Modeling uncertainty in burn severity using the Composite Burn Index and remotely sensed data

Abstract

Methods to link ground-based measurements of burn severity with remotely sensed data vary, and there is no consensus on the best approach. The objectives of this dissertation were to: 1) summarize methodological pathways used to model continuous estimates of burn severity (based on the Composite Burn Index) using remotely sensed data, 2) dive into the fundamental theory regarding spatial alignment and temporal synchrony of field observations and remotely sensed data, and 3) perform an in-depth case study of key methodological decisions using the King Fire to assess tradeoffs in accuracy and sensitivity of the resulting outcomes. A literature review provided the basis for identifying a “decision menu” of analytical choices and a set of rationale for future studies to consider when constructing a workflow. I concluded that considerable challenges arise when comparing results across studies that used different methods due to uncertainty at each step in the analyses and lack of comprehensive comparative studies. Regarding spatial alignment, I introduced a framework of pixel and plot homogeneity assumptions that is useful for conceptualizing how study design and inherent geometric inaccuracies in data may influence model results. I suggested approaches for future studies to consider when matching field plot size and pixel grain and investigating the effect of this characteristic. Concerning temporal synchrony, I found that a close match between the timing of field observations and remotely sensed data acquisition, while conceptually important, did not, in most cases, reduce model performance. Finally, the sensitivity analysis includes robust assessment of key methodological decisions. I found that overall, while the accuracy and sensitivity of individual choices vary, no single decision drives model outcomes. However, I highlighted pathologies that can arise during model selection and suggest that studies generate criteria for assessing model quality in terms of the specific ecological objectives rather than uncritically picking a model based on performance metrics alone. The results suggest that little additional information can be extracted from Landsat optical bands based on the broad suite of known methods and, instead, focus may turn to ensuring alignment of methodologies with desired ecological applications.