1. Introduction
Projected increases of heavy rainfall, based on climate models, are expected to aggravate local floods [
1]. Thus, effective spatial tools are required by governments and societies to take action against increasing exposure to natural hazards [
2]. Geospatial products, such as digital elevation models (DEMs) are useful topographic representations of space and have some specifics for flood studies [
3,
4]. In a further explanation, the DEM concept has been considered in the same way proposed by Polat et al. [
5], which refers to DEM as the
Z-dimension of the terrain digitally. There is also the digital surface model (DSM) that include natural and man-made objects. Highly detailed terrain modelling is usually produced from data obtained by active sensors such as airborne light detection and ranging (LiDAR) [
6,
7]. The bare ground representation in the form of DEM from these sources is the basis of urban [
8,
9,
10] and peri-urban local flood studies [
11]. LiDAR technology have as main advantages of its laser energy penetration to the ground, for instance, through canopies [
7]; however, the cost and complexity of the data acquisition involved implies that such airborne data is not always easy to update, or sometimes is only partially available [
5].
The development of photogrammetry techniques based on Structure from Motion Multi-View Stereo (SfM-MVS) of images acquired by low-cost cameras in unmanned aerial vehicle systems (micro UAV, ≤ 2 kg) has seen a strong development in the last decade [
12]. This, together with the SfM-MVS processing in a single workflow, allows the DSM and DEM generation [
13,
14]. Although one of its main technical drawbacks is the time required for image processing [
15]. However, the (relative) flexibility in image acquisition and the increasing offer of robust SfM-MVS processing software, have made UAV photogrammetry a valid low-cost alternative to piloted airborne LiDAR technology [
5,
14]. Studies show that image based UAV-derived DEMs are comparable to LiDAR for fluvial flood assessment applications [
16,
17,
18,
19], such as flood extent and volume estimations. Leitao et al. [
20] showed that is possible to obtain detailed DEMs in urban environments from image based UAV platforms with quality comparable to LIDAR data (in terms of the difference between DEMs), and found that a realistic representation (resolution < 1 m), plays a fundamental role in the surface flow modeling. Therefore, it concludes that micro UAVs are a useful solution for describing urban landscapes. In the literature there are more examples of the functionality of UAV images in 2D urban hydrodynamic modelling [
21,
22], flood risk management and emergency response [
23,
24], and mapping of difficult-to-access areas [
25]. It is widely accepted that the UAV-derived DEM accuracy from SfM-MVS, i.e., aerial or terrestrial photogrammetry processing, is influenced by flight design and planning factors, such as GSD (ground sample distance), inclusion (or not) of oblique images, sensor and camera lens, flight pattern and georeferencing method, etc. [
26]. As a rule of thumb in UAV photogrammetry, vertical accuracy for a DEM obtained must be between one and three times the GSD of input imagery [
27,
28,
29]. The impact of the georeferencing method on the accuracy of SfM-MVS products is critical, and also well established in the literature. Georeferencing is usually classified as: (i) direct, by means of UAV navigation (GPS/ IMU) instruments, and sometimes corrected in real time by GPS-RTK [
30,
31]; and (ii) indirect, through established ground control points [
32]. Usually, the classical indirect georeferencing is considered the most accurate method [
33]. Depending on the size of the UAV SfM-MVS project, ground control point determination can become a challenging task due to its time-intensive nature [
34], and constraints found on the terrain [
14].
Using existing elevation data, for example from airborne LiDAR, can be an alternative for georeferencing a UAV photogrammetric project. The literature shows the complementarity between LiDAR, as an alternative source for ground control points, and photogrammetry of airborne imagery [
35,
36,
37,
38]. Liu et al. [
35] and James et al. [
37] suggested the use of non-physical or virtual control points called "LiDAR-derived control points". This complementarity has been recently exploited with high-resolution imagery by a multi-rotor UAV platforms and terrestrial LiDAR data for 3D city modelling [
39]. Persad et al. have proposed the use of LiDAR data and SfM-MVS image processing for modelling applications and DEM generation in a deltaic area [
40,
41]. However, there is no reference that shows the contribution of LiDAR data in DEM generation from fixed-wing UAV imagery, especially in flood assessment applications for estimation of areas and volumes. To validate the present contribution, it is necessary to compare UAV photogrammetric products with independent external elevation data [
42,
43] and with standard reference surfaces (e.g., LiDAR) for flood analysis [
16,
20]. If this complementarity is confirmed, the use of existing airborne remote sensing data (e.g., LiDAR databases) will prove to be an alternative georeferencing method for UAV researchers and flood specialists in order to obtain useful DEMs for local-level studies.
This work aims to investigate if LiDAR elevation data can be used in DEM generation from fixed-wing UAV imagery for flood applications. More specifically, it aims to (i) assess the accuracy achieved in DEM from the SfM-MVS processing chain using LiDAR-derived control points (LCPs); (ii) test the performance of two software applications used for DEM processing and (iii); compare flood estimations of volume and area between DEMs based on UAV and LiDAR data (reference).
This paper is organized as follows:
Section 2 describes the equipment employed and the methods followed: UAV surveys, LCP collection, image processing for DSM and DEM generation (SfM software comparisons), flood applications and assessment methods.
Section 3 discusses the accuracy of the method, and then, after its validity is confirmed, the flood results.
Section 4 is devoted to the discussion, and
Section 5 presents the conclusions.
4. Discussion
Our results show that airborne LiDAR-derived control points are useful in obtaining accurate DEMs from UAV-based RGB imaging, with a resolution of two times the pixel size of input imagery. PhotoScan offers better interactivity, especially in DEM generation. Although its DSM accuracy turned out slightly inferior than Pix4D's, it was compensated when DEM is generated. The UAV-based DEMs are in fact as accurate as LiDAR DEMs, and this is in agreement with the work of Polat and Uysal [
5]. In general, DEMs obtained by a SfM-MVS processing chain are within the expected ranges reported in the literature (
Figure 8). This is also confirmed when comparing the relative accuracies with the ones in
Table A2 in the
Appendix A. Therefore, the input of control points from airborne LiDAR to SfM-MVS processing of fixed-wing UAV imaging is justified [
35,
36,
37]. Furthermore, our results contribute to broadening UAV photogrammetry applications when the determination of control points is a burden, for example, in emergency situations [
23,
24]. It also enables exploiting automatic integration, as shown in the literature [
38,
41,
73,
74]. In addition, it makes quick and efficient DEM generation possible, as well as to carry out multitemporal analysis, which is one of the main advantages of UAV platforms [
75]. Finally, based on the trends of the abovementioned literature and results, the increasing offer of geospatial products is promising, especially in order to achieve UN Sustainable Development Goal 11, “Sustainable Cities and Communities” by 2030 [
2].
Important limitations for the replication of the described method are the current international regulations for civil UAV operation, in particular, the flight altitude. However, by reducing it, similar or better accuracies (in relative terms) as those reported in the literature are to be expected at the expense of a smaller area coverage per flight (
Appendix A,
Table A2), and therefore, the need for longer flights [
76]. This requires that the LiDAR reference and the pixel size of the UAV images maintain a relative accuracy of at least 2:1. For example, for a 5 cm pixel (~150 m AGL using the same equipment), the LiDAR must have a vertical accuracy equal to or less than 10 cm. The ever-increasing availability of terrestrial LiDAR elevation data can become an additional source of control points for SfM-MVS UAV photogrammetry, as has been recently shown in the literature [
39,
77].
Results for flood estimations compared to LiDAR show the usefulness of DEMs generated from SfM-MVS dense point clouds, when the user is actively involved in their classification. These findings are in agreement with those of Leitão et al. [
20], Coveney et al. [
16] and Schumann et al. [
17], who based their comparisons on previous reference LiDAR surfaces. Outcomes of flood analysis showed the suitability of using DEMs from SfM-MVS as a tool to support local flood studies in urban catchments or peri-urban floodplains [
75]. Specifically, this allows obtaining useful input elevation data for 2D hydrodynamic modelling in urban areas, as suggested by Yalcin et al. [
21]. On the other hand, the estimation of extreme flood events makes it possible to investigate the generated DEM beyond the streets where the precise altimetric information was available. This warrants carrying out a more general UAV DEM assessment.
While discrepancies are evident between UAV models and the reference LiDAR, they are due to the DEM generation strategy, which is highly sensitive to the filtering method for ground point extraction and dense cloud edition. Flood map outputs show that for Pix4D DEM there exists a tendency of rendering a certain residual urban fabric, owing to deficient quality of the determination of the DEM by the software (
Figure 8). The inclusion of residual urban fabric in the Pix4D-derived DEM influences the flood extent by volume displacement. Consequently, in Pix4D DEM, flooding tends to propagate along the streets, in contrast to PhotoScan DTM, where a broad water surface is observed. This agrees with the conclusions of Shaad et al. [
78] and Hashemi-Beni et al. [
23], who showed that fully automated ground extraction algorithms generate worse flood estimates than those obtained from a manual or semi-automated classification. A thorough user knowledge of the area, together with the availability of additional field data (profiles and/or observations of flood depths) is essential to ensure adequate utilization of ground filtering algorithms [
79]. Discrepancies between flood assessments may also be due to the use of an outdated reference surface (e.g., LiDAR reference in 2008 vs. UAV surveys in 2016, as in our case), particularly in low-lying areas, where terrain variations or changes in hydraulic infrastructures (e.g., in channels or near box culverts) play an important role in flood propagation. The above suggests an opportunity to study terrain evolution with high-resolution UAV surveys [
19].
This paper only solved the processing of UAV raw data, but differences in the resolution of UAV DEM and reference LiDAR could have an impact on flood estimates [
16,
20]. Further works might focus on finding an optimal DEM resolution by resampling methods for flood comparisons between UAV and LIDAR data. Additionally, the elevation dataset of the presented case study can be applied for the implementation of a local early warning system to estimate possible flood volume detection and water distribution in micro-morphology of streets.
The main UAV advantage is their flexibility to acquire image data [
20], especially for small to medium size areas (< 1 km
2, up to 7 km
2, see
Table A2 in the
Appendix A). On the other hand, the major disadvantages of UAV technology include a limited coverage area (e.g., flight time, payload and weather conditions) and the requirements for data processing. Piloted airborne platforms are more suited up to national scale surveys, while UAV is naturally better suited for an urban sub-basin scale. From the viewpoint of data processing, the larger size UAV project, the longer the time to processing. Processing time effort is about 45% of the UAV workflow [
27], contrary to airborne LiDAR, in which 3D data is obtained automatically.
Economic analyses by Jeunnette and Hart [
80] concludes that piloted platforms lead to a lower cost of operation at 610 m AGL. The UAVs would become cost-competitive at approximately 305 m, flight height close to that used in the present study (~325 m AGL). Yurtseven [
76] confirms the above at 350 m AGL, and also, it would be providing reasonable vertical accuracy and minimize the potential for systematic errors such as "doming effect" on the elevation products.
The increasing operational capabilities of civil micro UAV will doubtlessly integrate with other technologies, as was here shown with LiDAR. However, current aviation safety regulations often pose limitations to research endeavors, especially because regulations seldom keep pace with technological development [
47]. This means that UAV operators, society, authorities and industry have to continue working together towards a continuous improvement of local regulations [
81,
82]. It is expected that innovations in UAV safety will allow the next generation of fully integrated platforms to enter the airspace [
83].
Finally, the use of LCPs as proposed in this work and according to James et al. [
37], might involve important limitations, including: (i) availability of LiDAR data, some regions have either partial coverage or none at all, (ii) resolution of LIDAR data should be sufficiently detailed to allow the human operator to identify the superficial features, (iii) loss of information due to interpolation from raw point cloud data to grid data, and (iv) possible variations due to the temporal differences between LIDAR data acquisition and UAV surveys.
5. Conclusions
In this study the contribution of existing altimetric data of airborne LiDAR in DEM generation from UAV-based images (10.3 cm pixel size) for flood applications was investigated. Georeferencing based on LiDAR-derived control points has been applied, DEM accuracy assessed, and further applied to flood estimations. Floods from the corresponding UAV DEM were compared to those from a LiDAR DEM reference.
The applied LCP georeferencing method contributes to obtaining DEMs with vertical accuracies comparable to those found in the literature, of approximately 2 times the pixel size of the input imagery. The DSM obtained with Pix4D is slightly more accurate than PhotoScan. However, the PhotoScan DEM is closer to the reference LiDAR, and therefore, more suitable for flood assessment applications (volume and area flooded estimations). In general, the feasibility of semi-automatically obtained UAV DEMs is confirmed. The hereby proved complementary nature between LiDAR and SfM-MVS photogrammetry will provide terrain modelers and flood scientists with an alternative tool for georeferencing their UAV (e.g., fixed-wing) photogrammetric products, in particular when ground control point determination is challenging.
The expected applications of micro-UAV systems and the increasing supply of LiDAR datasets are promising for floods studies at the local level. Future work might focus on assessing the DEM accuracy of detailed terrestrial LiDAR-georeferenced UAV flights, and testing the impact of spatial resolution on flood estimates.