Temporal Stability of Global Burned Area Products
Padilla, M.1; Warrens, M.J.2; Chuvieco, E.1
1University of Alcalá, SPAIN; 2Leiden University, NETHERLANDS
Validation is a critical step of every remote sensing project, as it provides a quantitative assessment on the reliability of results, while facilitating critical information for end users. Accuracy is commonly defined as the closeness of the measurements to the true value of the quantity being measured. Most commonly, accuracy assessment is only performed for single time periods. However, the temporal consistency of accuracy is very relevant, as the Committee on Earth Observing Satellites’ Working Group on Calibration and Validation states. Temporal stability refers to the change of accuracy throughout time. In the ideal case the change should be low over any timescale, implying that the product performs consistently over time.
This paper develops a method to measure the temporal stability of burned area products, comparing the stability of the most commonly available. Temporal stability was derived from accuracy measures in eight study sites located in Angola, Australia, Brazil, Canada, Colombia, Portugal, South Africa and Kazakhstan. Reference burned area was extracted from multitemporal pairs of Landsat scenes for seven years (form 2000 to 2006). Accuracy was measured trough cross tabulation analysis, using the most common metrics used in burned area and land cover validation works, such as overall accuracy, Cohen's kappa, commission error and omission error. We present an innovative method to estimate accuracy metric variances, taking into account the spatial autocorrelation of errors in burned area estimates and the cluster sampling scheme implied when using satellite image scenes (i.e. spatially contiguous observations, from Landsat scenes in our case). The temporal stability has been assessed separately for each study site, providing the probability that two random years holds significantly different level of accuracy. The differences of accuracy between the pairs of years available on the sample dataset have been evaluated through the nonparametric Kruskal-Wallis test.
Preliminary results showed that the different burned area products tend to hold similar levels of accuracy, except for some study sites. As expected, results vary depending on which accuracy measure is used, as each one focusses on a specific aspect of the agreement.