Abstract

ARIZA, Alexander; SALAS REY, Javier  and  MERINO DE MIGUEL, Silvia. Comparison of maximum likelihood estimators and regression models for burn severity mapping in Mediterranean forests using Landsat TM and ETM+ data. Rev. cartogr. [online]. 2019, n.98, pp.145-177.  Epub Mar 14, 2022. ISSN 2663-3981.  https://doi.org/10.35424/rcarto.i98.145.

During the last decade, there has been a growing number of published works about burn severity of forest fires using remote sensing data for both natural resources management and research purposes. Many of these studies quantify changes between pre- and post-fire vegetation conditions from satellite images using spectral indices; however, there is an active discussion about which of the most commonly used indices is more suitable to estimate burn severity, and which methodology is the best for the estimation of severity levels. This study proposes and evaluates a Maximum Likelihood Estimation (MLE) Automatic Learning Algorithm for mapping burn severity as an alternative to regression models. We developed both these methods using GeoCBI (Geometrically structured Composite Burn Index) field data, and six different spectral indices (derived from Landsat TM and ETM+ images) for two forest fires in central Spain. We compared the capability to discriminate burn severity of these indices through a spectral separability index (M), and evaluated their concordance with GeoCBI-based field data using the coefficient of determination (R2). Afterwards, the selected index was used for the regression and MLE models for estimating burn severity levels (unburned, low, moderate, and high), and validated with field data. The RBR index showed a better spectral separability (average between two fires M= 2.00) than dNBR (M= 1.82) and RdNBR (M= 1.80). Additionally, GeoCBI had a higher adjustment with RBR (R2= 0.73) than with RdNBR (R2= 0.72) and dNBR (R2= 0.71). Finally, MLE showed the highest overall classification accuracy (Kappa= 0.65), and the best accuracy for each individual class.

Keywords : Regression models; Maximum likelihood; GeoCBI; dNBR; RdNBR; RBR.

        · abstract in Spanish | Portuguese     · text in English     · English ( pdf )