Accurate terrain elevation information is important in many applications of land surface modeling, such as flood, volcanology, ecology, and glaciology modeling [1
]. Space-borne radar or air-borne laser scanning are widely applied to retrieve data on topography that is used to develop the digital elevation model (DEM) [4
]. A DEM can be used to depict the terrain of the earth and is an organized array of the numbers which represent the elevations of spatial distributions above an arbitrary datum [7
]. The principle of a DEM is to describe the elevations of various points in a given area in a digital format. The term DEM is usually applied to land surface topography, but it is a general term that is used to depict the spatial patterns of various surfaces, e.g., surface water, ground surface, canopy, and so on. Digital surface model (DSM) and digital terrain model (DTM) are the two other terms which are frequently used for the ground terrain. DTM is referred to as the Earth terrain, i.e., bare ground, while DSM includes objects on the ground such as buildings and trees.
The Shuttle Radar Topography Mission (SRTM) is a publicly accessible DEM, at a global scale. Although it is provided at no cost, its accuracy is limited, with a root mean square error (RMSE) of more than 8 m in Singapore’s dense urban/forest areas [8
]. It was reported that SRTM suffers from inaccuracy especially in areas covered by the canopy, as the 5.6 cm wavelength used does not penetrate vegetation well [9
]. The absolute vertical SRTM error was found to be 22.35 m across 255,646 samples in the Amazon rainforest [10
], whilst in open areas of South America the equivalent error was at 6.2 m [5
]. Also, due to the rapid development of the urban area and coarse resolution, SRTM cannot capture the current building characteristics (SRTM collected the radar imagery in 2000 with approximately 30 m resolution) [11
]. There have been many studies on improving/correcting satellite DEMs using various methods. Data fusion is one of the techniques used for eliminating errors from space-borne DEMs [11
]. Muhadi et al. (2019) used a data fusion technique for deriving DEM that exploits two or more data to create a new data set for the planning and management of an oil farm plantation [16
]. The idea is the limitation of one sensor could be compensated by the other sensors, so the combination of different data sets overcome the limitations.
There are other types of studies using artificial neural networks (ANNs) for DEM improvement. Wendi et al. (2016) presented a promising DEM improvement scheme and showed substantial improvement of SRTM DEM with a RMSE reduction of 52–68% over two different forested areas in Singapore [9
]. The author used the ANN together with Landsat 8 multispectral imagery and 92 m resolution of SRTM to eliminate the error caused by dense canopy level in original SRTM. The application of ANN to improve SRTM was used in coastal areas where the elevation varies from 1 m to 20 m [17
]. The author used various input nodes in ANN which represent the characteristics of terrain such as slope, population density, canopy height, ICESat (Ice, Cloud and Land elevation satellite), and vegetation density. In the testing set, the RMSE between ground truth and derived SRTM were reduced by approximately 50%, and trained ANN applied to global scales where it showed reduced errors. Although both these methods were applied successfully to building an error regression model, they were limited to forested areas and/or coarse resolution, which do not represent the dense urban areas. Bagheri et al. (2018) fused two different sets of DEM data (TanDEM-X and Cartosat-1) using ANN to enhance the quality of both DEM datasets [18
]. The authors trained an ANN to learn the pattern of the relationship between height errors and features from the two datasets. The relative accuracy of derived DEM was improved both DEMs up to 50% in the validation. The authors drew the usage of ANN with strength in pattern recognition, which is the core idea in this study for DEM enhancement. Figure 1
shows two main limitations of SRTM DEM: (1) as sensors partially do not penetrate the vegetation area, the top of the canopy level represents the elevation of the forested area; and (2) with its coarse resolution, it does not represent particularly the dense urban cities well—for example, a grid could present the average of elevations of a low lying road/area and high rise buildings within that grid. The impacts of these limitations can be significant, for example, inaccuracy of flood simulations that affect mitigation measures [19
This paper presents significant improvements to the SRTM DEM using an ANN with remote sensing data. The improvement is particularly significant for dense urban areas. Figure 2
demonstrates the schematic diagram of our DEM improvement methodology. Generally, it requires four types of data; multispectral imagery, the DEM to be improved (SRTM DEM in this study), the building footprint for sorting the building areas, and a reference DEM (ground truth elevation).
Multispectral imagery is produced by the sensors which measure the reflected energy within several specific bands/sections of the electromagnetic spectrum. This can be defined as the acquisition of images in hundreds of contiguous, registered, spectral bands such that for each pixel a radiance spectrum can be derived [23
]. Sentinel-2 multispectral imagery has 13 bands of wavelengths. In this study only eight bands, band 2–band 8A, are used while the remaining five bands (bands 1, 9, 10, 11 and 12) are not relevant in this study, as they are for aerosol, water vapor, snow, ice, and clouds correction [26
]. Two sets of experiments with ANN training were conducted. One ANN was trained for non-building areas, while the other ANN was only for building areas. The high spatial resolution (1 m) and high-accuracy (40 cm) surveyed DEM was used as the reference elevation. Once the performance of the trained ANN was acceptable, it could be applied to areas where their SRTM DEMs are to be improved.
In this research, high-resolution and high-accuracy surveyed DEMs are available for Nice (France) and Singapore. The ground truth DEMs were used as reference DEMs to train and validate the ANN. SRTM DEM and Sentinel 2 data were used as common inputs for all areas. The trained ANN was later be applied to other areas in Nice and Singapore to verify its applicability. The performance of the DEM improvement scheme was evaluated over two dense urban cities using various matrices, e.g., visual clarity, scatter plots, root mean square error (RMSE), and flood maps.
4. Proof of Concept and Application of the Approach
This section evaluates the performance of the DEM derived using the method developed in this study, as described in Section 3
. DEMs in Nice (France) and Singapore were taken into consideration. Two scenarios of test cases were introduced in dense urban areas: (1) the ANN model trained and validated in Nice, France; and (2) the ANN model trained in Nice and validated in Singapore. The second case was essential, as we needed to ascertain the applicability of the ANN model, trained in Nice, at other places where no high-quality DEM, except satellite imagery, is available.
For the case of urban areas in Nice, the training area has an area of 12.0 km2
while the validation area was 5.2 km2
. Figure 4
shows the satellite image of the training (box with blue comb pattern) and validation (box with red comb pattern) areas. The areas are mainly urbanized with buildings, and the elevation profiles are from 0 m to 200 m. The average building height is 19.1 m (maximum 60.8 m) and buildings occupy 34% of the total area.
The ANN was trained in the training area with 1 m reference DEM data as the target layer. The iSRTM DEMwais obtained from two ANN trainings, one with and one without building heights. The trained ANNs were then applied to the validation area and the performances were evaluated against the reference DEM. Figure 5
shows the comparison of elevation maps of various DEMs.
a is a satellite image of the test area depicting the land shapes; Figure 5
b is the area from 1 m reference DEM; Figure 5
c is the area from the original SRTM DEM with 30 m resolution; and Figure 5
d is the area resulting from iSRTM DEM with 10 m resolution. The reference DEM shows most clear land shapes (i.e., buildings and roads); iSRTM DEM also shows clearer land shape visibility than the original SRTM DEM. iSRTM DEM (Figure 5
d) most matched the reference DEM (Figure 5
b). The significant improvements are reflected in statistical analysis in Figure 6
as well. The RMSE of iSRTM DEM reduced to 5.18 m from 8.36 m of SRTM DEM (a 38% reduction). Figure 6
c shows the frequency error distribution of iSRTM DEM and SRTM DEM. The percentage of absolute errors between −5 m to 5 m was 33.4% in SRTM DEM, while for iSRTM DEM it was 63.5%.
For the case of urban areas in Singapore, the interest was to investigate the quality of DEM generated by an ANN, trained in Nice, when it is applied in areas far away. The quality of DEM generated by the trained ANN was first validated in Singapore where good quality DEM is available. The area is a very dense urban area with high-rise buildings (Bukit Timah area, Singapore). The elevation ranges from 0 m to 100 m. The average building height is 18.8 m (maximum 90.5 m) and buildings occupy 28% of the total area.
shows a comparison of elevation maps between various DEMs. Figure 7
a is a satellite image of the test area depicting the land shapes; Figure 7
b shows the 1 m reference DEM; Figure 7
c shows the original SRTM DEM with 30 m resolution; and Figure 7
d shows the iSRTM DEM resulting from the ANN trained in Nice with 10 m resolution. The comparisons show that iSRTM DEM matches the 1 m reference DEM more than the original SRTM DEM. The improvements are reflected in statistical analysis in Figure 8
as well. The RMSE of iSRTM DEM is reduced from 10.70 m (SRTM DEM) to 6.93 m (35.2% reduction). Figure 8
c shows the frequency error distribution of iSRTM DEM and SRTM DEM. The percentage of absolute errors between −5 m to 5 m was 14.9% in SRTM DEM, while for iSRTM DEM it was 49.3%.
shows the error patterns in SRTM DEM and iSRTM DEM for different land-uses. The building areas in SRTM show the biggest RMSE, followed by impervious and pervious areas in Nice and Singapore. The RMSE of all the different land-use areas were reduced in iSRTM DEM for both validation areas. The most improvement occurred in the impervious area of Nice with a 44.7% RMSE reduction, while building areas in Singapore showed the most RMSE reduction, with a 42% improvement. It would be interesting in future work to consider other error patterns, such as that in [40
It is an interesting finding that the SRTM DEM of a place, where no good-quality surveyed DEM is available, can still be significantly improved with ANN trained in a faraway dense urban area where high-quality ground truth data are available.
SRTM DEM and iSRTM DEM were used for flood modeling to verify the applicability of iSRTM DEM in the hydrodynamic modeling application. The MIKE 21 flow model developed by DHI Water & Environment [41
] was used for a two-dimensional flow modeling system. The main purpose of this experiment was to identify which DEM represents the better fictitious flood maps based on common phenomena of inundation (flooding on low-lying areas such as roads in the urban areas). Fictitious rainfall (300 mm per 6 h) and a free flow boundary condition were used. The flood maps from different DEMs are compared in Figure 9
and Figure 10
and Figure 10
illustrate the inundated areas in Nice, France, and Singapore respectively. Different DEMs, SRTM DEM and iSRTM DEM, were used in the flood model to investigate the flood patterns. Flood maps resulting from iSRTM DEM capture the flooding in the flood prone areas, i.e., roads/low-lying areas, while flood maps resulting from the original SRTM DEM do not. Due to the coarse resolution of SRTM DEM (30 × 30 m) and inaccurate terrain elevation, flood patterns do not follow the real topographic characteristics.
The main objective of this research was to develop the SRTM DEM improvement scheme using ANN along with other remote sensing data. As discussed in Section 1
, SRTM has limitations due to its sensor and coarse resolution. The errors were able to be corrected using ANN with multispectral imagery after it was trained with ground truth data. This generalized neural network was applied to Singapore, which is far away from the training area in Nice France; as mentioned earlier the main purpose was to check whether the generated DEM in Singapore matches well with the surveyed DEM. Although the RMSE was significantly reduced, the building heights (higher than 60 m) did not clearly match with ground truth (Figure 7
). The reason for this is that buildings in the study area of Nice are mostly less than 60 m (only 0.1% is between 60–100 m), which is not as high as Singapore’s buildings (4% between 60–100 m). This means that the trained ANN mainly learned patterns of buildings of up to 60 m heights. Table 4
shows the percentage of impervious area and building characteristics in the study areas. This implies that the pattern of terrain shapes and building heights learned from similar areas and/or more variable patters would generate better performance in improved SRTM.
The building areas were filtered using the Open Street Map (OSM) building footprint. It has been reported that OSM data may have some inaccuracy with its positioning in a few meters [42
]. In this research, we used 10 m resolution for the DEM so that this error would not be significant. However, the presence of buildings in the data set is important, in that it is necessary to use the latest building information.
The 2D flood maps were generated using different DEMs with fictitious hydrologic data. Flood maps may be different from the actual situations as drainage networks are not considered. Also, this research did not use the actual rainfall characteristics of the areas. The flood maps from iSRTM DEM, however, showed that it captures the flooding on the roads (low-lying areas) better than that of SRTM DEM. This finding is not surprising, as the finer resolution iSRTM DEM (10 m) incorporates the terrain characteristics similar to those of the real condition.
The data fusion technique can be applied to this DEM improvement scheme using more data from other satellites (e.g., TanDEM-X, ASTER DEM, AW3D DEM, Landsat 8, Sentinel 1 and ASTER imagery). This technique would increase the performance of the output as the limitation of one sensor could be compensated by the other sensors [16
]. Also, the methodology developed is quite flexible in data selection and can be applied to the other space-borne DEM data sources mentioned above for their improvement.
This study used a classical neural network regression method to reduce the error between ground truth data and SRTM. More complicated and different architecture of neural networks could allow improved performance by reducing errors.
6. Summary and Conclusions
A new DEM improvement scheme for SRTM DEM in dense urban cities was suggested and described in this paper. The scheme was developed using an artificial neural network (ANN) with SRTM DEM and Sentinel 2 multispectral imagery as the input nodes, while high resolution and accuracy surveyed DEM was used as the target layer. The trained ANN was able to classify the land-uses and land-covers with the assistance of different bands of Sentinel 2. Based on the various land characteristics in the training, different weights were calculated to reduce the errors between the elevations of SRTM DEM and surveyed DEM.
Two scenarios were taken into consideration for training and validation: (1) an ANN model trained and validated in Nice, France; and (2) an ANN model trained in Nice and validated in Singapore. In both scenarios the performance of improved SRTM (iSRTM DEM) was shown to be significantly better than its counterpart, SRTM DEM. In the dense urban city of Nice, the RMSE reduction of the iSRTM DEM was 38% and its visibility (land shapes, buildings and roads) was clearer than SRTM DEM. As one of the interests in the study was to improve the SRTM DEM of faraway locations, where no high-quality surveyed DEM is available, the ANN trained in Nice was used to generate the DEM of a dense urban area in Singapore to test its applicability (scenario 2). The test performance again showed significant improvement over SRTM DEM, with a RMSE reduction of 35.2%. It is interesting to note that a well trained ANN somewhere with a high-accuracy DEM can be applied to generate DEM at other far way places, so long as their patterns are similar to the pattern of the place where ANN is trained. Flood simulations were conducted using fictitious hydrological data and different topography from SRTM DEM and iSRTM DEM. Flood map resulting from iSRTM DEM captured better flooding on low-lying areas such as roads. The scheme developed is able to be used in hydrodynamic applications where topographical information is crucial.
This study has shown that the quality of SRTM DEM can still be significantly improved with the DEM improvement scheme proposed in this paper. The DEM improvement scheme can be applied to the areas where high-quality DEM is not available. Also, the improved SRTM can be used in many types of applications (i.e., flood, groundwater modeling) to allow the modeling performance to proceed with high confidence.