A comparison of Gaussian process regression, random forests and support vector regression for burn severity assessment in diseased forests

Remote sensing has been widely adopted to map post-fire burn severity over large forested areas. Statistical regression based on linear or simple non-linear assumptions is typically used to link post-fire forest reflectance with the degree of burn severity. However, this linkage becomes complicated if forests experienced severe mortality caused by pre-fire disease or insect outbreaks, which is likely to occur more frequently as a result of rapid climate change. In an effort to improve the understanding of the relationship between forest reflectance and fire-disease disturbances, this study explored the efficacy of three machine learning techniques, that is, Gaussian process regression (GPR), random forests (RF) and support vector regression (SVR), within a geographic object-based image analysis (GEOBIA) framework to assess burn severity in a forest subject to pre-fire disease epidemics. MASTER [MODIS (Moderate Resolution Imaging Spectroradiometer)/ASTER (Advanced Spaceborne Thermal Emission and Reflection Radiometer)] airborne sensor was applied to collect relatively high spatial (4m) and high spectral (50 bands) resolution reflectance data. Results show that RF, SVR and GPR models outperformed conventional multiple regression by 48%, 29% and 27%, respectively. Compared to SVR and GPR, RF not only achieved better performance in burn severity assessment, but also demonstrated lower sensitivity to the application of different combinations of remote sensing variables. In addition to Normalized Burn Ratio (NBR), this study further revealed the importance of image-texture (representing spectral variability within and between neighbourhood forest patches) in burn severity mapping over diseased forests.