Introduction

Water movement, generation of surface runoff and other hydrological processes depend upon surface topography. Digital Elevation Model is the surface elevation of any point in digital format. It models the continuous earth surface by discrete points at regular grid interval. The correctivenss of representation of continuous surface depends upon grid spacing. Various studies have been carried out using easily available SRTM DEM data at different resolutions. Most comparisons were done using a single source of DEM with changing grid size to lower the resolution. Coz et al. (2009) evaluated different method of aggregation of DEM data to generate different grid size images. Jenson (1991), Hutchinson and Dowling (1991) have compared results from DEM of 100 m and 1,000 m resolution. With 10 m USGS DEM, Wu et al. (2008) produced a set of DEMs up to 200 m grid size and found out that all the terrain variables tested vary with the grid size change. Li and Wong (2010) have shown that highest resolution data may not perform the best as the scale of data may not effectively capture the phenomenon under investigation or to be modeled. They also investigated the effects of digital elevation model resolutions (40 m, 50 m, 60 m, 70 m, 80 m and, 90 m) on elevation, number and area of sub-basins and area of watershed. Based on the analysis of digital elevation model at resolutions (40 m, 50 m, 60 m, 70 m, 80 m and 90 m), Xiao et al. (2010) found that data aggregation has little influence on elevation of the study area. He showed that there is definite correlation that can be found between number and area of sub-basins and DEM grid size in this study area.

Different elevation data sources have different resolutions and different accuracies. Interferometric Synthetic Aperture Radar system used on Shuttle Radar Topography Mission (SRTM) currently provides most complete and robust source of elevation data at the global scale (Schumann et al. 2008).

The launch of Cartosat – 1 satellite in 2005 with an aim to generate high resolution along track stereo imagery with B/H ratio 0.62 has made possible to generate DEM with high accuracy. Accuracy of DEMs generated from photogrammetric methods depends primarily on the quality of Ground Control Points (GCPs) used. GCPs of high quality leads to the generation of highly accurate DEMs (Krishna et al. 2008). Digital photogrammetrey techniques are used to generate DEMs of varying resolution. Previous studies have also indicated that different sources of DEMs have an influence on the delineation of hydrologic features (Guo-an et al. (2001), Schumann et al. (2008), Li and Wong (2010), Vaze et al. (2010). Hence, examining the effect of different DEMs on derived hydrologic features is of significant importance. Saran et al. (2010) generated hydrologic response units from different DEMs and proved that the DEM derived from optical stereo pairs (ASTER and CARTOSAT) provided higher vertical accuracies than the SRTM and topographic map-based DEM. Hydrologic and flood inundation modelling is essentially dependent on digital elevation models. However, in most of the hydrologic studies there is a significant influence of grid size (Zhang and Montgomery 1994), horizontal accuracy and vertical accuracy (Walker and Willgoose 1999) of elevation data.

The main objective of this study is to evaluate the consistency of Cartosat-I derived DEMs of different grid size generated using photogrammetric method for extraction of hydrologic features and comparing them with other sources like SRTM 90 m DEM and ASTER 30 m GDEM. Muralikrishnan et al. (2012) validated the Indian National CartoDEM and reported absolute height accuracy of 4.7 m in flat terrain and 7 m in hilly terrain. Different grid size data result in elevation variation. The effect of which could be analysed by extracting stream network form the DEMs which is of significant importance in delineating watersheds, surface runoff modelling etc. Considering the importance of DEMs in flood simulation studies an attempt has been made to simulate the flood event for different grid size of Cartosat-I derived DEMs and compare the inundation area results across the DEM sources.

Methodology

Study Area and Data

Damanganga Basin which lies in the Western Ghats Region in Western India is selected as study area. It is bounded by 19°52′ to 20°19′ North latitude and 73°2′ to 73°38′ East Longitude and is characterized by high and low level plateau areas containing majorly basaltic rocks. A large area coverage (1,812 sq km) was selected to generate DEM using number of scenes. The study area is chosen considering the complexity of the terrain and its influence on quality of generated DEM. Figure 1 shows the index map of the study area.

Fig. 1
figure 1

Index map of the study area

Full size image

The different DEM Sources considered in the study are Cartosat-I derived DEM, SRTM 90 m DEM and ASTER 30 m GDEM. Eleven Cartosat Stereo pairs were used for generation of DEM of the entire Basin. Vertical Ground Control Points (GCPs) were taken from CartoDEM chip (Reference chip) generated by ISRO. Survey of India (SOI) topomaps at 1:50,000 scale are used for drainage comparison.

DEM Generation Using Cartosat-1 Stereo Pairs

DEMs were generated from Cartosat-1 stereo pairs in Leica Photogrammetry Suite (LPS). The LPS 2011, which is a complete suite of photogrammetric production tools for triangulation, generating terrain models, producing orthomosaics and extracting 3D features (http://www.erdas.com/products/LPS/LPS/Details.aspx) was used for the identification of horizontal and vertical GCPs and tie points. Firstly, Cartosat −1 stereo pairs were added to the block and the respective Rational Polynomial Coefficient (RPCs) were attached to each of the images followed by computation of pyramid layer for each image. Block (or aerial) triangulation is the process of establishing a mathematical relationship between the images contained in a project, the camera or sensor model, and the ground. After the pyramid layers were created, GCP’s and Tie points were measured using Classic Point Measurement tool in LPS. GCPs are identifiable features located on the Earth’s surface whose ground coordinates in X, Y, and Z are known. A horizontal GCP only specifies the X, Y coordinates, while a vertical GCP only specifies the Z coordinate. Horizontal and vertical ground control points were taken from Cartochip (reference image) generated by ISRO which serves as reference DEM for entire country. Cartochip is available with varying quality and best quality cartochips were used to identify GCPs in 2D and 3D mode. 32 GCPs were identified in this process. Apart from GCPs, tie points are needed at overlap portions and other parts of the image. A tie point is a point whose ground coordinates are not known, but is visually recognizable in the overlap area between two or more images. Total 74 tie points were taken for the entire block. The block was triangulated with rmse of 0.32 pixels. Ground coordinates for tie points were computed during block triangulation.

DEM at varying grid size can be generated using two different methods. First method uses simple re-sampling method to generate DEM at different grid size from available DEM. Second method uses photogrammetric method to generate different grid size DEM using same GCPS and tie points obtained in the process of triangulation. Vertical accuracy of generated DEM varies with varying grid size while keeping the number of GCPs and tie points same during the process of triangulation due to the procedure involved in the analysis. Hence to understand the effect of different grid size DEM, photogrammetric method was used to generate CartoDEM at grid size of 10 m, 20 m, 30 m, 40 m, 50 m and 90 m. Each of the DEM was produced independently using same GCPs and tie points so that even at the common point elevation could be different.

Evaluation of Generated DEMs

Digital Elevation Model (DEM) are models that approximate a topographic surface and can have both random and systematic errors that create uncertainty or unreliability in elevation measurement from DEM (Holm 2000). The simplest method to evaluate DEM is by visual inspection, which many times may not give correct evaluation but major errors may be analysed using this method. Another method is to compare the generated DEM with some high quality DEM by statistical methods. Root mean square error is the most simple and commonly used method to describe the vertical accuracy of DEM which calculates Root Mean Square Error (RMSE). RMSE which is expressed in m is a standard deviation which assumes that the errors are random and their distribution is normal. It indicates how good the coincidence between DEM and reference DEM is. To evaluate the quality of generated DEM at different grid size, Cartochip (Reference DEM) was used. 200 random points covering all elevation bands were selected throughout the study area and the associated z-values for each of the point, from all the DEMs were extracted. All the elevation values are above WGS 84 datum.

Stream Network Extraction and Flood Simulation

Arc GIS 9.1 Spatial Analyst extension, hydrology tools (ESRI 2003) was used to extract stream network for each of the DEMs. Flow direction and sinks were identified from DEMs followed by flow accumulation. For extraction of drainage, a suitable threshold value needs to be decided which affects the extracted drainage lines and the catchment area. A larger threshold reduces the number of extracted streams, avoiding redundant or artificial streams, whereas applying a smaller threshold creates more stream features, some of which may not actually exist (Tarboton 1997; Tarboton and Ames 2001). Since the objective of the study was to evaluate IRS CatoDEM for stream network extraction and the available stream data available is from SOI topomap at 1:50,000 scale, a threshold value which generated 8th order stream (As available on topomap for the catchment) on 10 m grid size DEM was selected. Same threshold value was used over all the DEMs to extract stream network.

The methodology adopted by (Li and Wong 2010) was considered for carrying out flood simulation study. The software module used was Virtual GIS in ERDAS V9.1. The simplistic flood simulation involved the consideration of the flooding parameter i.e. the water stage and the inundated area. The flooding parameter was varied from 300 m to 600 m at 100 m interval and the variations in the inundated area for each of the DEMs were examined. The minimum flooding parameter was selected to be 300 m since the downstream region of the basin was found to have average elevation of 300 m. Similarly the maximum flooding parameter was selected to be 600 m since the upstream region of the basin was found out to have average elevation of 600 m. This simplistic flood simulation study considered no water movement. Hence the lateral movement of water, percolation of water into the ground and evaporation of water were not considered. Although this method does not actually simulate the flooding conditions but for the objective to compare different grid size DEM, it removes the effect of other parameters.

Evaluation of Stream Networks

Evaluation of stream networks was done based on raster based approach suggested by Pontius et al. (2008) and Li and Wong (2010) . The reference used was the stream networks taken from SOI Topomaps at 1:50,000 scale. The Correctness Index was calculated as

$$ \mathrm{CI}=\mathrm{N}\left( {_{\mathrm{A}}{\cap_{\mathrm{B}}}} \right)/{{\mathrm{N}}_{\mathrm{B}}} $$
(1)

where N(A∩B) is the number of cells found in both the extracted network and the drainage digitized from SOI topomaps as well. NB represents the number of cells in the drainage digitized from the SOI Topomaps. In order to take into account that how well the entire extracted network can resemble the entire reference network, Figure of Merit Index suggested by Li and Wong (2010) was calculated as

$$ \mathrm{FM}=\mathrm{N}\left( {_{\mathrm{A}}{\cap_{\mathrm{B}}}} \right)/\mathrm{N}\left( {_{\mathrm{A}}{\cup_{\mathrm{B}}}} \right) $$
(2)

Where N (A∩B) is the number of cells found in both the extracted network and the drainage digitized from the SOI topomaps and N (A∪B) is the number of cells in the two networks combined with overlapping cells being counted once.

Results

RMSE Statistic for Each DEM

Elevation values were noted down for randomly selected 200 points for all CartoDEMs, ASTER GDEM and SRTM DEM in different elevation bands. Cartochip was used as reference DEM for comparison of the results. The elevation values obtained by ASTER DEM and SRTM DEM were very high giving larger RMSE. Based on the elevation values, five ranges were selected as shown in Fig. 2. Band wise RMSE was calculated for different grid sizes. The RMSE value is low for all the grid size CartoDEM in <100 m elevation band. Lower grid size CartoDEM (10 m to 30 m) performed well till 400 m elevation range and the accuracy slightly decreased beyond 400 m elevation range. Although the RMSE increased from 400 m elevation onwards, all the CartoDEM performed similarly and hence the RMSE difference was less. Near similar performance by all the DEMs is evident from less rmse difference at >600 m elevation range in comparison to performance at elevation band of <100 m. The performance of 50 m and 90 m CartoDEM was less in comparison to others and more variation in the RMSE error was visible on 200–400 m elevation band. The RMSE values for ASTER 30 m and SRTM DEM are higher than other DEMs for all the elevation ranges indicating over estimation of elevation values on these DEMs. When compared to SRTM over Indian landmass, 90 % of pixels of CartoDEM reported were within ±8 m difference (Muralikrishnan et al. 2012)

Fig. 2
figure 2

RMSE statistics in different elevation bands for CartoDEMs at varying grid size, ASTER and SRTM DEM

Full size image

Stream Network Extraction

The extracted stream networks and the drainage digitized from SOI topomaps are shown in Fig. 3. The extracted networks are accessed for the positional accuracy in comparison to reference drainage digitized from 1:50,000 scale SOI topomap.

Fig. 3
figure 3

Stream networks generated from all DEMs and SOI Topomaps

Full size image

Visual inspection of the stream networks clearly shows that the networks derived through DEM accurately depict the main river channels. Lower order streams which are near the edge of catchments are not delineated on low resolution DEM. Even reducing the threshold to very low value, lower order streams could not be delineated on low resolution DEM. Very high threshold value also resulted in extraction of nonexistent streams on 10 m and 20 m DEM. The threshold value selected in the present analysis corresponds to 8th order drainage extraction on 10 m DEM for the purpose of comparison. No common threshold can give satisfactory result across entire study area and different grid size. Visual inspection of delineated drainage shows a very good comparison between 10 m CartoDEM and toposheet. These two networks are quite similar inspite of being from altogether different sources. Barring the omission of few lower order streams, 20 m DEM also depicts similar drainage. Overall pattern and structure of drainage lines is same for toposheet and Carto10m and 20 m DEM. Drainage density is highest for 10 m CartoDEM due to digital extraction of even very small drainage lines. In comparison to ASTERDEM 30 m, CartoDEM 30 m shows more curvilinear drainages similar to that on topomap with more drainage density. Similarly the drainage derived from 90 m CartoDEM has more drainage density than SRTM derived DEM. Many studies suggest that the pattern and structures derived from DEM of different source are quite dissimilar but in the present analysis no such major dissimilarity was observed. To study the effect of threshold on drainage density, drainage extraction was done for two threshold values. Figure 4 shows the drainage densities across DEMs for threshold 50 and 200.

Fig. 4
figure 4

Drainage densities from different DEMs at threshold values of 50 and 200

Full size image

The difference in drainage densities obtained using two threshold values is high for higher resolution DEM which goes on reducing with decrease in resolution. The difference in drainage densities is negligible for 90 m CartoDEM and SRTM DEM. Higher drainage density at lower grid size is due to extraction of very small nonexistent drainage lines which are nothing but potential flow lines. The drainage order from threshold 50 on CartoDEM 10 m became 9 due to all the additional extracted drainage, where as it remained at 6 on CartoDEM 90 m and SRTM DEM inspite of reducing the threshold further. For similar drainage order on topomap and CartoDEM 10 m, the results from 10 m and 20 m are quite comparable.

Evaluation of Drainages Derived from all DEMs

To evaluate the quality of extracted network, comparison with reference network derived from SOI topomap was done in raster format. The drainage lines from topomap were converted into ratser layer by choosing appropriate grid size. A constant single cell was taken for entire drainage irrespective of its order or width and so all the drainages have width as one cell size. This method has its advantage as well as disadvantage. Increasing the width of river while rasterization is likely to produce reasonable correspondence between reference image and extracted drainage. More confidence on positional accuracy is given when we select single cell size for rasterization. Nine sets of raster river grid at single cell river width were generated at different grid cell size. Available Digital Elevation Models (DEMs) from Cartosat, ASTER and SRTM were resampled on different gridcell size ranging between 10 m to 90 m. Correctness Index and Figure of Merit Index values were calculated for all the DEMs at all the grid cell size. Figure 5 shows the Correctness Index and Figure of Merit Index. It has been observed that the DEM cell size near to its original grid size shows better correlation in general. The relationship between effect of different grid cell size on correctness index of DEMs could also be seen from Fig. 5. From the figure one can divide the DEMs into two groups. First group which comprises of CartoDEM 10 m to 40 m shows initial increase in CI value with grid cell size till 40 m cell size. Beyond 40 m the correctness reduces which may be attributed to comparison of less number of delineated drainage (Fig. 3) with reference drainage. Second group which comprises of DEM above 40 m grid size, ASTER DEM and SRTM DEM show certain decreasing trend of CI with increase in grid cell size. 90 m CartoDEM, ASTER DEM and SRTM DEM perform poorly at all grid cell size. Performance of 40 m CartoDEM is good particularly for small grid cell size (<30 m). 10 m CartoDEM performed best with cell size 30 m. The best performance of DEM could be seen at cell size 20 m. At large cell size the performance of DEMs are inferior but comparable.

Fig. 5
figure 5

a Relation between correctness index and different grid cell size for CartoDEM generated at varying resolution. b Relation between Figure of Merit Index and different grid cell size for CartoDEM generated at varying resolution

Full size image

The study of correctness values provide many observations. Low resolution DEM data provide almost constant results across grid cell sizes. Performance of higher resolution data at high grid cell size is comparable with performance of low resolution data at similar grid cell size. As per Li and Wong (2010) higher resolution data will perform well only when the analysis is conducted at a cell size similar to the data resolution. In our study when the CartoDEM 10 to 40 m is resampled to 10 to 40 m, the best performance is observed at 20 m and 30 m.

The denomination in the Figure of Merit index includes the unmatched drainage from DEMs. Due to this the FM values are naturally lower than CI values but the trend observed in the study is slightly different. Almost for all the DEMs there is increase in FM with cell size till 40 m grid cell size. The FM values tend to reduce for all the DEMs beyond 40 m. CartoDEM 20 m and 40 m perform best among all. SRTM DEM and CartoDEM 90 m perform poorly.

Visual comparison of drainage delineated from different DEMs at different grid cell size from 10 m to 90 m is shown in Fig. 6. Delineated drainage is curvilinear in shape for the DEMs till cell size 40 m. DEM generated at lower resolution and resampled at lower grid size do not perform that well.

Fig. 6
figure 6

Visual comparison of drainage delineated from different DEM at 10 m and 90 m grid cell size

Full size image

Flood Simulation Results

To study the effect of flooding, inundation area maps were prepared between elevation 300 m to 600 m for all the datasets. The inundation area for all the case studies was calculated and displayed in Fig. 7. Inundation areas observed from SRTM and ASTER 30 m exhibit clear difference from others. The area coverage for SRTM is fragmented due to higher elevation recording by SRTM at all the levels. Wang et al. (2012) also showed similar results of SRTM in comparison to LIDAR and NED data. Results are closer at 600 m elevation as the average elevation difference reduced at this elevation range as shown in Fig. 2. Figure 7 also shows the relationship between inundated area, grid size and elevation bands. The inundation area increases with level of flood. There is very minor variation of around 6 % in inundated area derived from lower and higher grid size CartoDEM. Inundation area derived from SRTM DEM and ASTER 30 m DEM is about 35 % less at 300 m elevation and 7 % less at 600 m elevation (Fig. 7), hence overall the inundation is less from SRTM and ASTER DEM. It is evident from Fig. 7 that the grid size of CartoDEM has very little bearing on the flood inundation result. Higher resolution data record undulations in height more correctly than others and usually the results are better as also evident from Fig. 7.

Fig. 7
figure 7

Results of flood simulation studies on different DEMs and impact of grid size on inundated area

Full size image

Conclusions

In the present study, we evaluated three sources of DEM data by river network extraction and flood simulations study towards hydrological applications. From all the dataset used the expected best DEM was to be CartoDEM at 10 m and 20 m. Our findings also showed the CartoDEM 20 m to be overall best. High resolution DEM derived from Cartosat showed better results but surprisingly CartoDEM 40 m also performed very well from point of consideration of positional accuracy. High resolution data performs well at cell size similar or near to resolution. Performance of low resolution data is in general less. The elevation levels obtained from SRTM and ASTER DEMs have overestimated the elevation level when compared with CartoDEM reference chip. Higher resolution CartoDEM performed very well at lower elevation bands but the accuracy degraded for the study area in higher elevation bands. Inundation area does not seem to be linked to DEM resolution and variation is very little between results from 10 m CartoDEM and 90 m CartoDEM. Due to overestimation of heights by SRTM DEM and ASTER DEM, the inundation results obtained from these grids are lower than others.

The poor performance of DEM is not necessarily due to spatial resolution. The landscape is represented by smooth surface in lower resolution data but has its own advantage. The derivatives like slope which may not require high undulations could be estimated better with low resolution data at grid size similar to resolution. Very high resolution data will also include artificial artifacts which may pose hindrance in understanding terrain and its analysis.

In many applications, to match different other data sources, high resolution DEM is resampled to lower resolution (High grid cell size) with assumption that the higher resolution DEM will always work better than lower resolution DEM. Our study suggests that this assumptions should be used with caution and instead of resamping the DEM it is advisable to generate the DEM at specified resolution by photogrammetric procedure.

The study has its own limitations. Only one reference network was compared with all the data sets at different grid cell size inspite of our knowledge of delineation of less drainage network from lower resolution DEM. To understand the impact of cell size while rasterizing the topomap derived drainage, the index analysis was carried out for different river width rasterization of topomap derived drainage. Figure 8 shows the results from that analysis and as expected it showed strong correlation between accuracy and cell size.

Fig. 8
figure 8

Relation between Figure of Merit Index and different resolution CartoDEM at varying grid cell sizes

Full size image

For the applications considered, the CartoDEM 20 m at 20 m grid cell size showed reliable results in comparison to others closely followed by CartoDEM 40 m at 20 to 40 m grid cell size. The analysis included horizontal as well as vertical accuracy of DEM and the two hydrological applications. This also tells that high resolution DEM may not always perform upto expectations in overall representation and analysis, or in other words the relationship between DEM quality, resolution and grid cell size is not a straight one. The results of the analysis pertain to highly undulating terrain of parts of Damanganga basin. Towards verification of the results obtained in the study, more work will be carried out in future in different types of terrain.