This is Version 2 of the Australian Soil Depth product of the Soil and Landscape Grid of Australia.
It supersedes the Release 1 product that can be found at https://doi.org/10.4225/08/546F540FE10AA
The map gives a modelled estimate of the spatial distribution of soil depth in soils across Australia.
It supersedes the Release 1 product that can be found at https://doi.org/10.4225/08/546F540FE10AA
The map gives a modelled estimate of the spatial distribution of soil depth in soils across Australia.
The Soil and Landscape Grid of Australia has produced a range of digital soil attribute products. Each product contains six digital soil attribute maps, and their upper and lower confidence limits, representing the soil attribute at six depths: 0-5Â cm, 5-15Â cm, 15-30Â cm, 30-60Â cm, 60-100Â cm and 100-200Â cm. These depths are consistent with the specifications of the GlobalSoilMap.net project - GlobalSoilMaps.
The digital soil attribute maps are in raster format at a resolution of 3 arc sec (~90 x 90Â m pixels).
Detailed information about the Soil and Landscape Grid of Australia can be found at - SLGA
- Attribute Definition: Depth of soil profile (A & B horizons);
- Units: metres;
- Period (temporal coverage; approximately): 1950-2021;
- Spatial resolution: 3 arc seconds (approx 90Â m);
- Total number of gridded maps for this attribute: 18;
- Number of pixels with coverage per layer: 2007M (49200 * 40800);
- Data license : Creative Commons Attribution 4.0 (CC BY);
- Target data standard: GlobalSoilMap specifications;
- Format: Cloud Optimised GeoTIFF;
Credit
We at TERN acknowledge the Traditional Owners and Custodians throughout Australia, New Zealand and all nations. We honour their profound connections to land, water, biodiversity and culture and pay our respects to their Elders past, present and emerging.
This work was jointly funded by CSIRO, Terrestrial Ecosystem Research Network (TERN) and the Australian Government through the National Collaborative Research Infrastructure Strategy (NCRIS).
We are grateful to the custodians of the soil site data in each state and territory for providing access to the soil site data, and all of the organisations listed as collaborating agencies for their significant contributions to the project and its outcomes.
Purpose
The map gives a modelled estimate of the spatial distribution of soil depth across Australia.
The aim is to operate an open national capability that provides access to verified, science-quality land surface dynamics data and soils information layers, plus high-end data analytics tools that integrated with other TERN observations can meet the needs of ecosystem researchers and actionable information for policy makers and natural resource managers.
Lineage
Rather than fitting a single model of soil thicknesses we went for a nuanced approach which entailed three separate models for:
Model 1. Predicting the occurrence of rock outcrops.
Model 2. Predicting the thickness of soils within the 0-2Â m range.
Model 3. Predicting the occurrence of deep soils (soils greater than 2Â m thick).
Models 1 and 3 used the categorical model variant of the Ranger RF which was preceded by distinguishing; for Model 1, the observations that were deemed as rock outcrops from soils. And for Model 3, distinguishing soils that were less than 2Â m thick (and not rock outcrops) from soils greater than 2Â m thick. Ultimately both Models 1 and 3 were binary categorical models. 50 repeats of 5-fold CV (cross-validation) iterations of the Ranger RF model were run for each Model variant.
Model 2 used the regression form of the random forest model. After removing from the total data set the observations that were regarded as rock outcrops and soil greater than 2Â m, there were 111,302 observations available. Of these, 67,698 had explicitly defined soil thickness values. The remaining 43,604 were right-censored data and were treated as follows. For each repeated 5-fold iteration, prior to splitting the data in calibration and validation datasets, values from a beta function were drawn at random of length 43,604. This value (between 0 and 1) was multiplied by the censored value soil thickness and then added to this same value, creating a simulated pseudo-soil thickness. Once the simulated data were combined with actual soil thickness data, the values were square-root transformed to approximate a normal distribution. Ranger RF modelling proceeded after optimising the Hyperparameter settings as described above for the categorical modelling. Like the categorical modelling, 50 repeated 5-fold CV iterations were computed.
All three model approaches were integrated via a simple ‘if-then’ pixel-based procedure. At each pixel, if Model 1 indicated the presence of rock outcrops 45 times or more out of 50 (90% of resampling iterations), the estimated soil thickness was estimated as rock outcrop, or effectively 0 cm. Similarly, for Model 3 which was the model based on prediction of deep soils (soils >2 m deep). In no situations did we encounter both Models 1 and 3 predict in the positive on 90% or more occasions simultaneously. If Model 1 or 3 did not predict in the positive in 90% of iterations, the prediction outputs of Model 2 were used.
After model integration, we derived a set of soil thickness exceedance probability mapping outputs. These were derived simply by assessing the empirical probabilities (at each pixel) and then tallying the number of occasions the estimated soil depth exceeded given threshold depths of 10Â cm, 50Â cm, 100Â cm, and 150Â cm. This tallied number was divided by 50 to give an exceedance probability for each threshold depth.
All processing for the generation of these products was undertaken using the R programming language (R Core Team, 2020).
Model 1. Predicting the occurrence of rock outcrops.
Model 2. Predicting the thickness of soils within the 0-2Â m range.
Model 3. Predicting the occurrence of deep soils (soils greater than 2Â m thick).
Models 1 and 3 used the categorical model variant of the Ranger RF which was preceded by distinguishing; for Model 1, the observations that were deemed as rock outcrops from soils. And for Model 3, distinguishing soils that were less than 2Â m thick (and not rock outcrops) from soils greater than 2Â m thick. Ultimately both Models 1 and 3 were binary categorical models. 50 repeats of 5-fold CV (cross-validation) iterations of the Ranger RF model were run for each Model variant.
Model 2 used the regression form of the random forest model. After removing from the total data set the observations that were regarded as rock outcrops and soil greater than 2Â m, there were 111,302 observations available. Of these, 67,698 had explicitly defined soil thickness values. The remaining 43,604 were right-censored data and were treated as follows. For each repeated 5-fold iteration, prior to splitting the data in calibration and validation datasets, values from a beta function were drawn at random of length 43,604. This value (between 0 and 1) was multiplied by the censored value soil thickness and then added to this same value, creating a simulated pseudo-soil thickness. Once the simulated data were combined with actual soil thickness data, the values were square-root transformed to approximate a normal distribution. Ranger RF modelling proceeded after optimising the Hyperparameter settings as described above for the categorical modelling. Like the categorical modelling, 50 repeated 5-fold CV iterations were computed.
All three model approaches were integrated via a simple ‘if-then’ pixel-based procedure. At each pixel, if Model 1 indicated the presence of rock outcrops 45 times or more out of 50 (90% of resampling iterations), the estimated soil thickness was estimated as rock outcrop, or effectively 0 cm. Similarly, for Model 3 which was the model based on prediction of deep soils (soils >2 m deep). In no situations did we encounter both Models 1 and 3 predict in the positive on 90% or more occasions simultaneously. If Model 1 or 3 did not predict in the positive in 90% of iterations, the prediction outputs of Model 2 were used.
After model integration, we derived a set of soil thickness exceedance probability mapping outputs. These were derived simply by assessing the empirical probabilities (at each pixel) and then tallying the number of occasions the estimated soil depth exceeded given threshold depths of 10Â cm, 50Â cm, 100Â cm, and 150Â cm. This tallied number was divided by 50 to give an exceedance probability for each threshold depth.
All processing for the generation of these products was undertaken using the R programming language (R Core Team, 2020).
Authors
Point of Contact
Publisher
Method Documentation