Assessment of Uncertainty on a Digital Soil Map: a sensitivity analysis on the uncertainty indicators
Résumé
Digital Soil Map uncertainty is usually evaluated from a set of independent soil observations
– i.e. that are not used for calibrating the DSM model -. As any statistical parameters derived
from a set of individuals, the uncertainty indicators – e.g., R2, ME, PICP,...- could be sensitive to
the number and the locations of these soil observations. To our knowledge, this issue has not be
considered yet in the literature since it would require performing a sensitivity study from a base
spatial sampling that had to be dense and extended enough for picturing the real underlying soil
pattern and allowing the test of multiple sampling schemes, which is not feasible in practice.
In this paper, such sensitivity analysis is performed from the virtual pattern of topsoil clay
content of bare soil surfaces at 5 meter resolution over 300 km2 in the Cap Bon region (Tunisia).
This pattern, derived from a hyperspectral image, was accurate enough (R2val <0.75), free of
visible artefacts and pedologically plausible for being considered as a fair representation of reality.
We estimated the uncertainty of a DSM model obtained by calibrating from virtual values of
clay content a Quantile Random Forest using relief soil covariates and geographical location (the r
and n of “scorpan”). Different sampling methods and numbers of validation sites were considered,
each time with 100 repetitions. The result showed that i) the range of variation on the uncertainty
indicators raised a lot below a given number/density of validation sites, whatever the sampling
method; ii) a non-negligible uncertainty range may remain for large/dense validation datasets.
We will discuss these results in the perspective of better assessing the quality of the Digital
Maps of soil properties that are currently being produced across the world.
Origine : Fichiers éditeurs autorisés sur une archive ouverte
Loading...
