![Hier klicken, um den Treffer aus der Auswahl zu entfernen](images/unchecked.gif) |
Titel |
Predictive mapping of soil properties at high resolution by component wise gradient boosting |
VerfasserIn |
Madlene Nussbaum, Andreas Papritz, Marielle Fraefel, Andri Baltensweiler, Armin Keller |
Konferenz |
EGU General Assembly 2015
|
Medientyp |
Artikel
|
Sprache |
Englisch
|
Digitales Dokument |
PDF |
Erschienen |
In: GRA - Volume 17 (2015) |
Datensatznummer |
250102712
|
Publikation (Nr.) |
EGU/EGU2015-2106.pdf |
|
|
|
Zusammenfassung |
Accurate spatial information on soils is crucial for sustainable usage of the resource soil.
Spatial planning, agriculture, forestry or natural hazards management need high resolution
maps of potentials of soils for particular functions (e. g. water storage, nutrient supply). Soil
functions are derived from basic soil properties like soil organic carbon or soil texture. For
many regions precise maps of basic soil properties are missing. Hence, as a prerequisite for
digital soil function mapping, maps of soil properties must be created with the desired
resolution.
A wide range of statistical approaches (linear and additive models, external drift kriging,
Random Forest) were used for this in the past. When numerous environmental covariates (e.
g. hyper-spectral remote sensing data) are available the selection of the model with best
predictive power is challenging. Besides the issue of covariate selection, one should allow for
non-linear effects of covariates on soil properties.
To handle these difficulties we used a gradient boosting approach that included besides
categorical covariates linear and smooth non-linear terms of continuous covariates as base
learners. Residual auto-correlation and non-stationary relationships were modeled by
smooth spatial surfaces. Gradient boosting of this flavor selects relevant covariates
in a slow learning procedure and inherently models non-linear dependencies on
covariates during the fitting process. The restriction to linear and smoothing spline base
learners retains the interpretability of the fitted predictive models. The number of
boosting iterations is the main tuning parameter and was determined by tenfold cross
validation.
To explore the feasibility of the gradient boosting approach we mapped pH of forest
topsoils in Canton of Zurich, Switzerland, at high (50 m) spatial resolution. Legacy pH
measurements were available from 1200 sites in the in the forests of Canton of Zurich.
Gradient boosting selected a sparse model with 11 out of 270 possible covariates. External
validation with data of 166 sites not used for calibration was found to be satisfactory with a
RMSE of 0.96 and R2 of 0.54. We computed also prediction intervals by a model
based bootstrap approach and validated externally the coverage of these intervals. |
|
|
|
|
|