Atmospheric effects on the performance and threshold extrapolation of multi-temporal Landsat derived dNBR for burn severity assessment

The Landsat derived differenced Normalized Burn Ratio (dNBR) is widely used for burn severity assessments. Studies of regional wildfire trends in response to climate change require consistency in dNBR mapping across multiple image dates, which may vary in atmospheric condition. Conversion of continuous dNBR images into categorical burn severity maps often requires extrapolation of dNBR thresholds from present fires for which field severity measurements such as Composite Burn Index (CBI) data are available, to historical fires for which CBI data are typically unavailable. Although differential atmospheric effects between image collection dates could lead to biased estimates of historical burn severity patterns, little is known concerning the influence of atmospheric effects on dNBR performance and threshold extrapolation. In this study, we compared the performance of dNBR calculated from six atmospheric correction methods using an optimality approach. The six correction methods included one partial (Top of atmosphere reflectance, TOA), two absolute, and three relative methods. We assessed how the correction methods affected the CBI-dNBR correlation and burn severity mapping in a Chinese boreal forest fire which occurred in 2010. The dNBR thresholds of the 2010 fire for each of the correction methods were then extrapolated to classify a historical fire from 2000. Classification accuracies of threshold extrapolations were assessed based on Cohenʹs Kappa analysis with 73 field-based validation plots. Our study found most correction methods improved mean dNBR optimality of the two fires. The relative correction methods generated 32% higher optimality than both TOA and absolute correction methods. All the correction methods yielded high CBI-dNBR correlations (mean R2 = 0.847) but distinctly different dNBR thresholds for severity classification of 2010 fire. Absolute correction methods could substantially increase optimality score, but were insufficient to provide a consistent scale of radiometric condition between multi-temporal Landsat images, which resulted in lower severity classification accuracies (Kappa = 0.53) than those relative correction methods (Kappa = 0.72) for the 2000 fire. Consistent radiometric response in remote sensing datasets proved essential for accuracy in regional burn severity trends monitoring. Extrapolation of empirical dNBR thresholds to historical conditions without relative normalization will likely lead to biased burn severity classifications.