Digital terrain model
.png&w=3840&q=75)
A global, bare-earth terrain model at 10 m
We created a model that produces a 10 m bare-earth digital terrain model (DTM) from an open global 30 m terrain model, enhanced with satellite data. We train it against high-resolution national aerial LiDAR surveys where they exist. Once trained, the model works everywhere, producing a globally consistent result, including places where nobody has ever flown a survey.
Because the model is guided by satellite data, it can generate a surface for any year with enough satellite coverage, back to 2017.
How it performs
We validated the output against national LiDAR from 19 countries, and with hold-out data from two countries and an entire continent the model had never seen. It beats the two most commonly used DTMs on everything we measured.
Steep, broken country is where it helps most
The coarse 30 m input does worst where the ground is most complicated, so that is where our model has the most opportunity to improve over the input. The hardest place we tried was the alpine karst of Slovenia, one of the countries the model never saw during training. There it took the shape error from 6.10 m down to 2.72 m, removing 3.4 m of error, or 55% of it, and recovered between 35% of the missing river detail in the gentlest catchments and 45% in the steepest. Australia showed the same pattern, from 19% in the flattest catchments to 48% in the most rugged.
Flat ground has the least to give back
Where the ground is flat, the coarse input is already close to the truth, so there is less to fix. In the Dutch lowlands, the model takes the shape error from 1.16 m down to 0.44 m. That is the largest proportional cut we measured, 62%, but in absolute terms it is only 0.7 m of elevation.
Flat ground also holds little river structure for the input to have missed, so only 27% of the lowlands' missing river detail is recovered. The lowest anywhere was 19%, in the flattest Australian catchments.
MindEarth DTM
- Grid posting
- 10 m
- Shortest resolved wavelength
- 46 m
- Shape error, SI-RMSE (m)
- 1.56
- Absolute error, std-RMSE (m)
- 2.54
- Mean absolute error (m)
- 1.36
- Bias (m)
- 0.17
- Slope error, RMSE (°)
- 2.08
- Aspect error, RMSE (°)
- 16.6
- Channel network delineation F1, higher is better
- 0.545
- Flood detection rate at 1 m stage, higher is better
- 78%
- Flood false-alarm rate at 1 m stage
- 24%
On 117 Australian catchmentsGEDTM30
- Grid posting
- 30 m
- Shortest resolved wavelength
- 113 m
- Shape error, SI-RMSE (m)
- 3.22
- Absolute error, std-RMSE (m)
- 3.79
- Mean absolute error (m)
- 2.00
- Bias (m)
- 0.17
- Slope error, RMSE (°)
- 4.07
- Aspect error, RMSE (°)
- 33.0
- Channel network delineation F1, higher is better
- 0.388
- Flood detection rate at 1 m stage, higher is better
- 62%
- Flood false-alarm rate at 1 m stage
- 33%
On 117 Australian catchmentsFABDEM
- Grid posting
- 30 m
- Shortest resolved wavelength
- 105 m
- Shape error, SI-RMSE (m)
- 3.16
- Absolute error, std-RMSE (m)
- 3.73
- Mean absolute error (m)
- 1.87
- Bias (m)
- 0.46
- Slope error, RMSE (°)
- 4.00
- Aspect error, RMSE (°)
- 32.7
- Channel network delineation F1, higher is better
- 0.362
- Flood detection rate at 1 m stage, higher is better
- –
- Flood false-alarm rate at 1 m stage
- –
On 117 Australian catchments
See the difference
10 m resolves detail that 30 m misses.
Explore different locations in the demo and see what changes in drainage, terrain shape and surface detail.
Request a DTM sample
Pick any 1 km × 1 km area on the map and we'll send you the bare-earth digital terrain model for it.