DeltaDTM: mapping coastal terrain elevation

A state-of-the-art global coastal digital terrain model (DTM) with a resolution of 1 arcsecond (~30m) has been created by researchers at Deltares and Delft University of Technology in the Netherlands. In their work, they extensively use elevation models as input for their numerical models, but they noticed that freely available elevation models were not sufficiently accurate for their purposes. To address this, they developed a new elevation model by integrating data from the latest ICESat-2 and GEDI missions with the CopernicusDEM data. This article details their methodology.

To accurately model future extreme water levels caused by sea level rise (SLR), subsidence and worsening storm surges, elevation data with high vertical accuracy (within 1m) is essential for all coastal areas worldwide. Local airborne Lidar data is sometimes used for this purpose, but this is expensive and only available in more affluent parts of the world rather than globally.

In areas where this data is missing, such as in Southeast Asia, global digital elevation models (DEMs) are utilized to assess coastal flood risk, among other things. However, these models are based on digital surface models (DSMs) that measure only the upper part of canopy and buildings, and thus do not represent the bare earth and height everywhere. In vegetated areas, the differences between the model and terrain can be tens of metres.

Challenges in high-resolution Lidar DEMs

Since 2018, the spaceborne Lidar missions ICESat-2 and GEDI have provided global, albeit sparse, terrain measurements. ICESat-2 alone has been used to create a global coastal digital terrain model (GLL_DTM v2 [Vernimmen & Hooijer, 2023]), achieving high accuracy (mean absolute error [MAE] of 0.34m) but with a low horizontal spatial resolution of 1km. When ICESat-2 and GEDI are combined, a 500m-resolution Lidar-based DEM could be achieved globally (Pronk, Eleveld & Ledoux, 2024).

To achieve higher-resolution DEMs with spaceborne Lidar, this data must be combined with global DEMs. Several efforts have been made to correct biases in global DEMs for areas covered by vegetation or buildings using spaceborne Lidar and auxiliary datasets like tree-cover or urban agglomeration maps. CoastalDEM, FABDEM and DiluviumDEM all use ICESat-2 data to correct the surface data present in global DEMs. Except for DiluviumDEM, however, these corrected DSMs are not in the public domain – they are only free for research purposes – and nor are the machine learning models used to generate them.

4-step methodology of DeltaDTM

The newly developed DeltaDTM is the first open and global coastal terrain model with a vertical accuracy within 1m. It corrects the vertical biases in surface data (e.g. canopy, buildings) present in CopernicusDEM by utilizing ICESat-2 and GEDI terrain elevation measurements. The method involves four key steps:

• Spatial filtering: removing pits and other outliers present in CopernicusDEM
• Co-registration: vertically aligning CopernicusDEM with ICESat-2 to eliminate any vertical bias
• Filtering of non-ground points: classifying CopernicusDEM into terrain and non-terrain using morphological filters and removing the non-terrain elevation pixels
• Void filling: spatially interpolating the values removed in the previous step using the adjusted inverse distance weighing (AIDW) method.

Spatial filtering

The base elevation model DeltaDTM used as the starting point is the CopernicusDEM GLO-30 dataset, provided under COPERNICUS by the European Union and the European Space Agency (ESA). The dataset is distributed in tiles of 1 by 1 degree, with a spatial resolution of one arcsecond (~30m at the equator). It is based on TanDEM-X interferometric synthetic aperture radar (SAR) data and is freely available for the entire globe.

CopernicusDEM contains many small low outliers, often the result of multi-bounce backscattering errors in urban areas, such as around electricity poles. A 25 by 25-pixel window (~750 x 750m) function is used to remove all values below two standard deviations of all elevation values in the window. The window size is sufficient to filter larger patches (3 by 3 pixels) of low outliers, as observed in CopernicusDEM.

Likewise, all elevation values are removed in case the height errors exceed 0.75m according to the CopernicusDEM height error data, or CopernicusDEM was infilled (patched) with another DEM. The 0.75m value was empirically chosen based on the outliers observed in validation areas. Furthermore, quality filters are applied on both the ICESat-2 and GEDI data. For ICESat-2, only data with the flag ‘subset_te_flag’ set to 1 is kept. For GEDI, the same filtering as used in standard processing for the derived GEDI L3A product is used, and only data with the ‘sensitivity’ flag above 0.95 is kept.

Co-registration

Any elevation dataset will have biases due to instrument and processing errors, and these biases can be determined and corrected by using a second – more accurate – elevation dataset. The ICESat-2 ATL08 data is used to correct the terrain elevation bias in the CopernicusDEM dataset. GEDI is not used for the bias correction, as it is less accurate for terrain elevation assessment than ICESat-2 and does not cover latitudes above 56°.

For each quarter of a CopernicusDEM tile (0.5 by 0.5 degree), the elevation of the ICESat-2 points is compared to the elevation of the CopernicusDEM data for all landcovers without trees or buildings. In this way, the distribution of CopernicusDEM minus ICESat-2 could be calculated, and the peak of this distribution was denoted as the bias for each tile.

The resulting point dataset, containing the bias at the centre coordinates of each quarter of a CopernicusDEM tile, was used to create a bias correction raster for the whole tile by interpolating using a nearest neighbour algorithm. Afterwards, this bias correction raster was applied to the original CopernicusDEM tile.

Filtering of non-ground points

Like any current global radar or optical-based DEM, CopernicusDEM measures the surface of the Earth and thus includes vegetation, building heights and other civil constructions. To remove these biases and determine the true ‘bare-earth’ surface, morphological surface filters are applied that are supported by terrain measurements from the ICESat-2 ATL08 and GEDI L2A data.

Morphological filters relate to the morphology (shape) of features and work on subsections (windows) of raster (image) data, to which non-linear (such as minimum) filters are applied. These filters are often used for terrain classification of airborne Lidar datasets, but require at least some terrain measurements in each area to work. On its own, CopernicusDEM is not suitable for such filtering, as it does not contain any terrain measurements in large parts of the world, such as tropical forests. Moreover, these filters normally operate on the scales of individual trees and houses, using raster resolutions of one metre, not ~30m as in the case of CopernicusDEM.

CopernicusDEM data is replaced with ICESat-2 ATL08 data (Neuenschwander & Pitts, 2019) and GEDI L2A terrain data (Dubayah, Hofton, Blair, Armston, Tang & Luthcke, 2021) when available, ‘burning’ the Lidar-derived elevations into the bias-corrected CopernicusDEM raster. This enables the use of morphological filters, albeit with much larger windows sizes than usual morphological filter operations. Specific algorithm settings – such as slopes and the initial height threshold – are dynamically derived per landcover class from ICESat-2 ATL08 and GEDI L2A data.

Void filling

The resulting non-terrain cells – on average 50% of a tile – are filled by AIDW interpolation of the remaining terrain points. The resulting interpolated surface is unrealistically smooth for a terrain. To create a more realistic visual landscape representation, the roughness of the surface – derived from the original CopernicusDEM – is added to the interpolated terrain values only. In the worst case, this adds random noise to the DEM, like the noise present in non-interpolated CopernicusDEM elevation values. In the best case, however, it represents actual topography patterns such as ditches or small canals underneath the canopy. Overall, the additions are small and balanced (roughly having a zero mean) and do not affect the accuracy.

Validation and conclusion

The DeltaDTM dataset has been validated against public local airborne Lidar reference datasets in Australia, the USA (Florida), Indonesia (see Figure 2), Latvia, the Marshall Islands, Mexico, the Netherlands, Poland and the United Kingdom. Of the corrected elevation models, DeltaDTM performs best for all land cover classes combined, with a bias of -0.03m, an MAE of 0.42m, and a root mean square error (RMSE) of 0.71m. 92% of DeltaDTM is within 1m of the reference surface, 98% within 2m and 100% within 5m. The next best DEM is DiluviumDEM, followed by FABDEM (although this is closely matched by CoastalDEM, but not for the percentage within 1m).

Each corrected DSM has its own strengths and performs differently per land cover class. For example, FABDEM has been optimized for urban areas and has a similar performance for ‘Built-up’ as DeltaDTM, with an MAE of 0.69m and 0.55m respectively. In areas with no vegetation or buildings, like ‘Wetland’ or ‘Cropland’, an uncorrected DSM such as CopernicusDEM performs like corrected DSMs. CopernicusDEM has an MAE of 0.43m for ‘Cropland’, whereas FABDEM and DeltaDTM have an MAE of 0.38m and 0.32m, respectively. As expected, the errors for ‘Tree cover’ are greatest, with 87% of DeltaDTM elevations within 1m, one of its lowest values overall. DiluviumDEM is next, with 60% within 1m, followed by CoastalDEM at 49% and FABDEM at 42%.

In the reference area with most ‘Tree cover’, all datasets have lower accuracies. Clearly, extensive and dense forest in the tropics is hard to correct for. DeltaDTM is closest to the airborne Lidar reference, having the smallest errors overall, but it still misses smaller patches of forest. The hillshades show artefacts in the processing of CoastalDEM and DiluviumDEM. Both CoastalDEM and DiluviumDEM display machine learning artefacts in the form of square patches of pixels with different elevations, whereas DiluviumDEM also has large differences between individual corrected pixels, resulting in a high overall slope.

It should be realized that given the overall RMSE of 0.71m, DeltaDTM can be used to model SLR in increments of 1.40m or higher at 68% confidence level (Figure 3). For 1m SLR increments, the confidence level will be 52%.

Accessing the DeltaDTM data

DeltaDTM is licensed under the CC BY 4.0 licence, which means that you are free to share and adapt the dataset, if you give appropriate credit.

DeltaDTM is available as a zipped (.zip) archive per continent (to a total of 35 GB) at https://doi.org/10.4121/21997565.

It is also hosted as a Google Earth Engine collection under the collection ID users/maartenpronk/deltadtm/v1-1.

An example on how to access the dataset is provided at

https://code.earthengine.google.com/?scriptPath=users/maartenpronk/deltadtm:v1.

Since the initial release, several updates to DeltaDTM have been made. Since version 1.1, it goes up to 30m above sea level. The ICESat-2 and GEDI missions are ongoing, so further updates and improvements to DeltaDTM are to be expected.

Acknowledgements

This article reuses text and figures from the paper: Pronk, M., Hooijer, A., Eilander, D. et al. DeltaDTM: A global coastal digital terrain model. Sci Data 11, 273 (2024), https://doi.org/10.1038/s41597-024-03091-9, licensed under a Creative Commons Attribution 4.0 International License. To see a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/

Further reading

Vernimmen, R.; Hooijer, A. (2023). New LiDAR-Based Elevation Model Shows Greatest Increase in Global Coastal Exposure to Flooding to Be Caused by Early-Stage Sea-Level Rise, Earth’s Future, Vol. 11, No. 1, e2022EF002880. doi:10.1029/2022EF002880

Pronk, M.; Eleveld, M.; Ledoux, H. (2024). Assessing Vertical Accuracy and Spatial Coverage of ICESat-2 and GEDI Spaceborne Lidar for Creating Global Terrain Models, Remote Sensing, Vol. 16, No. 13, 2259. doi:10.3390/rs16132259

European Space Agency; Airbus. (2022). Copernicus DEM, European Space Agency. doi:10.5270/ESA-c5d3d65

Neuenschwander, A.; Pitts, K. (2019). The ATL08 land and vegetation product for the ICESat-2 Mission, Remote Sensing of Environment, Vol. 221, 247–259. doi:10/gf9wmm

Dubayah, R.; Hofton, M.; Blair, J.; Armston, J.; Tang, H.; Luthcke, S. (2021). GEDI L2A Elevation and Height Metrics Data Global Footprint Level V002, NASA EOSDIS Land Processes Distributed Active Archive Center. doi:10.5067/GEDI/GEDI02_A.002