Stratification: a Problem in Rangeland Monitoring

Resource stratification for ground-based and remotely-sensed methods of rangeland assessment is essential for locating appropriate sampling sites. This is particularly the case when site comparisons are required, as in the case of grazing gradient analysis, wherein the objective is to limit the number of environmental variables. This study compared stratified units derived from supervised, unsupervised and object-based classification with pre-defined sites from visual interpretation of satellite imagery and discusses the advantages and disadvantages of these technologies for stratification as pre-consideration in ground-based and remotely-sensed studies. Landsat Thematic Mapper (TM) images have been used for this study. Derived strata were controlled by the collected data prior to image interpretation from 30 systematic sampling sites (5 paddocks and 6 distances from water points) at Middleback Field Center in South Australia. Allocated time for pixeland object-based stratification is very low compared with the manual interpretation. Overall accuracy and Kappa statistics for pixel-based classification was very low (<33%). Overall, the object-oriented method provided results with more accuracy (50%) in comparison with pixel-based classification, indicating that object-oriented image analysis has more potential for stratification of arid rangelands. However, results from an ecological perspective need further investigation using high spatial resolution imagery and with more samples for ground-truthing. This may create more reliable and robust results.