Confidence of suitability data for the FGARA project

Brief description

The Mahalanobis distance is a raster dataset used to report a reliability measure of the prediction of the land suitability data of the FGARA project (Mahalanobis, 1936). The Mahalanobis distance is a generalised distance function that measures how similar samples are based on their covariate information and has been used to assess prediction reliability in the context of land suitability prediction (adapted from Sanderman et al., 2011). In this application it is used to represent the spatial covariate information at a given point in the landscape. If a point in the catchment is very similar to regions that were sampled then the model predictions for that point will be more reliable. This raster data represents a modelled surface of values representing distance from known and covariate data. The value range is from 4 - 116 with lower values representing a higher confidence and higher values representing a lower confidence in the suitability data. Processing information is contained in Python scripts (not supplied).
The project is described in: Bartley R, Thomas MF, Clifford D, Phillip S, Brough D, Harms D, Willis R, Gregory L, Glover M, Moodie K, Sugars M, Eyre L, Smith DJ, Hicks W and Petheram C (2013) Land suitability: technical methods. A technical report to the Australian Government for the Flinders and Gilbert Agricultural Resource Assessment (FGARA) project, CSIRO.
References: Mahalanobis PC (1936) On the generalised distance in statistics. Proceedings of the National Institute of Sciences of India 2(1), 49-55.
Sanderman J, Baldock J, Hawke B, Macdonald L, Massis-Puccini A and Szarvas S (2011) National Soil Carbon Research Programme: Field and laboratory methodologies. CSIRO Sustainable Agriculture Flagship.

Lineage

This Mahalanobis reliability data has been created from a range of inputs and processing steps. Below is an overview. For more information refer to the CSIRO FGARA published reports and in particular ‘Land suitability: technical methods. A technical report to the Australian Government for the Flinders and Gilbert Agricultural Resource Assessment’. Broadly, the steps were to:
1. Collate existing data (data related to: climate, topography, soils, natural resources, remotely sensed etc. of various formats; reports, spatial vector, spatial raster etc.)
2. Select additional soil and attribute site data by Latin hypercube statistical sampling method applied across the covariate space
3. Carry out fieldwork to collect additional soil and attribute data and understand geomorphology and landscapes
4. Build models from selected input data and covariate data using predictive learning via rule ensembles in the RuleFit3 software
5. Create Mahalanobis distance (Mahalanobis, 1936)