The validation of low-resolution remote sensing products using high-resolution data may be affected by different error factors that would eventually imply unrealistic accuracy estimations. The usual validation methodologies were designed for high or medium resolution but may be not very adequate for coarse resolution, particularly when trying to separate those errors associated to classification from those related to the actual pixel size. The Pareto Boundary methodology can be a good alternative to discriminate between those two sources of errors. We tested its application to a recently released global burned area product based on AVHRR data. This product was developed within the Fire_cci project of the European Space Agency (ESA). The product, named FireCCITL11, has the coarsest resolution (0.05°) and the longest time series (1982-2018) compared to all other global BA products. Furthermore, FireCCILT11 is the only global BA product without a dichotomy classification which detects BA proportions. The accuracy of the FireCCILT11 was validated by Pareto Boundary and an independent reference dataset of Landsat at 0.05°. FireCCILT11 was usually close to boundary curve or below it, which indicates suitable performance. Commission errors (Ce) were usually lower than Omission errors (Oe) in the time series, like other BA products such as those based on MODIS sensor. Both types of accuracy errors present low values, although there were unbalanced years. Year 2014 showed the lowest errors for the entire time series with balanced errors (Ce = 0.12 and Oe = 0.14).
|