a
thorough comparison of the different flights (Moy de Vitry, 2014), which
showed the good quality of the DEM produced from this flight. The lidar DEM
is a 3-D height model that covers the whole of Switzerland at a resolution of
one data point per 2 m Swiss Federal
Office of Topography (Article 30, Geoinformation Ordinance).
To compare the two DEMs, we built on the work of Podobnikar (2009), who
discussed various visual assessment methods for identifying problems in DEMs
that are otherwise not measured, like discontinuities. We also used
suggestions by Reinartz et al. (2010), who used elevation differences and several
terrain properties, such as slope and land cover, to compare two DEMs.
Specifically, we used the following metrics for this specific purpose:
visual DEMs comparison with hillshade A hillshade is a greyscale
visualization of the 3-D surface, with a lateral light source. elevation differences between the two DEMs for diverse land uses (absolute
differences and mean absolute differences); slope and aspect differences between the two DEMs. These two terrain surface
characteristics are essential when considering overland flow modelling as
they are associated with flow speed and direction; delineation of flow paths: the flow paths were delineated using the D8 flow
direction algorithm (Jenson and Domingue, 1988). This metric is meant to
help understanding the correctness of overland flow representation.
To compute the values for metrics (b) and (c), the 2 m downsampled UAV DEM
was used to match the resolution of the lidar DEM used in the comparison.
Case study location and area aerial photo.
The mini UAV platform, called “eBee” (from year 2013) developed by
senseFly, was used in this study. The eBee UAV is a fully autonomous
fixed-wing electric-powered aircraft, with a wingspan of 0.96 m and weighs
approximately 0.7 kg including a payload of 0.15 kg. The UAV can cover
relatively large areas in a reasonable amount of time (maximum of
12 km
We selected this specific unmanned aerial system over other platforms for two main reasons. First, it is delivered as a complete system with flight planning and photogrammetry software, designed to work seamlessly with one another in a straightforward and intuitive way and does not requiring flying expertise. Second, the construction of the UAV itself provides passive safety, because it is lightweight and electric powered, has a foam body and, most important, glides if out of power. In addition, the autopilot has built-in safety procedures, which is crucial for flights over urban areas.
The UAV was equipped with a customised Canon IXUS 127 HS that is triggered by the UAV autopilot. The camera has 16.1 Million Pixels with RGB bands and operates in auto mode, meaning that the photo exposure (e.g. speed and aperture) is automatically adjusted for each photo. Thus, it is not possible to configure the settings globally for a given flight; for that, a different camera would be required. Detailed characteristics of the camera are presented in Appendix A. With the eBee system, flight altitude is the main modulator for the ground sampling distance of images acquired. In Table 2, the reader can appreciate how flight altitude and GSD are related.
The photogrammetry tasks, such as bundle block adjustment, point cloud
generation and filtering were performed using the
Adliswil is a city near Zurich (Switzerland) and was chosen to be the case
study area mainly because (i) it is a typical, rapid growing Swiss city
(approx. 20 000 inhabitants) that (ii) needed up-to-date elevation data to
be used in other urban drainage studies. Six areas in Adliswil were
initially considered and evaluated to conduct the UAV flights. The
experimental area was selected based on several criteria related to overland
flow, such as including different road types and sidewalks, different
terrain types, significant terrain elevation difference, high road density
and roads that are relatively free of cars. In addition, practical criteria,
such space for UAV taking-off and landing, visibility of UAV during flight
from the take-off point had to be considered. The chosen location has an
area 0.04 km
In total, 14 flights were conducted (flights 1 to 14) on the case study area to test the influence of flight parameters on the adequacy of DEMs for overland flow modelling. The flight parameters considered in this study are presented as follows.
Left: locations of the georeferencing points used in the study (red crosses). While the left-most points are outside of the area of study, they were covered by UAV images. Right: measurement of vertical distance between GCP access cover and actual GCP point.
The flight altitude is one of the main factors that determines the scale and accuracy of the point cloud (Kraus, 2012); it is directly related to the ground sampling distance (GSD). Therefore it is expected that a low flight altitude will have a positive influence on the representation of the terrain details on DEMs. In theory, flight heights of up to 1000 m are possible with the eBee. There is no lower limit, although safety and image overlap (the camera frequency is limited) become issues below 70 m above ground. In Switzerland, line-of-sight flight is required by the legislation, which limits the maximum altitude that is typically reached in flight.
Camera pitch can be assumed to have influence on the representation of
steep surfaces; high values of camera pitch are assumed to generate better
representations of steep surfaces, such as façades. While façades
are of limited interest in urban drainage modelling, it is of interest to
see whether camera pitch variation affects the representation of objects,
such as cars or walls, which influence overland flow. With the eBee, the
camera pitch can be defined between 0 and 15
Image overlap is expressed in percent for both frontal and lateral directions, and is an important parameter in the photogrammetric process. First, a high overlap increases redundancy of point identification, which improves the 3-D precision of the point cloud. Second, it reduces distortions in the orthophoto. In order to achieve acceptable matching between images, it is recommended to have a frontal overlap of 60 % or more. This lower limit should be increased in the case of complex terrain (for example forest), or in the case of unstable platforms (for example UAVs).
Lighting and the presence of shadows may have a strong effect on photogrammetry results. We deliberately did not adjust the flight plans to weather conditions. All the flights took place within a 2-day time interval; some of the flights were performed under cloudy conditions whereas others were performed with direct sunlight.
The flights were conducted on 29 and 30 January 2014 between 11:30 and 13:30 LT (local time) (solar noon on those days was around 12:40 LT). In addition to the 14 flights, 2 virtual flights (flight 15 and flight 16) were generated from 2 of the 14 flights to simulate the effect of image overlapping. Flight 15 was generated from every third flight line of flight 14. Similarly, flight 16 was generated from every third image from flight 11. These two additional virtual flights were created to (i) increase the number of “flights” used in the statistical analysis with different parameters and (ii), specifically, to investigate the effect of image overlapping on the quality of UAV imagery DEMs. This contributed to a more robust statistical analysis of the impact of UAV flight parameters on DEM quality (based on the selected DEM evaluation metrics). The parameters of all 16 flights are presented in Table 2.
Five ground control points were used to geo-reference the digital elevation models (see Fig. 3). The control points used were official survey points (LFP3) with a vertical accuracy of 3.7 cm and a horizontal accuracy of 3 cm. Since the points are protected with access covers, it was the access covers that were used for georeferencing the images. It was assumed that the cadastral points were directly underneath the centre of their cover. For this reason, the elevation difference between the points and their covers was measured and compensated for (Fig. 3), but any horizontal discrepancies were neglected.
Location used to calculate the metric values.
The settings that were used to generate the UAV DEMs with the Pix4D software for the steps of feature extraction, point cloud generation, and point cloud filtering are shown in Table 3. The reader can refer to the Pix4D user manual (Pix4D Support Team, 2014) for detailed information. For the assessment of the influence of UAV flight parameters on DEM quality, default settings were used. For the DEM comparison, settings were chosen through trial and error.
Co-registration of the UAV DEM with the lidar DEM is done implicitly by georeferencing the point clouds with the official survey points. By doing so, the generated UAV DEM is also georeferenced and can be directly overlaid with the lidar DEM, which is provided in the same coordinate reference system.
Figure 4 presents the locations surveyed to then allow for calculating the (a) qualitative and (b) quantitative metrics.
In this section we first present the results of the influence of parameters on DEM quality, and second the results from the comparison of the UAV DEM to the lidar DEM.
Pix4D settings that were used to generate the UAV DEM.
Results of the statistical models for the qualitative metrics. Bold
figures represent
Results of the statistical models for the quantitative metrics. Bold
figures represent
Relationship between the quality of the representation of building edges and flight altitude and between wall representation and weather conditions. The size of the dots is proportional to the number of observed metrics with identical quality class and altitude or weather condition.
The statistical models set-up for the qualitative metrics showed that, as expected, lower flight altitude produces better DEMs for overland flow modelling; lower flights tend to increase the quality of the DEM (Fig. 5a). Also, flights performed under overcast conditions led to better results (Fig. 5b), most likely due to the more uniform illumination and absence of hard and moving shadows. The influences of other flight conditions are clearly not significant; Table 4 contains the summarised statistical results.
Surprisingly, none of the quantitative metrics could have been related to the flight parameters. This may indicate that the variability of the metrics between flights with the same parameters is larger than the influence of the parameters; one can also say that the performance of the UAV is robust regarding the flight configuration. The results of these statistical models are presented in Table 5 (see also the visualisation in the supporting information). The significant result of the overcast weather condition for terrain elevation should not be over interpreted: first, only one flight was conducted under such conditions; second, the model suggests that the error is larger for overcast than for clear conditions, which is counterintuitive and contradicts the result for the qualitative metrics.
Other flight parameters than the ones considered in this study may have contributed to these results; these factors could be external, such as wind conditions and time of the day, or internal, such as the camera quality and operation mode. The camera mounted in the UAV is a modified point-and-shoot consumer camera; we expect that we would have observed larger differences if a professional camera had been used. For example, a better camera could have been operated with manual exposure, settings, and would have produced more equally exposed images. This alone could have substantially improved the identification of characteristic points. These additional factors may be worth further investigations (a different experimental design) that go beyond the scope of this study.
The objective of comparing the UAV DEMs and a nationwide available and commonly used DEM is to evaluate whether UAV DEMs have a similar or better quality, especially in the urban areas, which are relevant for overland flow modelling.
We expect that DEMs made available nationwide (e.g., data sets provided by
Swisstopo: the Swiss Federal Office of Topography
Because the two DEMs have different resolutions and we wanted to compare the two data sets on a pixel by pixel basis, we downscaled the UAV DEM to match the resolution of the lidar DEM, using the arithmetic average to compute new pixel values.
Qualitative (visual) assessment of DEM quality can use
Visual comparison of the UAV DEM and lidar DEM.
As a result, only a few areas below the vegetation can be regenerated in the photogrammetric point cloud. Because of the visual noise caused by overhead branches, the 3-D accuracy of the point cloud in these areas is compromised, which predisposes the points to be removed during the automatic point cloud filtering process. The tree-leaves-off conditions during the UAV flight in early March makes it difficult to identify matching points in the canopy/on bare thin branches, which are often less wide than the GSD. In our experiments with the image data set, it was fully possible to reconstruct the tree trunks and branches of many of the trees in the above-mentioned area, but it required an image overlap far superior than what is common for cartographic photogrammetry missions. Though not having the tree canopies represented is not a problem for overland flow modelling.
Apart from differences due to the presence and better representation of vegetation in the lidar DEM, there are also mobile objects such as vehicles that differ between the two scenes.
When looking at the insets, it appears that the quality of the two DEMs is very similar, with the exception that the lidar DEM has more noise and sharper edges than the UAV DEM. This can be at least partially explained by the averaging performed when downsampling the UAV DEM.
Because the two DEMs represent different seasons, there are a number of
differences between the two DEMs that are due to physical changes in the
environment and not due directly to the characteristics of one DEM
generation process or the other. Therefore, the comparison of the two
elevation data sets using the whole area is not meaningful. Due to this fact,
the comparison of the DEMs presented in the following sections will be
limited to a selected road area (area marked with the red line polygon in
Fig. 7). This area was defined based on visual analysis of the aerial
orthophoto associated to the UAV DEM. This area covers approximately
1500 m
The map of the elevation differences between the UAV DEM and the lidar DEM
(Fig. 7) was calculated subtracting the lidar DEM from the UAV DEM (Eq. 4)
with 2 m pixel
The red line polygon represents a road area selected based on visual analysis of the UAV orthophoto in order to quantitatively compare elevation, slope, and aspect of the two DEMs without the influence of objects such as vegetation, cars, or other features that are known to differ categorically between the two DEMs.
As can be seen in Fig. 8, the differences between the two DEMs in this
area are almost negligible. The minimum, maximum, mean, and standard
deviation of the elevation differences between the two DEMs are
The slope differences were calculated for the selected road area (see red
line polygon in Fig. 6a) using (Eq. 5) with 2 m pixel
As can be seen in Fig. 9, the slope differences between the two DEMs are
almost always below 10 %; it is noteworthy that the larger slope
differences are located along the boundary of the red-line polygon. The
value of the descriptive statistics of the slope differences between of the
two DEM are
minimum: maximum: 74.41 %; mean: standard deviation: 14.16 %.
The terrain aspect distribution of the selected road area of the two DEMs is
also very similar, as presented in Fig. 10.
Elevation differences between the UAV DEM and the lidar DEM (both
with 2 m pixel
Slope differences between the UAV DEM and the lidar DEM.
Flow paths were delineated using the conventional D8 flow direction
algorithm (Jenson and Domingue, 1988) for the three UAV DEMs at different
resolutions (0.5, 1.0 and 2.0 m pixel
Distribution of terrain aspect. The aspect values are in degrees. The outer number represent the cardinal directions in degrees.
In this study, we demonstrate the applicability and the advantages of using UAVs to generate very high resolution DEMs to be used in urban overland flow and flood modelling. To address this objective, we assessed (i) the influence of flight parameters in the quality of the DEMs produced using UAVs technology, and (ii) the quality of the UAV-based DEM in comparison to the conventional lidar-based DEM available in Switzerland. We concluded that
DEM-based flow path delineation.
UAV platforms and software are a mature technology that deliver basic data
leading to satisfactory results for urban overland flow modelling. Interestingly, only few dependencies between the flight parameters and DEM
quality could be identified. This might be due to variability introduced by
other external and internal factors not investigated in detail in this
study. Although, at first sight, this might leave only little potential for
optimal experimental design, at second sight this also means that the
technology is rather robust against flight altitude, camera pitch settings,
image overlapping parameters and thus suitable for practitioners. As expected, the most influential flight parameter was the flight altitude,
where Comparing the UAV DEM to a commonly available lidar-based DEM, we found that
the quality of both DEMs is comparable. The differences between the two DEMs
are not substantial, especially when the comparison is conducted in a
selected road area without cars, buildings, trees, or vegetation. When
comparing flow paths delineated using the different DEMs, it could be seen
that the flow paths obtained using a DEM downsampled (2 m pixel size) from
the finer resolution UAV DEM (0.05 m pixel size) retained the major flow
path patterns. The flow paths obtained using the lidar DEM were slightly
different from those obtained using the UAV DEMs; this is mostly due to the
presence of vegetation and trees in the first DEM. The UAV DEM has two
main/practical advantages over the lidar DEM, despite the similarities
mentioned above. First, it is more flexible to acquire elevation data using
UAVs, especially for small to medium size areas (or catchments), and the
second is that if the UAV flights are conducted during winter with
tree-leaves-off conditions, DEMs with no tree canopies represented can be
produced, which are especially beneficial for land use classification and
overland flow processes. It is, however, important to mention that there are
other solutions to generate DEMs other than nationwide airborne lidar-based
and UAV-based solutions, such as ground-based lidar. In particular,
a ground-based lidar solution is as flexible as the UAV solution, capable of
producing very fine resolution DEMs and may not have the problem of
obstruction by tree leaves as photogrammetric mini UAV solutions. However,
it also has disadvantages: the major one is perhaps related to the
limitation of covering areas located behind the buildings, i.e. it does not
allow for covering the whole area of interest (e.g. an urban catchment).
Our findings suggest that UAVs can greatly improve overland flow modelling
by increasing the detail of terrain representation and also by their
inherent flexibility to update existing elevation data sets. The very high
resolution that is possible to obtain using UAV DEMs is also an advantage for urban
overland flow and flood modelling purposes. Further research should be
carried out towards the development of an urban drainage modelling
application in order to assess the real benefit of using very high
resolution DEMs and hydraulic models.
In addition to the generation of DEMs, UAV imagery can also be used to
generate other very interesting data sets for urban drainage modelling
applications based on image classification. These are, for example,
identification of pervious/ impervious areas (Tokarczyk et al., 2015),
automatic identification and location of sewer inlets and manholes, and other
man-made features relevant to overland flow (Moy de Vitry, 2014).
The mini UAV platform used in the study is a fully autonomous fixed-wing
aircraft developed by senseFly SA
The specifications of the IXUS 127 HS camera part of the unmanned aerial system used in this study are presented in Table A2.
Detailed characteristics of the UAV.
Specifications of the Canon IXUS 127 HS.
The authors are grateful for the expert advice received from Konrad Schindler, ETH Zurich, during the development of this study, especially regarding photogrammetry. Edited by: G. Di Baldassarre