Abstract for presentation at The 13th Australasian Remote Sensing and Photogrammetry Conference

In Australia, continental mapping and monitoring of the extent and change in perennial vegetation using Landsat satellite imagery is routinely performed as part of the National Carbon Accounting System - Land Cover Change Project (NCAS-LCCP). Since its original inception, the methods within the operational LCCP system have been progressively developed by the CSIRO Mathematical and Information Sciences division in collaboration with the Australian Greenhouse Office (AGO). Under a framework of contracts and Quality Assurance (QA) procedures, commercial companies apply these methods to the growing archive of Landsat images to produce time-series continental coverages of the presence and absence of perennial vegetation cover at a pixel resolution of 25m. The raw data archive currently consists of approximately 5000 Landsat images having an approximate data volume of 2×1012 bytes (2 terabytes), which is transformed into information products having similar data volumes. Given the above operating environment, accuracy, interpretability (for outsourcing and QA), computational efficiency, the ability to incorporate “better” algorithms, and reliability when applied through space and time, are important aspects for consideration during methodology development.
In this paper, we examine the potential benefits and costs associated with using several popular classification techniques within (as subcomponents) the operational classifier. Our key criteria for benefit/cost comparisons are classification accuracy versus computational requirements and interpretability. Our main findings are: that the current operational subcomponent is within 1.5% on average of the benchmark; adopting the benchmark may allow the earlier identification of new plantations, at the expense of an order of magnitude computation; the choice of method for the subcomponent has less effect than choices made elsewhere in the process.