Published works

In this report we describe a simple model that can be used to estimate carbon

(C) stocks in surface fuel layers for C accounting purposes. We used empirical data collected from dry sclerophyll forests from a range of sites in Victoria, New South Wales and the Australian Capital Territory. This information was used to develop an easy‐to‐use tool to improve estimates of C emissions from prescribed burning. Models developed using data from each state have been reported previously – here we present an evaluation of a universal model developed using the complete empirical dataset for all sites in all three states, and two separate models (‘universal’ models) developed using data from all the sites burnt by prescribed fires and nearby unburnt sites.

Samples of the near‐surface fuel layer were separated into three fractions: fine fuel (<9 mm diameter), intact leaves, and twigs and other material such as fruits, flowers and bark. The dry weight and C content of each fraction was determined. To model biomass and C content of surface fuels, a mixture design was used. For each site, the proportion of the total fuel load of each of the three surface litter fractions was used as an independent factor (x1, x2, and x3), and the corresponding total fuel load (t ha-1) or C content (t C ha‐1) was used as the dependent factor. A response surface was fitted to the mixture design using a Generalised Blending Mixture model (GBM) and a polynomial equation for each response was generated by running the GBM with varying numbers of terms included in the response surface equation. To determine the best fitting equation, Akaike information criterion (AICc) was used as a measure of the relative quality of the response surface for a given set of data in relation to other model iterations. Data were randomly assigned into an 80:20 split for training and testing of the response surface of the model. Models were also validated against a second set of data collected from high and low productivity forest sites. This additional information improved data spread and, thus, model testing.

The response surfaces fitted to data showed reasonable agreement with the data but the universal model (burnt and unburnt data from all sites combined) tended to be unreliable with both over- and underpredictions depending upon which dataset was being used for testing or validation. Universal models created using data from all burnt or unburnt sites were better than other trained models for predicting of biomass or C content in relation to fire history.