Benchmarking Sentinel-2-derived predictors for long-term burn severity modelling: the 2016–17 Chilean firestorm

Abstract

ABSTRACT Remotely sensed burn severity modelling is commonly addressed by regressing fire-induced forest biophysical damage measured in the field, using the Composite Burn Index (CBI) method on a pre- to post-fire difference of the normalized burn ratio (∆NBR) representing early fire effects. Although less often addressed, it may also be modelled on a wider post-fire temporal range to assess longer-lasting ecological effects. However, this approach may involve a larger set of potential predictors, posing a high-dimensional statistical issue. This study examines the performance of different spectrotemporal combinations of remotely sensed predictors for long-term burn severity modelling of the two largest wildfires during the central Chilean firestorm in the summer of 2017, in order to assess their predictive contribution and to support their optimal selection. Sentinel-2 images acquired annually in 2016–2020 were used to create different spectrotemporal combinations of predictor sets, encompassing ∆NBR, differenced normalized difference water index (∆NDWI) and differenced normalized difference vegetation index (∆NDVI), as well as differenced bands. Thereafter, they were modelled by CBI data collected in the field, using conventional (least-squares-based) and ridge-regularized linear regressions, and assessed by spatial cross-validation. Among the multiple-variable models, ridge regression outperformed multiple linear regression, the highest accuracies being achieved by the multitemporal sets comprising any of the differenced index types. Among these, ∆NBRs reached the lowest average root mean square error (RMSEavg = 0.591). Results confirm the importance of ∆NBR as burn severity predictor in a long-term assessment context, top-performing for both, single- and multiple-variable models. They also highlight the value of tracking the temporal trajectory of this differenced index when enlarging the number of post-fire image dates involved, and hence, no a-priori knowledge about what predictor to select exists.