Improving forest burn severity estimations with partial least squares regression and orthogonal signal correction methods in Daxing'an Mountains, China.

Several indices and simple empirical models and ratios of single band from pre- and post-fire Landsat images have been developed to estimate and/or map burn severity. However, these models and indices are usually site-, time- and vegetation-dependent and their applications are limited. The Daxing'an Mountains range has the largest forested area in China and is prone to wildfires. Whether or not the existing models can effectively characterize the burn severity over a large region is unclear. In this study, we used the orthogonal signal correction method based on partial least squares regression (PLSR) to select those variables that better interpret the variance of burn severity. A new index and other commonly used indices were used to construct a new, multivariate PLSR model which was compared with the popular single variable models, according to three assessment indices: relative root mean square error (RMSE%), relative bias (RE%) and Nash–Sutcliffe efficiency (NSE%). The results indicate that the multivariate PLSR model performed better than the other single variable models with higher NSE% (68.2% vs. 67.8%) and less RE% (3.7% vs. − 8.7%), while achieving almost the same RMSE%. We also discuss the spectral characteristics of the four selected variables for constructing the multivariate PLSR model and their correlation with the field burn severity data. The new model developed from this study should help to better understand the patterns of forest burn severity and assist in vegetation restoration efforts in the region. [ABSTRACT FROM AUTHOR]