Time series analysis of MODIS EVI data for regular land cover mapping in Victoria, Australia

The objectives of this research were to develop a MOD13Q1 EVI based land cover classification protocol to generate a land cover map of Victoria and evaluate the accuracy of this map. In 2009, an extensive geographically stratified random sampling design was constructed and undertaken across the whole state. Twenty-three MOD13Q1 EVI image products were stacked and then smoothed using the Asymmetric Gaussian filter within the TIMESAT program where 11 phenology metrics that characterised the smoothed seasonal growth curves were created. In addition, the standard deviation of the 23 smoothed time slices was computed. Values for each of these 12 metrics were extracted for each homogenous area sampled on the ground and were used to create a rule set for classification using the C5.0 program. These rules were applied to the 12 metrics to produce a land cover classification of Victoria with 11 discrete classes. A 100-fold cross validation trial was used (as part of the C5.0 program) and determined the predictive accuracy of the rule set to be 76.8%.