Accounting for stochastic factors in predictive vegetation modelling: The role of remotely sensed image texture

Authors

  • Catherine Simpson


Accounting for stochastic factors in predictive vegetation modelling: The role of remotely sensed image texture. Catherine Simpson, School of Resources Environment and Society, The Australian National University, Acton, Canberra, ACT 0200, Australia. Email: catherine.simpson@anu.edu.au.

Key words: image texture, Landsat, regression, stochastic.

Effective conservation of native vegetation relies on the sympathetic management of remnant vegetation on private lands. Financial, time and policy constraints on landowners threaten the conservation values of native forest on private lands and, in turn, regional-scale conservation. Regional-scale assessment and monitoring of the conservation values of forests is required to inform policy and facilitate to the use of alternative policy instruments for the sustainable management of forests.

Feasible methods for the operational mapping of forest-stand scale surrogates of biodiversity over regional scales remain in their infancy. Comprehensive field-based assessments are expensive to undertake across regions, whilst modelling using environmental variables is limited by the spatial resolution and temporal repeat capture of GIS variables. Remotely sensed imagery provides an alternative data source for predictive vegetation modelling, and is particularly suited for rapid, repeated measurements over large areas and lands with restricted access.

It has previously been observed that, where differences in forest attributes are subtle, stands can have the same or similar spectral signatures in remotely sensed imagery. Where spectral analysis has proven to be inadequate, measures of image texture – the degree to which a phenomenon is clumped or dispersed in spatially explicit data – may improve the results of analysis based on remotely sensed imagery.

The current research explored the use of remotely sensed imagery for estimating a stand-scale structural complexity index (McElhinny 2005) for dry sclerophyll forests in the Southern Tablelands of New South Wales, Australia. Prediction of the individual stand attributes used to calculate the index was also explored, given the varying combination of surrogates and indices that are used to assess vegetation condition.

The relationship between Landsat imagery (bands 1–5 and 7), its spectral (NDVI, EVI, IFVI and first principal component) and spatial transformations (Variance, G* and Moran's I), environmental variables (elevation, aspect, topographic position, and site fertility) and field measurements of vegetation attributes and the index was investigated. Stepwise least-squares linear regression was used to build models that were later inverted to obtain spatial predictions. Root mean square error (RMSE) was used to assess the accuracy of spatial predictions against validation data withheld from the field data.

Improvements in the regression models were achieved where image texture was included as an explanatory variable. Miller (2005) identified that, in predictive vegetation modelling, remotely sensed image texture accounted for spatial patterning resulting from causal factors that were excluded, unintentionally or deliberately, from the regression models; included at incorrect scales; or included without sufficient representation. Image texture indices also enhanced the prediction of observations influenced by local stochastic factors without interfering with the prediction of observations resulting from deterministic factors from environmental variables. This suggests that factors additional to the environmental parameters included in the modelling were influencing forest stand structural complexity. Disturbance history is a key stochastic factor known to influence native vegetation structure and composition, but was not incorporated into the modelling because of insufficient data to compile a spatial representation.

The role of image texture indices in representing unexplained spatial patterning has important consequences for accounting for stochastic factors in predictive vegetation modelling. For example, image texture may assist in accounting for future stochastic factors in modelling, such as vegetation enhancement activities, which influence the structure and composition of remnant forest patches, yet whose location is determined by conservation policy and land management instead of environmental characteristics (Zerger et al. 2006).

Ancillary