Item

Abstract

Burnt area (BA) products are usually provided as a binary mask, indicating whether within a particular time interval, a pixel has or has not burnt. However, this is an inference derived from assessing e.g. the change in reflectance due to the fire. These calculations are prone to uncertainty from a number of sources: thermal noise in the sensor, residual atmospheric correction shortcomings or insufficient temporal sampling, etc. In this contribution, we aim to provide a framework for uncertainty characterisation of BA products. The uncertainty framework is Bayesian in nature, and provides a way to propagate uncertainty from the observations, across scales, but also allows one to propagate uncertainty in algorithm parameterisation. We illustrate the framework with a simple example based on logistic regression. Finally, we discuss how the uncertainty at the pixel level can be aggregated to the climate modeller grid (CMG), providing a consistent way to treat uncertainty from the observations and algorithm parameters to the final products.