A high resolution map of soil types and physical properties for Cyprus: A digital soil mapping optimization

- We created a digital soil map of Cyprus (1:25,000) using Random Forest.
- We developed a reliability index to qualitatively show prediction uncertainties.
- The model showed greater sensitivity to the number of input points than trees grown.
- Soil geochemistry predictors had a prevalent role in identifying soil properties.

Fine-resolution soil maps constitute important data for many different environmental studies. Digital soil mapping techniques represent a cost-effective method to obtain detailed information about soil types and soil properties over large areas. The main objective of the study was to extend predictions from 1:25,000 legacy soil surveys (including WRB soil groups, soil depth and soil texture classes) to the larger area of Cyprus. A multiple-trees classification technique, namely Random Forest (RF), was applied. Specific objectives were: (i) to analyze the role and importance of a large data set of environmental predictors, (ii) to investigate the effect of the number of training points, forest size (ntree), the numbers of predictors sampled per node (mtry) and tree size (nodesize) in RF; (iii) to compare RF-derived maps with maps derived with a multinomial logistic regression model, in terms of validation error (test set and independent profiles) and map uncertainty, using the confusion index and a newly developed reliability index. The optimized RF model was run using half of the input points available (over a million) and with ntree equal to 350. The mtry parameter was set to 5 (close to half the number of the environmental variables used) for both soil series and soil properties. The nodesize calibration showed no relevant performance increase and was kept at its default value (1). In terms of environmental variables, the model used 10 predictors, covering all the soil formation factors considered in the scorpan formula, to derive the three maps. Soil properties, derived from geochemistry data, showed a high importance in deriving soil groups, depths and texture. Random Forest constructed a better predictive model than multinomial logistic regression, showing comparable predictive uncertainty but much lower validation error. The RF-derived maps show very low out of bag (OOB) errors (around 10% for both soil groups and soil properties) but relatively high validation error from independent profiles (45% for soil depth, 51% for soil texture). The resulting reliability index was low in the main mountainous area of Cyprus, where predictions were extrapolations as indicated by the multivariate environmental similarity surface, but medium to high in the main agricultural areas of the country.

177