Which elevation model: SRTM, Copernicus, CartoDEM, lidar

You need a digital elevation model (DEM) for a catchment, a flood map or a bridge study, and the free options all look alike: a grid of heights at about 30 m. Yet the question of SRTM vs Copernicus DEM matters, because the two differ in age, in sensor and in how they treat trees and buildings, and neither is a substitute for a surveyed cross-section. This article compares the models in common use in India, states what each one measures and how well, and suggests which to pick for which job.

What is a DEM, and what does it actually measure?

A DEM is a grid of heights. The word hides an important split. A digital surface model (DSM) records the top of whatever the sensor sees first: tree canopy, roofs, snow. A digital terrain model (DTM), also called a bare-earth model, records the ground. Almost every free global model is a surface model. The Copernicus DEM describes itself as a DSM that represents the surface "including buildings, infrastructure and vegetation" [1], and NRSC states that CartoDEM, "although termed CartoDEM, it is a DSM and not a bare earth surface" [7].

Three more facts decide how far you can trust any model: the sensor (radar, optical stereo or laser), the date it was acquired (a river can move, a slope can slide, a road can be built since), and the vertical datum (the surface heights are measured from; mismatches can be metres). See gauge zero, datum and benchmarks for the same problem in water levels.

The main elevation models compared

SRTM: the 2000 radar baseline

The Shuttle Radar Topography Mission flew on Space Shuttle Endeavour for 11 days in February 2000 and mapped land between 60° N and 56° S at 1 arc second (about 30 m) [3]. Version 3 filled the voids of earlier releases using other elevation models, ASTER GDEM v2 among them [3]. It records the first surface the radar meets, so it is a surface model.

How good is it in the Himalaya? Mukul and colleagues tested SRTM against about 200 high-precision GPS points across India. For the 30 m C-band data (the product most people download), the raw error was 23.53 m RMSE, falling to 10.14 m after removing voids and outliers. In the higher Himalaya (above 2,000 m) the filtered 3 arc-second data had an RMSE of 18.63 m and the 1 arc-second data 13.71 m, against about 8 to 16 m in the foreland and foothills [4]. Errors grew with elevation and slope.

Copernicus DEM: GLO-30 and GLO-90

The Copernicus DEM is built from the TanDEM-X radar mission, with data acquired between 2011 and 2015, edited and gap-filled with older models where needed. It is offered at 30 m (GLO-30) and 90 m (GLO-90), free of charge worldwide, with the source credit to DLR and Airbus required for GLO-30 [1]. The specification asks for an absolute vertical accuracy better than 4 m (90 % linear error) and a relative accuracy better than 2 m on slopes up to 20 % and better than 4 m on steeper slopes [1].

The validation report checks the model against ICESat laser heights. Over its global comparison it reports an absolute vertical accuracy of 2.17 m at 90 % confidence, a mean error of −1.29 m and an RMSE of 1.68 m [2]. One caveat: the validation removed reference points in mountainous terrain with a standard-deviation filter [2], so the headline figure describes gentler ground than most Himalayan catchments. The report itself notes that cells with larger errors cluster along mountainous terrain and forests [2].

Independent tests elsewhere point the same way. Against airborne lidar in Cape Town, Copernicus had the lowest error of four global models, with an RMSE of 2.34 m. In dense tropical forest in Gabon, though, ASTER and NASADEM did better, and the study concluded that Copernicus was more accurate where tree cover was under 40 % and the others where it exceeded 70 % [12]. We found no published test of GLO-30 against GPS or lidar in the Indian Himalaya, so treat its better global figures as a good prior, not a result for your catchment.

ASTER GDEM: older and noisier

ASTER GDEM version 3 comes from optical stereo images acquired between March 2000 and November 2013, at 30 m, from 83° N to 83° S [5]. Optical stereo struggles in cloud, shadow and snow, and NASA warns that the product may contain anomalies and artefacts [5]. In the United States, version 3 had an RMSE of 8.52 m and a mean error of −1.20 m against more than 23,000 control points [6]. In steep Hunan Province, China, ASTER GDEM2 was the worst of the open models tested (RMSE 10.1 m against ICESat-2, against 8.0 m for SRTM) [13].

CartoDEM: India's own stereo DSM

CartoDEM is produced by ISRO from Cartosat-1 stereo images, with Cartosat-1 launched on 5 May 2005. The product is a DSM with an absolute vertical accuracy of 8 m (LE90) and a horizontal accuracy of 15 m (CE90), stated to hold for slopes under 20°. NRSC's own disclaimer adds that accuracy "may vary" on slopes above 20°, that sinks and spikes occur in hill shadows and snow areas, that no hydrological conditioning has been done, and that about 1 % of the area has gaps because cloud-free stereo pairs were unavailable [7]. The specification lists a 10 m posting, and NRSC also offers an updated 2.5 m surface model; NRSC reports that, in its comparisons for hydrological and hydrodynamic modelling, the 2.5 m version was much better than the 10 m one and showed more detail than a 5 m TanDEM-X model [7].

Two practical points. First, the specification states heights relative to the WGS84 ellipsoid [7], so check the datum before mixing CartoDEM heights with levels from a gauge or with a model referenced to a geoid. Second, the 8 m figure is a specification, and independent assessment in steep terrain is thin.

Airborne lidar: the one that measures the ground

Lidar fires laser pulses from an aircraft and records the returns. Some pulses pass through gaps in the canopy, so the lowest returns give the ground, and the result can be a true DTM. The US Geological Survey's lidar base specification (version 1.3, 2018) shows what is routinely achieved: its common "quality level 2" asks for at least 2 pulses per square metre and a vertical RMSE of 10 cm on open, non-vegetated ground [10]. Accuracy under vegetation is looser, which is why that specification assesses vegetated and non-vegetated ground separately [10].

In India, NRSC runs airborne lidar surveys and says high-resolution DEMs have been generated for the Indian coast and a few flood-prone river reaches [9]. The Survey of India, under the National Hydrology Project, lists DEMs of 0.5 m, 5 m and 10 m for river-basin areas, up to 5 km either side of the rivers, for plain areas, and calls for local high-resolution surveys such as lidar in flood-prone areas [8].

SRTM, Copernicus and CartoDEM at a glance

The accuracy measures differ (LE90, RMSE, different reference data), so the numbers are a guide, not a ranking.

Which DEM should you use?

Use the Copernicus DEM GLO-30 as the default free 30 m model for catchment delineation, slope and rough hydrological work, and the SRTM or ASTER models as a cross-check for gross errors. Use CartoDEM 2.5 m where its finer posting matters and the slopes are gentle, and lidar wherever the water level at a building, bridge or bank matters.

Catchments and watersheds

Catchment boundaries, flow directions and stream networks depend on relative heights over many kilometres, and a 30 m model is usually adequate for catchments of tens of square kilometres and above. Small basins, narrow gorges and sinks are more sensitive; fill sinks and check the drawn stream against satellite imagery. Our article on delineating a basin from a DEM shows the steps.

Flood maps

Here the vertical error converts directly into horizontal error on gentle ground. The worked example below is illustrative, with the arithmetic shown.

lateral error ≈ vertical error ÷ ground slope

Floodplain slope 1 in 200 (0.005):   2 m DEM error  → 2 / 0.005 = 400 m
                                     8 m DEM error  → 8 / 0.005 = 1,600 m
Floodplain slope 1 in 20  (0.05):    2 m DEM error  → 2 / 0.05  =  40 m
                                     8 m DEM error  → 8 / 0.05  = 160 m
Lidar, 0.1 m error, slope 0.005:                       0.1 / 0.005 =  20 m

On a flat plain, a 2 m error can shift the edge of a mapped flood by hundreds of metres, and a flood is often only a few metres deep to begin with. That is why a free 30 m surface model gives a useful screening map, not a basis for deciding which houses or which road section will flood. Hawker and colleagues found that removing trees and buildings from the Copernicus model with machine learning cut the mean absolute error in forest from 5.15 m to 2.88 m and, in built-up areas, from 1.61 m to 1.12 m [11]; the bare-earth product, however, carries a non-commercial licence, so check the terms before using it in a project. More on how hazard maps are built and used is in flood hazard maps and floodplain zoning.

River channels and stations

No satellite DEM sees the river bed, and radar echoes from water make the channel look flat. For discharge you need a surveyed cross-section at the gauge, not a DEM; see river cross-section surveys. A DEM does help at the next stage: choosing the site, estimating the upstream catchment area and slope, and placing the first sensors.

Steep Himalayan terrain

Every model degrades on steep slopes, so compare two models over your site and treat disagreements larger than the stated accuracy as uncertain: errors rise with slope in the Hunan study, with RMSE peaking near 15 m above 25° [13], and with elevation in the Indian Himalaya [4].

Key points

• Nearly all free global models are surface models: they include canopy and buildings. Only lidar can give a true bare-earth surface.
• Copernicus GLO-30 is the strongest free 30 m default, with a validated absolute accuracy of 2.17 m (LE90) on gentler ground, but its mountain performance in the Himalaya is not yet independently documented.
• SRTM in the Indian Himalaya had RMSE of 13.71 m above 2,000 m even after filtering [4]; ASTER is noisier still.
• CartoDEM is specified at 8 m (LE90) on slopes under 20°, with a 2.5 m posting and heights above the WGS84 ellipsoid.
• On flat ground a 2 m vertical error can displace a flood edge by hundreds of metres; use lidar or surveyed sections where it matters.

Where ClimMaTech fits

A DEM tells you where the water might go; a station tells you what it is doing. At ClimMaTech we work on the measurement side: river monitoring stations that record level and surface velocity directly, because no elevation model replaces a measured stage at the gauge.

References

1. Copernicus Data Space Ecosystem, Copernicus DEM: Global and European Digital Elevation Model (COP-DEM), collection description, European Space Agency / European Commission. dataspace.copernicus.eu
2. Airbus Defence and Space for ESA/European Commission, Copernicus Digital Elevation Model Validation Report, document GEO1988-CopernicusDEM-RP-001, version 3.0, 9 November 2020. PDF
3. NASA/USGS LP DAAC, NASA Shuttle Radar Topography Mission Global 1 arc second, version 3 (SRTMGL1.003). doi:10.5067/MEASURES/SRTM/SRTMGL1.003
4. M. Mukul, V. Srivastava, S. Jade, M. Mukul, Uncertainties in the Shuttle Radar Topography Mission (SRTM) heights: insights from the Indian Himalaya and Peninsula, Scientific Reports 7, 41672, 2017. doi:10.1038/srep41672
5. NASA/METI LP DAAC, ASTER Global Digital Elevation Model V003 (ASTGTM.003), 2019. doi:10.5067/ASTER/ASTGTM.003
6. D. B. Gesch, M. J. Oimoen, J. J. Danielson, D. Meyer, Validation of the ASTER Global Digital Elevation Model version 3 over the conterminous United States, ISPRS Archives XLI-B4, 143–148, 2016. doi:10.5194/isprs-archives-XLI-B4-143-2016
7. National Remote Sensing Centre (ISRO), 2.5 m CartoDSM products (data sheet, February 2024) and CartoDEM specifications, Bhoonidhi portal. CartoDSM sheet; CartoDEM specifications
8. Survey of India, National Hydrology Project page. surveyofindia.gov.in
9. National Remote Sensing Centre, Aerial Services & Data Management. nrsc.gov.in
10. H. K. Heidemann, Lidar Base Specification, U.S. Geological Survey Techniques and Methods 11-B4, version 1.3, 2018. pubs.usgs.gov
11. L. Hawker et al., A 30 m global map of elevation with forests and buildings removed, Environmental Research Letters 17(2), 024016, 2022. doi:10.1088/1748-9326/ac4d4f
12. C. J. Okolie, J. P. Mills, A. K. Adeleke, J. L. Smit, M. V. Peppa, A. O. Altunel, I. D. Arungwa, Assessment of the global Copernicus, NASADEM, ASTER and AW3D digital elevation models in Central and Southern Africa, Geo-spatial Information Science 27(4), 1362–1390, 2024. doi:10.1080/10095020.2023.2296010
13. Z. Liu, J. Zhu, H. Fu, C. Zhou, T. Zuo, Evaluation of the vertical accuracy of open global DEMs over steep terrain regions using ICESat data: a case study over Hunan Province, China, Sensors 20(17), 4865, 2020. doi:10.3390/s20174865