Abstract
This work presents a new extension to BSplines that enables them to model functions on directed tree graphs such as nonbraided river networks. The main challenge of the application of Bsplines to graphs is their definition in the neighbourhood of nodes with more than two incident edges. Achieving that the Bsplines are continuous at these points is nontrivial. For both, simplification reasons and in view of our application, we limit the graphs to directed tree graphs. To fulfil the requirement of continuity, the knots defining the BSplines need to be located symmetrically along the edges with the same direction. With such defined BSplines, we approximate the topography of the Mekong River system from scattered height data along the river. To this end, we first test and validate successfully the method with synthetic water level data, with and without added annual signal. The quality of the resulting heights is assessed besides others by means of root mean square errors (RMSE) and mean absolute differences (MAD). The RMSE values are 0.26 m and 1.05 m without and with added annual variation respectively and the MAD values are even lower with 0.11 m and 0.60 m. For the second test, we use real water level observations measured by satellite altimetry. Again, we successfully estimate the river topography, but also discuss the short comings and problems with unevenly distributed data. The unevenly distributed data leads to some very large outliers close to the upstream ends of the rivers tributaries and in regions with rapidly changing topography such as the Mekong Falls. Without the outlier removal the standard deviation of the resulting heights can be as large as 50 m with a mean value of 15.73 m. After the outlier removal the mean standard deviation drops to 8.34 m.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
In geosciences, observations are often available either on the twodimensional (2D) surface of the Earth or as a onedimensional (1D) time series at a specific spatial location, or as a combination of both. Bsplines (basic splines) are often used to describe the physical relationships underlying these observations in a purely mathematical model. BSplines are piecewise polynomial basis functions with limited support and certain continuity properties (Prautzsch et al. 2002).
In the last decades a large number of modifications brought the 1D Bspline modelling approach to higher dimensions. To be more specific, whereas socalled endpointinterpolating Bsplines allow for modelling a 1D target function defined within a finite interval (Stollnitz et al. 1995a, b), trigonometrical Bsplines are defined on a circle. Consequently, the above mentioned 2D observations globally distributed on the Earth’s surface can be represented by a series expansion in tensor products of endpointinterpolating Bsplines for describing the latitude dependency and trigonometric Bsplines for describing the longitude dependency (Schumaker and Traas 1991; Jekeli 2005; Schmidt et al. 2015). As described e. g. by Schmidt (2012), the tensor product approach can even be extended to a representation of a multidimensional (nD) signal. As two examples, the vertical total electron content and the electron density of the ionosphere are represented as a 3D and a 4D series expansion (e.g. Durmaz and Karslioğlu 2011; Schmidt et al. 2015; Goss et al. 2019); other examples are rainfall interpolation (Gallo et al. 1998), air pollution (Leitenstorfer and Tutz 2007), and magnetic field modelling (Jiang and Zhang 2013) just to name a few.
In other applications continuous quantities are only observable along a network such as a river system or a road network embedded in a 2D surface. Air pollution in cities, for example, is often only measurable along the street network; the continuously changing water temperature of a river can be monitored only along the same (Laaha et al. 2013). Another example is the topography along river networks which is the topic of this work. The height of the mean water level and the topography of the shore line change continuously along the path of the river, water falls excluded. At river junctions, the height in all branches are continuous albeit more or less independent from each other.
In this paper we expand endpointinterpolating Bsplines to facilitate modelling on a river network. We prove the concept with the estimation of the topography along the Mekong River system from water level observations of satellite altimetry. Also Villadsen et al. (2015) and Bercher et al. (2013) estimated river topography from altimetry observations but only with low degree polynomial functions which is not feasible for a more complicated topography or large river systems. Also, neither of them considered tributaries of the river network. Zakharova et al. (2020) also used BSplines for the topographic modelling from altimetric observations, although only along a limited stretch of the main river. The here presented Bspline method has already been successfully applied to model the nonconstant unknown mean in Universal Kriging (Boergens et al. 2019).
In the sequel, we will first provide an introduction to endpointinterpolating Bsplines in Sect. 2 followed by our extension to Bsplines on directed tree graphs Sect. 3. In Sect. 4 we apply the new method to estimate the river topography for the Mekong River system. The manuscript closes with a discussion and a conclusion in Sect. 5 where also the application of the method to other possible graphs is considered.
2 Endpointinterpolating Bsplines
The normalised Bspline functions \(B^m_{k}(x)\) of degree \(m \in {\mathbb {N}}\) and shift \(k = 0,1, \ldots , K1\) depending on the variable \(x \in {\mathbb {R}}^1\) are calculable recursively via the relation
from the initial values
see e.g., Stollnitz et al. (1995a, 1995b) or (Schmidt 2007). The positive integer value K defines the number of given nondecreasing values \(t_0, t_1, \ldots , t_{K+m+1}\), denoted as knots. A Bspline is compactly supported, i.e. its values are different from zero only in a finite range on the real axis. Since \(B^m_{k}(x) \ne 0\) for \(t_k \le x < t_{k+m+1}\) and \(B^m_{k}(x) = 0\) otherwise, this finite range corresponds to an interval \([t_k,t_{k+m+1}]\).
For regional modelling, i.e. \(x \in [x_l, x_r]\), where \(x_l\) and \(x_r\) with \(x_r > x_l\) denote the left and right finite boundary values, we introduce the endpointinterpolating Bsplines of degree m ; see e.g. (Stollnitz et al. 1995a, b). For that purpose we set the first \(m+1\) knots to the value \(x_l\) and the last \(m+1\) knots to the value \(x_r\). Hence, the knot sequence for endpointinterpolating Bsplines of degree m is given as
where the distances \(\Delta x_k = t_{k+1}  t_k\) for \(k = m, \ldots , K1\) and \(m<K\) between two neighbouring knots could vary. Note that in Eq. (1), the factors are taken as zero if their denominators are zero. As can be seen from Fig. 1, the endpointinterpolation modifies the first and the last m Bspline functions which have a different shape than the other Bspline functions.
In the following we use the endpointinterpolating Bsplines of degree \(m=2\) – denoted as quadratic Bsplines – for \(k = 0, \ldots , K1\). Each of the these functions is different from zero only in the subinterval
and thus, called is a localising function. Considering the set of the altogether K Bspline functions the condition
holds. Figure 1 shows the \(K = 6\) Bspline functions \(B^2_{k}(x)\) within the interval \([x_l=0,x_r=10]\). The black dots on the xaxis mark the altogether 9 knots: the first 3 knots \(t_0 = t_1 = t_2 = x_l = 0\) at the left boundary, the regular knots \(t_3\) to \(t_5\) as well as the last 3 knots \(t_6 = t_7 = t_8 = x_r = 10\). The redcoloured Bspline \(B^2_{3}(x)\) is different from zero in the interval \([t_3,t_6]\) according to Eq. (4). As can be seen the distances \(\Delta x_k = t_{k+1}  t_k\) for \(k = 2, \ldots , 5\) are different, since the knots are nonequally spaced along the xaxis within the interval [0, 10].
3 Bsplines on directed tree graphs
Bsplines are well defined in \({\mathbb {R}}^1\) as was shown in the previous section, as well as to higher dimensions \({\mathbb {R}}^n\) for \(n > 1\) as was mentioned in Sect. 1; see also Schmidt (2012). In this paper we show how to use Bsplines on a directed tree graph G embedded in a metric space \(({\mathbb {R}}^2, d)\). We can think of d as the Euclidean or spherical metric. We will demonstrate this using the example of river topologies.
In order to apply our method to the river topography, we model the geographical river shape as a directed tree graph denoted as G.
The river network is usually provided as polygonal chains describing the shore line, i.e. both the left and right river bank are given. In a first step, these polygonal chains are skeletonised (Jiang et al. 2011). The reaches between two confluences are discretised with less than 1 km distance. This yields a planar embedding of a graph G such that vertices of the skeletons correspond to vertices of G. In addition, we introduce vertices for both the upstream ends and the downstream ends of all tributaries as well as for the downstream end of the river. We direct edges in flow direction and obtain a directed, acyclic, planar graph in which each vertex has outdegree 0 or 1. Outdegree is the number of outgoing edges in a vertex, i. e. an outdegree of 0 or 1 means that we do not consider multichannel and braided rivers. We call such graphs directed tree even though this term is usually used for digraphs which are directed in the opposite way. A vertex with more than one incoming edge is called junction vertex.
To realise Bsplines on a directed tree G, the knots \(t_k\) have to be placed on the branches of G, which is the part of G between two junction vertices. By identifying G with its image under the planar embedding, we define the BSpline function \(f: {\mathbb {R}}^2 \rightarrow {\mathbb {R}}\) as
where \(\alpha _k\) are the BSpline coefficients.
Figure 2 displays a schematic example of BSplines in the neighbourhood of such a junction. As can be seen in the figure, a Bspline centred below the junction can reach up into the upper branches and vice versa.
The difficulty of applying Bsplines on directed graphs arises at a junction as we aim at a continuity of the function f. As explained in Sect. 2 and shown in Fig. 1 the shape and width of a Bspline depends on the distance between the knots. Thus, to ensure a continuous function across a junction the knots have to be placed symmetrically on the two upper branches. As written before, a Bspline centred below the junction reaches in both upstream branches and the distance between the Bspline knots define the shape of the spline. Such a junction crossing BSpline (e. g. the red spline in Fig. 2) can be either defined by the knots downstream of the junction together with the knots on the left upstream branch or with the knots on the right upstream branch. However, the shape of the Bspline downstream of the junction should be the same for both alternatives. This can only be ensured by placing the knots upstream of the junction symmetrically on both branches. Additionally, a function f(x) (see Eq. 6) modelled with such defined Bsplines will be continuous both for the path downstreamleft upstream branch and for the path downstreamright upstream branch.
These considerations lead to the following rules for placing the BSpline knots \(t_k\):

1.
At the sink, according to endpointinterpolating Bsplines (see Sect. 2);

2.
At all sources, according to endpointinterpolating Bsplines as well;

3.
At each junction vertex;

4.
Along the branches equidistant to the next lower junction vertex with a fixed width.
The distances between each two knots have to be chosen carefully such that enough knots can be placed on each branch. The interval \([t_k,t_{k+m+1}]\) of a Bspline \(B^m_k(x)\) of degree \(m = 2\) covers \(m + 2 = 4\) knots and thus, at least \(m = 2\) knots need to be placed symmetrically on the two branches. Thus, the distance \(d_k\) between the knots needs to obey
where \(d_{b_i}\) is the length of the \(i{\text {th}}\) branch. The dense spatial resolution of G allows to place all knots on vertices with a reasonable accuracy.
4 Bsplines for river topography modelling
In this section we will demonstrate the usage of BSplines on graphs for modelling the river topography including tributaries. The mean river heights in a river basin cannot in general be described with any polynomial function due to a rapidly changing river slope. The topography of the main river and of the tributaries upstream of the confluence have to be modelled separately while ensuring continuity at the junctions.
In the here presented experiment we use satellite altimetry data of various missions to measure the river topography and to estimate the Bspline coefficient for the whole basin. As a test case we chose the Mekong River Basin in SouthEastAsia.
4.1 Method
For this study the Bspline knots t are automatically placed along the main river and the major tributaries of the Mekong River following the rules above (c. f. Sect. 3). The maximum distance between two knots is 30 km, which we found an appropriate compromise between the river reaches with dense and sparse data coverage. We excluded river reaches upstream of dams, as they constitute a discontinuity of the river topography. This results in 221 knots in the whole river basin. Figure 3(a) displays the knot placement in the whole river basin which is zoomed in to show the placement around two river junctions in Fig. 3(b). Most notably is the smaller distance downstream of the left junction compared to the distance of the knots in the two upstream branches. Whereas, around the right junction the distances between the knots in the two upstream branches are nearly the same to the downstream knot.
The such defined knots lead to the definition of BSplines following Eq. 1, i. e. 221 BSplines. With these BSplines we can interpolate any function on the graph G by estimating all BSpline coefficients \(\alpha _k\) with a sufficient number of observations of f(x) (c. f. Eq. 6). In the two following experiments, we have around 10 000 height observations for f(x) for which we assume equal uncertainty in a leastsquareadjustment.
4.2 Synthetic data experiment
4.2.1 Synthetic data
For the synthetic data set we simulate water level heights along the Mekong River and its tributaries based on the HydroSHEDS topography model (Lehner et al. 2008). The simulation consists of three steps: First, we choose to simulate the water levels at the same location as the altimetry observations in the second experiment (see Fig. 6). However, the location of the river in HydroSHEDS and the altimetry observation do not always coincide. Thus, for each point we use the minimum height in HydroSHEDS in a 2’ by 2’ window around the location. These heights are mostly located on the river but still do not form a more or less continuous height profile along the river. Thus, secondly, we employ these heights to estimate a first set of BSpline coefficients, with which we gain interpolated improved heights at the locations. The such derived heights are taken as the "true" river topography. Unfortunately, this method does not ensure monotonous rising heights along the river.
In the third and last step, we manufacture two synthetic data sets. For the first only noise is added to the true heights to simulate measurement uncertainties. We choose a constant normal distributed noise level with a standard deviation of 1 m. For the second synthetic data set, we add besides the noise an additional annual sinusoidal signal to the heights with a constant amplitude of 5 m. To determine the annual signal, every location gets a random date. It should be noted that the choice of seasonal amplitude has an influence on the quality of the results with higher amplitudes decreasing the quality. However, an amplitude of 5 m is for most parts of the Mekong River basin an upper bound to the amplitudes observed in reality.
4.2.2 Results
The two synthetic data sets, with and without annual signal, are used to estimate the BSpline coefficients which are then employed to model the river topography. The resulting heights of the data set with annual signal are displayed in Fig. 4 (a). We refrain from additionally showing the results from the data set without annual signal, as differences are not visible at this resolution. In Fig. 4 (b) the differences between the estimated heights and the true heights are shown. Most differences are below 1 m but can be as large as 9 m. The differences tend to be larger in reaches with sparse data coverage and at the upstream ends of tributaries. Besides this, no systematic effect is observable.
Further, we employed quality measures to compare the estimated heights with the true heights: the root mean square error (RMSE), correlation coefficient (\(\rho \)), the mean absolute difference (MAD), and both the standard deviation and maximum value (std(AD) and max(AD)) of the absolute differences (AD). The results of this validation is summarised in Table 1. For both synthetic data sets the correlation is nearly perfect. The RMSE is as low as only 26 cm for the case without added annual signal and about 1 m with added annual signal. The MAD values in comparison are lower showing that a few very large differences distort the RMSE to larger values. The std(AD) of the synthetic data without added annual signal is significantly smaller even while the maximum observed absolute differences are very similar for both synthetic cases.
Lastly, we investigate in more detail the height estimation in two river reaches: The Tonle Srepok in the SouthEast, and the upstream main river reach South of Luang Prabang (see Fig. 3). For these two river reaches, we plot the true heights, the synthetic heights, and the resulting estimated topography for both data sets in Fig. 5. The resulting topography estimated with the synthetic data without added annual signal shows nearly no difference to the true heights in both investigated reaches. On the other hand, the results of the data with annual signal reveals an interesting point. In both reaches some observations close by each other seem to have a similar phasing (Tonle Srepong between 500 km and 530 km and between 750 km and 780 km, upstream between 1580 km and 1625 km) which leads to a uniform shift of the data and, thus, the resulting estimated topography. Further, we can see the continuity of the estimated topography across the river junctions (marked in grey in the figure). Only at the second junction of Tonle Srepok a small step is visible. However, around the junction only few observations are present which might lead to this feature.
4.3 Altimetry data experiment
4.3.1 Altimetry data
We use river water level height data of the satellite altimetry missions Envisat (2002–2010), Envisat Extended Mission (EM) (2010–2011, drifting orbit), SARAL (2013–2016), SARAL Drifting Phase (DP) (2016present), Jason2 (2008–2016), Jason3 (2016present), and CryoSat2 (2010present). The missions Envisat, SARAL, Jason2, and Jason3 are socalled shortrepeat orbit missions which lead to water level height time series at fixed locations with a temporal resolution of 10 to 35 days but suffering on a sparse spatial resolution. The other missions are either on nonrepeat orbits (Envisat EM and SARAL DP) or on longrepeat orbits (CryoSat2, repeat time 369 days). This leads to a very dense spatial coverage.
All data have been processed with DAHITI (Database for Hydrological Time Series of Inland Waters, (Schwatke et al. 2015)), except for the CryoSat2 data which were processed following the method described by Boergens et al. (2017). The uncertainty, expressed as standard deviation, of the altimetric observations depends on the width of the river as well as the observing satellite mission and is estimated to be between 10 cm and 1.5 m.
Figure 6 displays the spatial distribution of the data in the river basin. Here the dense spatial distribution of the long and nonrepeat orbits missions (visualised by black symbols ) compared to the sparse distribution of the shortrepeat orbit missions (visualised by redcoloured squares) is visible.
4.3.2 Results
Fig. 7(a) shows the resulting estimated topography heights at all points of the topology. The BSpline modelled topography mostly fits well to the topography model. However, at some points the estimated heights do not correspond with the surrounding heights. Most prominently, many upstream end points of tributaries are not correctly modelled, which is likely caused by missing altimetry data in these reaches. Also in river reaches where only sparse altimetry data is available, especially with only a single observations, the quality of the estimated heights deteriorates (see Fig. 6) as the estimation of \(\alpha _k\) in Eq. 6 gets numerically unstable. In these parts a denser or more evenly distributed data coverage could improve the results. Additionally, the abrupt height change around the Mekong Falls (13.95°N, 105.94°E) cannot be modelled sufficiently with only the sparse data available. To correctly model such a reach a dense data coverage would be needed and possibly more Bspline knots to enable the modelling of small scale changes. Consequently, in these reaches also the estimated standard deviation of the heights increases (Fig. 7(b)). The standard deviations are estimated from the a posteriori uncertainties, as no globally reliable data uncertainties are available.
Based on these observations we removed points with a standard deviation larger than 25 m and points with less then 50 altimetry observations in their 50 km neighbourhood. However, this leads to gaps in the heights along the river. The result of this procedure w.r.t. the estimated heights is shown in Fig. 7(c). The mean of all standard deviations is 15.73 m before the outlier removal and 8.34 m afterwards.
As in the synthetic data experiment, we also investigate the estimated heights for two river reaches in more detail in Fig. 8. Two things are notably along the Tonle Srepok. First, the dip in river topography around 650 km which is not explained by the altimetric observations. As we saw the same decline in the synthetic data experiment this might be caused by the knot distribution in this region causing unexpected behaviour in the BSplines. Second, the step around the second junction is visible in the estimated topography, which we also already observed in the synthetic data results. Again, this might be caused by either the knot placement or the missing altimetry data around the junction. It might also be possible, that the river topography in changing that rapidly in this region such that the BSplines cannot model it sufficiently.
5 Discussion and conclusions
This study introduces a method which allows to apply Bsplines on directed tree graphs. This procedure allows for modelling a continuous quantity along graphs. Most notably, it ensures the continuity of the Bsplines around junction vertexes, i. e. vertices with more than one incoming edge. This is established by a symmetrical knot placement on the two incoming edges incident to the same vertex. To allow for a proper modelling of all sources and the sink of the graph we applied endpointinterpolating Bsplines.
We demonstrate the new method to the example of topographic modelling along a nonbraided river network. Thus, the knots defining the Bsplines have to be placed symmetrically on both river reaches upstream a confluence. The method was successfully tested to estimate mean water levels along the Mekong River system from both synthetic and altimetric water level observations. The synthetic data case allows a validation of the estimated river topography. The validation proves the suitability of the method for topographic modelling. Even if annual water level variations are not removed from the observation, the method is able to estimate an accurate topography. Further, we found in additional investigations that the noise level of synthetic altimetry observations if of minor influence to the quality of the resulting topography. This proves the stability of the proposed method. In the second investigation, we employed altimetric water level observations. Here the main challenge are outliers in the data set which needs to be removed. Further, for both the synthetic and real data set cases the unevenly distributed observation hinder the topography estimation in regions with sparse data coverage. Another problem of real altimetry data can be the systematic biases between different satellite altimetry missions. To this end, we used crosscalibrated corrections provided via OpenADB (https://openadb.dgfi.tum.de/en/) following the method of Bosch et al. (2014) to remove these biases from the altimetry data before hand. Alternatively, one could add an additional bias parameter for each satellite mission in the parameter estimation of the BSplines coefficients. However, this would imply that the biases remain per mission constant over space and time, which is not necessarily the case.
In this study we only considered Bsplines on directed tree graphs, but the question arises if it is possible to extend these ideas to more general graphs.
First, let us again consider a tree shaped graph (i.e. one or three edges in each vertex), but investigate the directionality of the whole graph. As mentioned in Sect. 3 the direction of the directed tree is of no importance for the proposed approach. In the next step, we can removeleave out the directionality altogether. Such an undirected tree graph might , for example, be used to model the density of biological specimens in a river network. So far, the directionality of the graph implies a hierarchy of the edges at each vertex with Bspline knots symmetrically distributed along two of the three edges. This is because BSplines centred below the junction reach into both upstream reaches, but BSplines centred upstream only reach into the downstream reach and not the second upstream reach. If the directionality is left out, Bsplines centred on any of the three edges reach into both other edges. Thus, the knots need to be placed symmetrically around the vertex on all three edges. With this we know how to place the BSpline knots locally around each junctionvertex but not how to combine the knot placement around several connected vertices. So far, in the case of the directed graph, the nonsymmetry of the knot placement on the downstream branch allowed to place the BSpline knots equidistant on each branch starting on its downstream end. This leads to a possible smaller distance between the last knot below the next junction and the junction (see the knots in Fig. 2). In the case of the nondirectional graph this is no longer a suitable solution.
A possibility for a nondirectional graph is shown in Fig. 9. With the aid of this example, we will provide you with an ideashow you how to proceed with the knot placement. Around each vertex (large coloured dots) we placed locally equidistant knots (small coloured dots). Let us assume, we choose the same distance d for the equidistant knot placement around all vertices (d will be later defined). In the case of an edge length larger than four times d (e. g. edge between red and yellow vertices) the BSpline knots are simply placed symmetrically around the red and yellow vertex. The remaining space on the edge is then filled up with as many knots as needed (Fig. 9 black dots). In the case of an edge length of exactly four time d (green and yellow vertices) we again place the knots symmetrically around each vertex. However, this time, one knot overlaps from each vertex (the greenyellow dot in the example). For even shorter edges, it is also possible to have two overlapping knots. This leads us to the definition of d which has to chosen such, that all edged can be divided in three or four d long pieces plus an optional remaining piece. Theoretically, it is not necessary to use the same d around each vertex although this simplifies the automatisation of the method.
The same idea can be even further expanded to cyclic graphs. As long as the knots are place symmetrically around the vertices the proposed method can be applied.
References
Bercher, N., Dinardo, S., Lucas, B.M., Fleury, S., Picot, N.: A review of cryosat2/siral applications for the monitoring of river water levels. In: ESA Living Planet 2013, vol 2013, pp 1–7 (2013)
Boergens, E., Nielsen, K., Andersen, O.B., Dettmering, D., Seitz, F.: River levels derived with CryoSat2 SAR data classificationA case study in the Mekong River Basin. Remote Sens. 9(12), 1238 (2017). https://doi.org/10.3390/rs9121238
Boergens, E., Dettmering, D., Seitz, F.: Observing water level extremes in the Mekong River Basin: the benefit of longrepeat orbit missions in a multimission satellite altimetry approach. J. Hydrol. 570, 463–472 (2019). https://doi.org/10.1016/j.jhydrol.2018.12.041
Bosch, W., Dettmering, D., Schwatke, C.: Multimission crosscalibration of satellite altimeters: constructing a longterm data record for global and regional sea level change studies. Remote Sens. 6(3), 2255–2281 (2014). https://doi.org/10.3390/rs6032255
Durmaz, M., Karslioğlu, M.O.: Nonparametric regional VTEC modeling with multivariate adaptive regression BSplines. Adv. Space Res. 48(9), 1523–1530 (2011). https://doi.org/10.1016/j.asr.2011.06.031
Gallo, G., Spagnuolo, M., Spinello, S.: Rainfall estimation from sparse data with fuzzy Bsplines. J. Geograph. Inf. Decision Anal. 2(2), 194–203 (1998)
Goss, A., Schmidt, M., Erdogan, E., Görres, B., Seitz, F.: Highresolution vertical total electron content maps based on multiscale Bspline representations. Annal. Geophys. 37(4), 699–717 (2019). https://doi.org/10.5194/angeo376992019
Jekeli, C.: Spline Representations of Functions on a Sphere for Geopotential Modeling. Technical Report, Ohio State University, Division of Geodetic Science (2005)
Jiang, B., Xie, Z., Wu, X., Wu, L., Zhou, L.: Topology constrain for river dimensionreducing generalization in landuse map. In: Proceedings  2011 19th International Conference on Geoinformatics, Geoinformatics 2011, (2011) https://doi.org/10.1109/GeoInformatics.2011.5980668
Jiang, Y., Zhang, N.: Applications of scattered interpolation with multilevel Bsplines to magnetic anomaly data. Near Surf. Geophys. 11(5), 579–585 (2013). https://doi.org/10.3997/18730604.2013016
Laaha, G., Skøien, J.O., Nobilis, F., Blöschl, G.: Spatial prediction of stream temperatures using topkriging with an external drift. Environ. Model. Assess. (2013). https://doi.org/10.1007/s1066601393733
Lehner, B., Verdin, K., Jarvis, A.: New global hydrography derived from spaceborne elevation data. Eos Commentarii Societatis philologae Polonorum 89(10), 93–94 (2008). https://doi.org/10.1029/2008EO100001
Leitenstorfer, F., Tutz, G.: Generalized monotonic regression based on Bsplines with an application to air pollution data. Biostatistics 8(3), 654–673 (2007). https://doi.org/10.1093/biostatistics/kxl036
Prautzsch, H., Boehm, W., Paluszny, M.: Bspline representation. In: Prautzsch, H., Boehm, W., Paluszny, M. (eds.) Bézier and BSpline Techniques, Mathematics and Visualization, pp. 59–75. Springer, Berlin, Heidelberg (2002)
Schmidt, M.: Wavelet modelling in support of IRI. Adv. Space Res. 39, 932–940 (2007)
Schmidt, M.: Towards a multiscale representation of multidimensional signals. VII HotineMarussi Symp. Math. Geodesy 137, 119–127 (2012). https://doi.org/10.1007/9783642220784_18
Schmidt, M., Dettmering, D., Seitz, F.: Using BSpline expansions for ionosphere modeling. In: Handbook of Geomathematics: Second Edition, pp 939–983, (2015) https://doi.org/10.1007/9783642545511_80
Schumaker, L.L., Traas, C.: Fitting scattered data on spherelike surfaces using tensor products of trigonometric and polynomial splines. Numerische Mathematik 60(1), 133–144 (1991)
Schwatke, C., Dettmering, D., Bosch, W., Seitz, F.: DAHITIAn innovative approach for estimating water level time series over inland waters using multimission satellite altimetry. Hydrol. Earth Syst. Sci. 19(10), 4345–4364 (2015). https://doi.org/10.5194/hess1943452015
Stollnitz, E., DeRose, T., Salesin, D.: Wavelets for computer graphics: a primer, part I. IEEE Comput. Graph. Appl. 3, 76–84 (1995a)
Stollnitz, E., DeRose, T., Salesin, D.: Wavelets for computer graphics: a primer. Part I and part II. IEEE Comput. Graph. Appl. 15, 75–85 (1995b)
Villadsen, H., Andersen, O.B., Stenseng, L., Nielsen, K., Knudsen, P.: CryoSat2 altimetry for river level monitoringEvaluation in the GangesBrahmaputra River basin. Remote Sens. Environ. 168, 80–89 (2015). https://doi.org/10.1016/j.rse.2015.05.025
Zakharova, E., Nielsen, K., Kamenev, G., Kouraev, A.: River discharge estimation from radar altimetry: assessment of satellite performance, river scales and methods. J. Hydrol. 583, 124561 (2020). https://doi.org/10.1016/j.jhydrol.2020.124561
Acknowledgements
We thank Daniel Rotter, who help with the mathematical description of graphs and proof read the manuscript. We also thank the two anonymous reviewers for their valuable remarks which helped us to improve the manuscript. The altimetry data including all corrections have been obtained from OpenADB (https://openadb.dgfi.tum.de/en/).
Funding
Open Access funding enabled and organized by Projekt DEAL.
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Conflict of interest
The authors declare that they have no conflict of interest.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Boergens, E., Schmidt, M. & Seitz, F. The use of Bsplines to represent the topography of river networks. Int J Geomath 12, 21 (2021). https://doi.org/10.1007/s1313702100188w
Received:
Accepted:
Published:
DOI: https://doi.org/10.1007/s1313702100188w