Skip to content
BY-NC-ND 3.0 license Open Access Published by De Gruyter Open Access December 30, 2016

Dem Local Accuracy Patterns in Land-Use/Land-Cover Classification

  • Wassim Katerji , Mercedes Farjas Abadia and Maria del Carmen Morillo Balsera
From the journal Open Geosciences

Abstract

Global and nation-wide DEM do not preserve the same height accuracy throughout the area of study. Instead of assuming a single RMSE value for the whole area, this study proposes a vario-model that divides the area into sub-regions depending on the land-use / landcover (LULC) classification, and assigns a local accuracy per each zone, as these areas share similar terrain formation and roughness, and tend to have similar DEM accuracies. A pilot study over Lebanon using the SRTM and ASTER DEMs, combined with a set of 1,105 randomly distributed ground control points (GCPs) showed that even though the inputDEMs have different spatial and temporal resolution, and were collected using difierent techniques, their accuracy varied similarly when changing over difierent LULC classes. Furthermore, validating the generated vario-models proved that they provide a closer representation of the accuracy to the validating GCPs than the conventional RMSE, by 94% and 86% for the SRTMand ASTER respectively. Geostatistical analysis of the input and output datasets showed that the results have a normal distribution, which support the generalization of the proven hypothesis, making this finding applicable to other input datasets anywhere around the world.

1 Introduction

A digital elevation model (DEM) is a 3D continuous representation of a terrain surface. In geographic information systems (GIS), a DEM can be represented in a grid of equally-sized cells, called raster, or in vector surface based on a network of interconnected triangles of various sizes, called Triangular Irregular Network (TIN) [1]. A DEM, like any other spatial data is subject to induced error that could result from data collection, representation, or analysis [2, 3]. The combination of all these errors is modelled as data accuracy. When using a DEM on a regional or nationwide scale, a single product cannot have the same accuracy equally in all the various sub-region, due the variation of the modeled terrain [4]. In order to be able to use a given product more efficiently, each sub-region must be evaluated separately.

One possibility to define sub-regions is based on landuse classification. Land use defines how humans are using the land in order to achieve a certain activity, whereas land cover defines the various physical materials that cover the surface of the Earth. While land use and land cover are often coupled and many times used interchangeably, there is a major difference between the two: land use refers to the purpose the land servers, while the land cover refers to the actual cover [5]. Nerveless, lands used similarly share some common characteristics [6], and can affect the accuracy of the terrain representation [7, 8].

The main objective of this study is to assess the existing and model altitude accuracy patterns in globally collected and generated DEMs related to land-use and landcover of the same area. A proposed procedure is presented to generate a vario-model of the local accuracy for any given DEM over any area of study, provided the required input data. This vario-model is intended to provide users of a DEM a more representative accuracy estimate of that DEM, compared to the conventional method of assigning a single root mean square error (RMSE) value for the whole area. The RMSE method was presented in previous studies such as [9].

At any specific location on the DEM, the global accuracy represents the difference between the actual and the modelled altitude at that location. On the other hand, the local accuracy represents the difference between the actual and the modelled variation of altitudes between that location and its surrounding. While theoretically, the global and local accuracies should have the same values, in reality they tend to be different. In fact, during the process of topographic data capture and data processing for DEM generation, several factors exist that will affect either accuracy separately from the other. Some of the major factors that would create a difference between the two accuracies are: The difference in height references between the model and the control points, including difference in geoid; and re-projection of the data from one coordinate system to another, especially if the two systems have different height references.

In the current tools available in the market today, the power of using GIS goes beyond visualizing a DEM as a terrain in 2D or 3D, to perform terrain based analysis. Most of these analyses, such as such as the slope and aspect, are considered focal ones rather than local. Their algorithms are based on processing each location and its surrounding in order to generate the derived dataset. Back to the main objective of this study, the proposed procedure focuses on the local accuracy rather than the global one as it is more significant for the interpretation of the results of the analysis.

2 Data and methods

This section covers the mathematical approach for calculating the local accuracy, a detailed description of the area of study and the input data, and the proposed procedure for analyzing the data and generating the vario-model in GIS.

2.1 The Local Accuracy

The local accuracy of the height between two points i and j is calculated as follows:

LocalAccuracy:LAij=|(ZjZi)(ZmjZmi)|(1)

where: Zj is the actual height at point j, Zi the actual height at point i, Zmj the modelled height at point j, and Zmi the modelled height at point i.

In reality, it is not practical to calculate the local accuracy between two points only; instead there should be a possibility to do this calculation for any number of points. This raises the following issues:

  1. For each point i, what is the criterion to choose the point j? If there are different criteria, which one is the best to choose from? And will this introduce additional uncertainty?

  2. If the selection criterion is reflexive, then if point j is chosen to calculate the local accuracy of point i, and point i to be chosen to calculate the local accuracy of point j. This reflexivity will result in pairs of points having the same local accuracy, which is misleading.

Instead of relying on a pair of points to calculate the local accuracy, this study proposes to adopt for each point the average local accuracy with respect to its surrounding points, as follows:

AverageLocalAccuracy:ALAi = j=1nLAijn(2)

Where: LAij is the local accuracy (1), and j is any elements from the subset of points that surround point i.

The proposed model must work properly on any set of points, irrespective of their densities and distribution patterns. Therefor the surrounding points cannot be selected using a fixed search radius nor using a fixed number of closest points, as the first method would result in some points having no surroundings, and the second in including points that are too far from each other. Instead, the proposed work, with an inspiration from the Natural Neighbor spatial interpolation, relies on the Voronoi tessellation to select the neighboring points to calculate ALA.

By definition, Voronoi tessellation is a mathematical method created by Georgy Voronoi, to divide a space into a set of sub-regions based on an input set of points. In a 2D system, these sub-regions are defined as a set of adjacent polygons, where each polygon defines the area that is closest to a point with respect to others [10]. Assuming that for each point i, the Voronoi polygon is defined as Vi, the neighbouring points will be chosen as those having their Voronoi polygons adjacent to the one of i. Accordingly, for a sample point i, the set of neighboring points are shown in Fig. 1:

Figure 1 Sample Voronoi decomposition, showing the neighboring points of a point i
Figure 1

Sample Voronoi decomposition, showing the neighboring points of a point i

2.2 Area of Study

The area of study for this research is Lebanon. Lebanon is situated on the eastern edge of the Mediterranean Sea. As a geographic extent, it is very interesting to study because, even though it is a very small country with an area of 10,452 km2, ranked 167th in the list of countries by area out of 249 countries, it has a rich variation of landscape: it consists of a narrow coast, followed by a rough chain of mountains (Western Chain), followed by a plain valley (the ‘Beqaa’ valley), then a less rough chain of mountains (Eastern Chain) toward the eastern side.

2.3 Input Data

For the proposed method to be performed, three inputs are required: a DEM, a land use – land cover (LULC) dataset, and a set of ground control points (GCPs).

DEM: The input DEM should be in raster format. The used DEMs in this study are the Shuttle Radar Topography Mission (SRTM) and Advanced Spaceborne Thermal Emission and Reflection Radiometer Global DEM (ASTER GDEM) with 90 meters and 30 meters spatial resolution respectively (Figure 2). SRTM was collected based on the technique of Interferometric Synthetic Aperture Radar (In-SAR), which is one technique of radargrammetry. The height values are based on Earth Gravitational Model of 1996 (EGM96) geoid, and the dataset is referenced horizontally to the World Geodetic System of 1984 (WGS84) ellipsoid [11]. ASTER GDEM was generated using stereo-pair images collected by the ASTER instruments, which is one of major photogrammetric techniques to generate DEMs or any topographic data [12]. Similar to SRTM, the height values are based on WGS84 EGM96 geoid, and the dataset is referenced horizontally to the WGS84 ellipsoid [13].

Figure 2 Hillshade generated from the ASTER showing the topographic representation of the terrain
Figure 2

Hillshade generated from the ASTER showing the topographic representation of the terrain

Statistically, the SRTM height values span between -24 and 3,080 meters, with an average height of 1,005.19 meters and a standard deviation of 618.74 meters. Similarly, the ASTER GDEM values span between 0 and 3,076 meters, with an average height of 995.74 meters and a standard deviation of 620.19 meters. Comparing the two DEMs shows that the SRTM has a slightlybigger range compared to ASTER, and a slightly higher average too. Additionally, calculating the correlation between them shows that there a high positive correlation, with a value of +0.9974, meaning that the height values from both DEMs are coherent and they represent to a high extent the same terrain.

GCPs: The input GCPs should be in vector point format. At least one GCP must be available within each of the available LULC classes within the area of study. A set of 1,105 GCPs were collected from various existing surveying projects, but all using the same land surveying technique based on total stations with reference to the Lebanese Geodetic Network. The height values are based on the Ohio State University 1991 Global Geopotential Model (OSU91) geoid, and the dataset is referenced horizontally to the Double Stereographic projection that is based on Clark1880 ellipsoid. According to the levelling standards in conventional topographic surveying [14], the maximum allowed error is approximately 1.2 centimeters of height per 1 kilometer of distance between the measured point and the reference one; with a maximum allowed spacing between the points being 3 kilometers, the minimum accuracy of the height of each GCP is about 34 centimeters. Given their high accuracy, the error in these points will be neglected, and will be used as a reference for this study (Fig. 3).

Figure 3 Distribution of the GCPs in the plane
Figure 3

Distribution of the GCPs in the plane

Statistically, the GCPs height values span between 0.16 and 2,888.12 meters, with an average height of 565.14 meters and a standard deviation of 455.85 meters. The distribution of the GCPs in the plane is not uniform, and the height values does not form a normal distribution, as shown in the Normal QQ-Plot in Fig.4. While such distribution will prevent a conventional study from generating a statistically confident result, the proposed procedure will not be affected as it is based on sub-regions, instead of the whole country.

Figure 4 Normal QQ-Plot for the heights of the input GCPs
Figure 4

Normal QQ-Plot for the heights of the input GCPs

LULC: The input LULC should be in vector polygon format, with one field containing the name or unique code of the LULC classification, as per the European’ Coordination of information on the Environment (CORINE) standard (Fig. 5). Other known standards could have been used, such as the US National Land Cover Database (NLCD) as described in [15], however for this study, CORINE was chosen as Lebanon shares more common geographic characteristics with the European continent than with the American one. The distribution of the area of study as per CORINE standard level 1 and level 2 is shown in Table 1:

Figure 5 LULC Level 1 Classification as per CORINE Standards
Figure 5

LULC Level 1 Classification as per CORINE Standards

Table 1

Total Area per LULC Class

L1L2Area (km2)%
Artificial Surfaces648.676%
Artificial, non-agricultural vegetated areas2.680.03%
Industrial, Commercial and Transport Units47.560.5%
Mine, Dump and Construction Sites111.84r1%
Urban Fabric486.595%
Agricultural Areas3,327.9732%
Arable Land1,588.3615%
Heterogeneous Agricultural Areas134.241%
Permanent Crops1,605.3716%
Forest and Semi-Natural Areas6,263.6761%
Forests1,366.0013%
Open Space with Little or no Vegetation3,681.5036%
Scrub and/or Herbaceous Vegetation Association1,216.1812%
Wetlands4.770.05%
Inland Wetlands2.570.03%
Maritime Wetlands2.200.02%
Water Bodies14.340.1%
Inland Waters14.330.1%
Marine Waters0.020.00%

As shown in the table above, out of the fourteen level 2 classes, the six most dominant ones make 88% of total area, as such: Open spaces and bare including mountains that get covered seasonally by snow (32%); Permanent crops including mainly involve fruits trees, olive trees, and vineyards (16%); arable lands including seasonal varying agriculture (15%); forests (13%); and scrubs and herbaceous vegetation (12%).

2.4 Exploratory Analysis of the Input GCPs

Before using the previously described GCPs to build the proposed procedure, this section statistically analyzes them, in comparison with the other input datasets, to understand the relations that are available between them.

Analysis of Variance (ANOVA) showed that the actual altitudes are highly affected by those of the two models, where the result is F (1.639, 1809.537) = 1500.17 and p <0.05. The assumption sphere of the affected models is (χ2 = 274.338, p <0.05), and therefore the degree of liberty of these measurements are corrected to the estimation of Greenhouse-Geisser sphere. Comparing each pair of altitude measurement shows that there exist significant differences between them; the difference between the actual altitude and the modelled from SRTM is (t (D) = 18.44, p <0.05), and the modelled from ASTER is (t (D) = 43.95, p <0.05). The difference between SRTM and ASTER is altitudes (t (D) = 42.15, p <0.05).

Additionally, a comparison between the mean altitudes and the LULC classification using the Factorial Analysis ofVariance Joint Design tool in ANOVA. The performed analysis resulted that there are significant differences between the means of each LULC class: F (10.1094), p <0.05, η2 = 0.2, and the three altitude values behaved somehow similarly.

2.5 Proposed Procedure

The proposed procedure is developed as a script that automates the steps to calculate the local accuracy as mathematically defined in section of local accuracy. The script is developed using Python on top of Esri ArcGIS 10.2. As described next, the procedure is divided into three main steps, and it can be performed using any GIS software package with raster-based spatial analysis tools:

  1. Extract Overlay Data: For each GCP in the dataset, the overlaying data from SRTM and ASTER DEMs, and the LULC dataset are extracted and stored in additional fields in the GCPs table of attributes. It can be achieve by creating two new fields in the layer table of attribute, name Zm for the modeled height from the DEM, and L2 for the LULC class from the LULC dataset. Then the values are stored using spatial join between each of the related datasets and the GCPs one. Figure 6 shows the resulting table of attributes for the GCPs layer when assessing the SRTM DEM.

  2. Calculate Local Accuracy: For each GCP in the dataset, the average local accuracy (ALA) is calculated as described in section “Local Accuracy”. For consistency, count distribution between model it must be implemented through scripting automation with the following logic, and the results are shown in Fig. 7:

    1. Generate the voronoi polygons layer that defines the neighborhoods as per the geographical distribution of the points. In ArcGIS, the tools is called Create Thiessen Polygons;

    2. For each GCP, search for the neighboring GCPs that have adjacent neighboring voronoi polygons. This can be achieved by using the Select By Location tool, and using the condition “Touches Boundary”;

    3. For each pair of the central and neighboring GCPS, calculate the local accuracy (LA) as defined in equation 1, using the mathematics library in the scripting language;

    4. For that central GCP, calculate the average local accuracy (ALA) as defined in equation 2, using the same library as in (iii).

  3. Select randomly 80% of the GCPs to generate the model, leaving 20% for validating it as per the process previously described. However, those points are not chosen totally random, but in a way that 80% are chosen from each LULC class. The table below summarizes the total number of GCPs available per LULC class, and their count distribution between model generation and model validation:

  4. Compile Data per LULC class: Statistically calculate the average difference per LULC class, and assign its regions this value as a continuous raster covering the whole area of study. This step can be achieved using the following sub-steps:

    1. Calculate the average ALA for each LULC class, using the Summarize tool; an output summary table is generated;

    2. Join the output table to the LULC layer using the Join Field tool, and the L2 field as a common field between the table and the layer;

    3. Convert the LULC layer from vector polygon into a raster using the Polygon to Raster tool, and specify the ALA field for the output raster values. For spatial consistency, defined the spatial resolution of the output raster the same as the related input DEM.

Figure 6 Screenshot of the GCPs table of attributes when assessing the SRTM DEM
Figure 6

Screenshot of the GCPs table of attributes when assessing the SRTM DEM

Figure 7 Screenshot of the GCPs table of attributes, highlighted the resulting ALA values per GCP
Figure 7

Screenshot of the GCPs table of attributes, highlighted the resulting ALA values per GCP

Table 2

Distribution of GCPs per LULC L2 Classes

L2ClassTotal GCPsModel (80%)Validation (20%)
11Urban Fabric21817444
12Industrial, Commercial and Transport Units36288
13Mine, Dump and Construction Sites20164
21Arable Land35328271
22Permanent Crops21216943
24Heterogeneous Agricultural Areas25205
31Forests38308
32Scrub and/or Herbaceous Vegetation Association38308
33Open Space with little or no Vegetation13010426
41Inland Wetlands321
51Inland Waters32257

2.6 Model Assessment

In order to assess the accuracy of the proposed model, and how much it enhanced the modeling of accuracy compared to the traditional methods, the height accuracy of both DEMs were computed at every GCP, then these spot accuracies are compared to the three different models:

Assumed accuracy: as defined by the DEMs’ data providers. For SRTM, the assumed accuracy is 20 meters [16], for ASTER ranges between 10 and 25 meters [17], averaged to 17.5 meters;

Average accuracy: computed as the average value of the accuracies of the GCPs. The average accuracy is computed as the room mean square error (RMSE) of all the GCPs participating in the model generation, which is equal to 9.65 meters for SRTM and 13.58 for ASTER.

Proposed vario model: computed as previously described. The result is an average accuracy per each LULC class, as summarized in the table below:

The deviations from the three defined models is calculated as the difference between the individual local accuracy calculated for the GCP, and the collected values as follows:

AccuracyDeviation=ALAGCPALAModel(3)

3 Results

The output of the proposed procedure in section “Proposed Procedure” is two raster datasets named vario-models, one representing the local accuracy of SRTM (Fig. 8), and one of ASTER (Fig. 9). The SRTM vario-model has a spatial resolution of 90 meters and a varying accuracy from 1.48 meters to 15.59 meters, and an average accuracy of 9.70 meters. On the other hand, the ASTER vario-model has a spatial resolution of 30 meters and a varying accuracy from 8.69 meters to 20.28 meters, and an average accuracy of 14.34 meters.

Figure 8 Relative accuracy vario-model for SRTM DEM
Figure 8

Relative accuracy vario-model for SRTM DEM

Figure 9 Relative accuracy vario-model for ASTER DEM
Figure 9

Relative accuracy vario-model for ASTER DEM

Calculating the correlation between the two vario-models, using the multivariate tools in ArcGIS, shows that both datasets have high positive correlation, with a value of +0.89. This high correlation proves that both DEMs’ accuracies behave similarly, even though the actual values are different; in fact, SRTM showed better accuracy than ASTER.

To test the validity of the vario-models, whether it provides a more representative model of the accuracy compared to the assumed and RMSE values, the deviations of the individual GCPs are calculated as per the proposed procedure, and compared. The summary of these deviations is shown in the below table, and visualized in the histograms in Fig. 10.

Figure 10 Histograms showing the frequency distributions of the deviation of the GCPs accuracies from: (a) SRTM from assumed accuracy, (b) SRTM from RMSE, (c) SRTM from vario-model, (d) ASTER from assumed accuracy, (e) ASTER from RMSE, and (f) ASTER from vario-model
Figure 10

Histograms showing the frequency distributions of the deviation of the GCPs accuracies from: (a) SRTM from assumed accuracy, (b) SRTM from RMSE, (c) SRTM from vario-model, (d) ASTER from assumed accuracy, (e) ASTER from RMSE, and (f) ASTER from vario-model

The results below in Fig. 10 show that for both SRTM and ASTER, the RMSE models are more representatives than the assumed accuracies, where the average deviation is closer to zero (meaning no deviation), by 78% and 69% respectively. Also, the proposed vario-models showed even more representative modeling of the accuracy, where the average deviation is closer to zero from the RMSE model by 94% and 86% for the SRTM and ASTER respectively. Furthermore, analyzing the deviations from the vario-models using geostatistical analysis tools such as the Normal QQ-Plot, shows that these deviations form a normal distribution, as shown in the two graphs in Fig. 11:

Figure 11 Normal QQPlot for the deviations from: (A) SRTM Vario-Model, and (B) ASTER Vario-Model
Figure 11

Normal QQPlot for the deviations from: (A) SRTM Vario-Model, and (B) ASTER Vario-Model

4 Discussion

As shown so far in this study, the proposed vario-model proven to be a more representative way of modelling the accuracy of the DEMs, compared to their assumed accuracies, and to the conventional method of calculating a single RMSE. Furthermore, as shown in Fig. 11, the deviations from the proposed model have a normal distribution. Therefor a general deduction can be made that the accuracy of a DEM do not vary totally randomly, but follows a pattern according to the classification of the proposed vario-model. Given that this vario-model was based on the different LULC classes, it can be deduced that the accuracy of a DEM follows a pattern that is related to the type of the surface, mostly defined by the LULC classification.

It is important to note that the objective of this study was not to compare the SRTM and ASTER DEMs accuracies in order to prove which one of them is better to use; instead these two datasets were used as an example to test and validate the hypothesis that the accuracy of any DEM has a variation pattern related to the classes of LULC. In fact, these two dataset have different spatial and temporal resolutions, were collected and generated using difierent techniques, and yet their accuracies showed similar behavior for same types of LULC, as the correlation shows in the section ‘Results’.

Back to the raised issue in describing the input GCPs in the section ‘Input Data’: these points do not have a uniform distribution in the plane, and their height values do no follow a normal distribution, which questions the validity of any derived model when assessed from a conventional point of view. It is important to note the following:

  1. Using the Near tool in ArcGIS shows that the distances between any GCP to its nearest have a range of 0.15 to 10,334.74 meters, with an average of 628.62 meters. According to (ASTER Home Page, 2004), the average altitude of the ASTER sensor was 750 kilometers above Earth when it was measuring the surface topography. Comparing the distances between any two GCPs on the ground to these points’ distances to the ASTER sensor that took the measurements, shows that the ground distances between the points are negligible. In fact the average distance between two GCPs makes only 0.08% of distance to the sensor. Even the largest distance between two GCPs makes only 1.4%, which can be neglected without causing any major effect into the calculations.

  2. Using the Multivariate Statistics tool in ArcGIS shows that there is a low or even no correlation between the height of a control point and its local accuracy. In fact the correlation factor between the ASTER DEM and its accuracy vario-model is +0.22, between the SRTM and its accuracy vario-model is +0.20. Accordingly, the actual height value of a GCP does not have a direct effect on the proposed model, and having a normal distribution in the heights or not will not affect the results found in this research. On the other hand, the results are highly dependent to the individual local accuracy for each GCP; these values follow a normal distribution for the ASTER, and ‘almost’ normal distribution for the SRTM, as shown in the below graphs in Fig. 12.

Figure 12 Normal QQ-Plot for the GCPs local accuracies of the model: (A) SRTM and (B) ASTER
Figure 12

Normal QQ-Plot for the GCPs local accuracies of the model: (A) SRTM and (B) ASTER

The proposed vario-model combines in its originality two aspects when representing the accuracy of a DEM: Defining the local accuracy rather than the global one, and detecting patterns in the accuracy related to the type of land coverage. While, as far as the authors researched previously published articles, no similar models have been proposed yet. However, other previous studies accounted the different types of land covers when assessing the accuracy of a DEM, even if sometimes not to a formal LULC classification. One of these studies is done by [18] in Turkey, showed that the ASTER GDEM have high accuracy in mining areas and low agricultural fields, and lowest accuracy in industrial areas. While the exact figures differ between the two studies, both of them showed similar behavior, where going back to Table 3 in section ‘Model Assessment’ shows that the ASTER GDEM has a lower accuracy in industrial areas (L2 Class = 12) compared to the agricultural fields (L2 Class = 21).

Table 3

Summary of the local accuracies per LULC class, expressed in meters

L2ClassASTERSRTM
11Urban Fabric11.948.04
12Industrial, Commercial and Transport Units13.677.71
13Mine, Dump and Construction Sites18.6313.65
14Artificial, non-agricultural vegetated areas17.759.80
21Arable Land10.124.35
22Permanent Crops10.885.86
24Heterogeneous Agricultural Areas8.696.38
31Forests17.5914.60
32Scrub and/or Herbaceous Vegetation Association20.2815.59
33Open Space with Little or no Vegetation14.9210.22
41Inland Wetlands9.611.48
42Maritime Wetlands9.611.48
51Inland Waters9.402.56
52Marine Waters9.402.56
Table 4

Summary of the accuracy deviations, expressed in meters

DEMDeviation From:MinMaxAvg.Std.
Assumed−20.0020.94−13.196.36
SRTMRMSE−9.6531.29−2.846.36
Vario Model−14.4530.19−0.185.75
Assumed−17.5013.75−5.676.79
ASTERRMSE−13.5817.67−1.756.79
VarioModel−18.6620.37−0.256.45

5 Conclusion

This study demonstrated that a DEM accuracy follows a pattern with respect to the land cover type, and proposed a vario-model to represent the variation of the local accuracy rather than the global one, given it is more significant when analyzing that DEM, and generating derived datasets. While the proposed model is applicable to any DEM anywhere around the world, provided the required inputs, the pilot study was based on two global and freely available DEMs: SRTM and ASTER. Also, a set of randomly spread 1,105 GCPs around the Lebanese territory, the pilot area of study, where used to assess the DEMs, generate their related vario-models, and test the validity of the results.

The conventional RMSE model showed a better representation of the accuracy than the assumed ones from the DEMs’ websites by being closer to the actual accuracies of the individual validation GCPs, by 78% and 69% for the SRTM and ASTER respectively. Additionally, the proposed vario-model showed a better representation of the accuracy than the RMSE one, by 94% and 86% respectively.

A major task to be considered as a future work is to design a procedure for propagating the developed vario-models to model the accuracy of the datasets derived from analyzing the related DEM. Other tasks to be considered is repeating this study over other areas around the world using the same SRTM and ASTER DEMs, along with the same CORINE LULC standard, and compare the results found in this study for consistency.

References

[1] Guth P.L., Geomorphometry from SRTM – comparison to NED. Photogrammetric Engineering & Remote Sensing, 2006, 72(3): 269–277.10.14358/PERS.72.3.269Search in Google Scholar

[2] Rodgriguez E., Morris Ch., and Belz J., A global assessment of SRTM performance, Photogrammetric Engineering and Remote Sensing, 2006, 72: 249–260.10.14358/PERS.72.3.249Search in Google Scholar

[3] Wechsler S., and Kroll Ch., Quantifying DEM uncertainty and its effect on topographic parameters, Photogrammetric Engineering and Remote Sensing, 2006, 72:1081–1091.10.14358/PERS.72.9.1081Search in Google Scholar

[4] Aguilar F., Agüera F., Aguilar M., and Carvajal F., Effects of Terrain Morphology, Sampling Density, and Interpolation Methods on Grid DEM Accuracy. Photogrammetric Engineering & Remote Sensing, 2005, 71(7).10.14358/PERS.71.7.805Search in Google Scholar

[5] Myint S., Gober P., Brazel A., Grossman-Clarke S., and Weng Q., Per-pixel vs. Object-based classification of urban land cover extraction using high spatial resolution imagery, Remote Sensing of Environment, 2011.10.1016/j.rse.2010.12.017Search in Google Scholar

[6] Thakur S, Singh A., and Suraiya S., Comparison of Different Image Classification Technique for Land Use Land Cover Classification: An Application in Jabalpur District in Central India, International Journal of Remote Sensing and GIS, 2012, 1(1).Search in Google Scholar

[7] Guo-an T., Strobl J., Jian-ya G., Mu-dan Z., and Zhen-jiang Ch., Evaluation on the accuracy of digital elevation models, Journal of Geographical Sciences, 2001, 11(2).10.1007/BF02888692Search in Google Scholar

[8] Su J. and Bork E., Influence of Vegetation, Slope, and Lidar Sampling Angle on DEM Accuracy. Photogrammetric Engineering & Remote Sensing, 2006, 72(11).10.14358/PERS.72.11.1265Search in Google Scholar

[9] Akira H., Welch R., and Lang H., Mapping from ASTER stereo image data: DEM validation and accuracy assessment, ISPRS Journal of Photogrammetry& Remote Sensing, 2003, 57: 356-370.10.1016/S0924-2716(02)00164-8Search in Google Scholar

[10] Aurenhammer F., and Klein R., Voronoi Diagrams. Chapter 5 in Handbook of Computational Geometry. Amsterdam, Netherlands: North-Holland, pp. 201-290, 2000.10.1016/B978-044482537-7/50006-1Search in Google Scholar

[11] Passini R. and Jacobsen K., Accuracy Analysis of SRTM Height Models. ASPRS 2007 Annual Conference.Search in Google Scholar

[12] Hohle J., DEM generation using a digital large format frame camera. Photogrammetric Engineering & Remote Sensing, 2009, 75(1): 87–93.10.14358/PERS.75.1.87Search in Google Scholar

[13] Reuter H., Neison A., Strobl P., Mehl W., and Jarvis A., A First Assessment of ASTER GDEM Tiles for Absolute Accuracy, Relative Accuracy and Terrain Parameters, IEEE IGRASS, 2009, 240.10.1109/IGARSS.2009.5417688Search in Google Scholar

[14] Moffitt F., Leveling, Surveying 10th Edition. Addison Wesley Longman, 1998, pp. 61-127.Search in Google Scholar

[15] Collin H., Huang Ch., Yang L., Wylie B., and Coan M., Development of a 2001 National Land-Cover Database for the United States. USGS Staff – Published Research, 2004, Paper 620.Search in Google Scholar

[16] NASA Jet Propulsion Laboratory: SRTM Home Page, 2009 http://www2.jpl.nasa.gov/srtmSearch in Google Scholar

[17] NASA Jet Propulsion Laboratory: ASTER Home Page, 2004 http://asterweb.jpl.nasa.gov/Search in Google Scholar

[18] Sertel E., Accuracy Assessment of ASTER Global DEM over Turkey, ISPRS Technical Commission IV, 2010, Orlando, Florida.Search in Google Scholar

Received: 2016-4-5
Accepted: 2016-7-13
Published Online: 2016-12-30
Published in Print: 2016-1-1

© 2016 Wassim Katerji et al.

This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 3.0 License.

Downloaded on 2.3.2024 from https://www.degruyter.com/document/doi/10.1515/geo-2016-0052/html
Scroll to top button