SoilGrids v.2.0.0 - Expert evaluation
SoilGrids is a set of global soil property maps, made with a Random Forest machine learning model using soil profile data (as served from the WOSIS database) and a large set of environmental data layers ('covariates') that are related to soil-forming factors as inputs.
So far, the layers have been submitted to a first draft expert evaluation at ISRIC. We are aware that some layers are prone to change prior to their final release in early 2020. While we continue to address some of the identified issues, you are most welcome to help us to improve SoilGrids by providing your expertise for the geographical areas you are most familiar with. For this, you are kindly invited to answer a few questions on the form below. Your answers will help us to qualitatively evaluate the general patterns of the different layers.
We would be grateful if you could provide your answers by end of January 2020.
Predictions are for a standard set of soil properties at the GlobalSoilMap standard depth slices (i.e. 0-5, 5-15, 15-30, 30-60, 60-100, 100-200) are on nominal 250x250 m pixels, consisting of the:
- mean prediction,
- median, i.e., 50th percentile,
- 5th percentile
- 95th percentile.
The ‘mean’ and ‘median (0.5 quantile)’ may both be used as predictions of the soil property for a given pixel. The mean represents the ‘expected value’ and provides an unbiased prediction of the soil property. The median yields that value for which there is a 50% probability that the true soil property value is greater or smaller. For symmetric distributions the mean and median will be identical, while the mean is greater than the median for distributions that are skewed to the right (such as soil organic carbon concentration).
The 5th and 95th percentile present the lower and upper boundaries of the 90% prediction interval that we used as a measure of prediction uncertainty following the GlobalSoilMap specifications. This interval presents a value range that contains the true soil property value for that pixel (which one would measure from a soil sample taken at the centre of the pixel) with 90% probability.
We solicit your expert opinion in evaluating SoilGrids, in particular its general spatial patterns.
If you find discrepancies with what you consider reality, we would appreciate knowing the details of these (where, what discrepancy) and what would be needed for SoilGrids to address these. For example, are there globally-available covariates that would help? What soil-landscape relations is SoilGrids missing?
When doing this assessment it will be useful to take the SoilGrids prediction uncertainty into consideration. For instance, perhaps a soil property prediction for some region and depth is much smaller or greater than what you know it to be. In such case it would be useful to check whether the 90% prediction interval is sufficiently wide and contains the ‘true’ value.
We ask you to qualitatively evaluate SoilGrids at different extents:
You can fill a different form for each of the extents. You can choose which properties and for which depths (one topsoil layer, one subsoil layer) you wish to do the evaluation; nothing is mandatory. There is also a box for general comments that you might wish to provide.
Please select areas where you know well the soil-landscape, preferably where there are reliable fine-scale maps and/or point observations.
If you have point observations and if you wish, you will have the opportunity to upload the results of a quantitative evaluation (or validation) using your points. For example you could provide the RMSE or other such measures in this separate form:
. However the main aim of this evaluation is to capture the qualitative expert knowledge about the spatial patterns that are predicted across the landscape
Never submit passwords through Google Forms.
This content is neither created nor endorsed by Google.
Terms of Service