HCAS Optimised SLGA Products

Ross Searle

With the development of the second version of the Habitat Condition Assessment System for Australia (HCAS) there was a requirement to develop some new products based on the Soil and Landscape Grid of Australia (SLGA) that are optimised for use within the HCAS system. This page describes the processing performed on the SLGA to generate HCAS optimised inputs.

Both the SLGA soil attribute products and the environmental covariate stack used to produce the SLGA soil attribute products were processed.

The HCAS optimised products formatted as Cloud Optimised GeoTIFFS (COGS) can be found at - https://esoil.io/TERNLandscapes/Public/Products/TERN/NonAnthropogenic/

There is also a QGIS project which allows for quick display of the 3 ArcSecond COG products - https://esoil.io/TERNLandscapes/Public/Products/TERN/NonAnthropogenic/Ecology%20Modelling%20Optimised%20SLGA%20Products.qgz

The R code used to perform the processing can be found at - https://github.com/AusSoilsDSM/SLGADevelopment/tree/main/Ross/NonAthropogenic . Contact ross.searle@csiro.au to get access to this

The SLGA Soil Attribute Surfaces

The main processing steps to generate HCAS optimised inputs from the SLGA soil attribute surfaces included remodelling the surfaces from the original observed point data using a revised set of covariates that minimised the effects of anthropogenic artefacts in the modelled surfaces, then filling any mainland NA values to produce a continuous raster eg no lakes or rivers, expanding the existing SLGA coastline slightly to allow for users to provide their own coastline, splining the SLGA standard depths at every pixel to generate a topsoil and a subsoil soil attribute product.

The original SLGA attribute surfaces that have been remodelled for use in HCAS include -  Total nitrogen, Total phosphorous, Bulk density, Soil organic carbon, Clay, pH - Water, Depth of Soil, Depth of Regolith, Available phosphorus and Available Water Capacity.

All processing was performed using R on the CSIRO Petrichor High Performance Compute Cluster.

The processing steps are :

1.  The original observed soil property data sets used to generate the SLGA soil attribute products were sourced from the original authors.

2. From the stack of 209 SLGA modelling covariates at 3 Arc Second resolution, a subset of these was manually chosen such that the spatial patterns of anthropogenic influences was minimised. https://esoil.io/TERNLandscapes/Public/Products/TERN/Covariates/Mosaics/covsToUse90m.csv (column "NonAnthropogenic" = 1). This removed a lot of the remote sensing covariates from the covariates to be used.

3. For each of the newly modelled soil attributes, the values of this subset of covariates were extracted at the locations of the observed soil data to generate a model development data set.

4. This set was split into a model training (90%) and an external model validation set (10%)

5. The training data set was used generate a random forest (ranger package)  regression model for each of the standard SLGA depths (0-5, 5-15, 15-30, 30-60, 60-100, 100-200cm) 

6. This model was then applied to the covariate stack to generate the new soil attribute maps across a set of dynamically generated covariate tiles (3070) to enable optimal processing on the HPC.

7. Each of these tiles then has any NA values filled using a 5x5 focal mean.

8. These filled tiles were then mosaiced together to generate single soil attribute map at each standard depth.

9. These mosaics are then expanded using a 3x3 focal raster to expanded the coastline 1km past the existing SLGA coastline.

10. These expanded mosaic rasters are then converted to Cloud Optimised GeoTIFFs and made available at - https://esoil.io/TERNLandscapes/Public/Products/TERN/NonAnthropogenic/

11. Model validation is then performed on each depth layer using the external validation data set by extracting the new model map values at each of the validation data set locations and Lin's Concordance and Regression coefficients reported. See validation summary statistics below in table 1 and figure 2.

12. All of these new soil attribute surfaces  at 3 Arc Second resolution was then spatially aggregated using a mean 3x3 kernel to generate a 9 Arc Second version of these products

Figure 1 shows a comparison of the 0-5 cm clay percent soil attribute from the original SLGA and the new HCAS optimised version.

Notes - Whilst the new HCAS optimised soil attribute surfaces generally have validation statistics reasonably similar the original SLGA soil attribute surfaces the SLGA surfaces have better validation statistics. This is due to a range of reasons, including a smaller number of covariates being used in the modelling, more sophisticated modelling approaches being used in the original SLGA products and expert knowledge being used to determine the best final product. Also while the spatial patterns are sometimes similar, the HCAS optimised products are always different to the original SLGA products.

Table 1 : Validation statistics for HCAS optimised soil attribute surfaces

Fig 2 : Model fits for the HCAS optimised soil attributes

Figure 3 : Topsoil (left) and Subsoil (right) depth aggregated soil water pH maps

SLGA Covariates

The SLGA raster environmental covariate stack is potentially suitable for use in the HCAS. To enable this they have been processed in a similar fashion to the soil attributes surface to make them suitable for use in HCAS. 

The processing steps include :

1. Each of the covariate rasters is dynamically tiled on the HPC and these tiles and any NA values filled using a 5x5 focal mean for continous data and modal value for categorical data.

2. These filled tiles were then mosaiced together to generate single new covariate raster mosaic.

3. These mosaics are then expanded using a 3x3 focal raster to expanded the coastline 1km past the existing SLGA coastline.

4. These expanded mosaic rasters are then converted to Cloud Optimised GeoTIFFs and made available at - https://esoil.io/TERNLandscapes/Public/Products/TERN/NonAnthropogenic/