Using 10 cm DEMs when the model was trained at 50 cm GSD — better or worse?

[elestirmen]

Hi, and thank you for this great project.

I’m running ADAF on DEM/DSM data and considering inputs at 10 cm GSD, while the model was trained around 50 cm. In practice, does moving to a finer resolution (10 cm vs 50 cm) help accuracy, or can it hurt due to domain shift and scale mismatch?

I understand that 10 cm reveals more fine-scale terrain detail, but it also changes pixel statistics and object scales relative to training. From your experience:

Do you recommend resampling to the training GSD (~50 cm) before inference, or is the model robust enough to benefit from native 10 cm inputs?

Any rule-of-thumb here (e.g., always normalize to training GSD; only go finer if you fine-tune; etc.)?

[NejcCoz]

Hi, and thanks a lot for the thoughtful question – and apologies for the slow response. The ADAF project was a bit on a side track for a while due to other, more pressing work, but over the next few months I should be able to put more focus back on machine learning and hopefully commit some updates to the ADAF codebase.

To answer your question directly: we have not run any systematic tests with different input resolutions. The models were trained on data 50 cm GSD, and using significantly finer resolutions (e.g. 10 cm) will almost certainly change both pixel statistics and the effective scale of objects relative to what the network has seen during training.

In addition, this is not only a CNN scale issue. The SLRM visualisation used in ADAF is computed using radius defined in pixel units. Changing the raster resolution therefore implicitly changes the physical scale at which terrain is being represented. At 10 cm, the same pixel-based radius corresponds to a much smaller neighbourhood in metres, which will affect how features are represented before they ever reach the model.

Because of this, I can’t really give a strong recommendation one way or the other at the moment. Resampling to the training GSD (~50 cm) is the “safer” option in terms of matching the training domain, but whether native 10 cm data could help or hurt in practice is something that really needs empirical testing. Without fine-tuning or retraining, there is a real possibility of domain and scale mismatch outweighing any benefit of higher spatial detail.

If you do end up running comparative tests (e.g. native 10 cm vs resampled 50 cm, possibly with adjusted visualisation parameters), it would be great to hear how that works out, and even better if you’re able to share or publish the findings – that kind of feedback would be very valuable for guiding future development.

Finally, just to note that the paper describing the design and implementation of the ADAF code is currently under review, so some of these aspects should be documented more clearly once that is published.

Thanks again for the interest and for raising an important point.