Pixel- and Object-Based ensemble learning for forest burn severity using USGS FIREMON and Mediterranean condition dNBRs in Aegean ecosystem (Turkey)

Forest fires cause aerosol emissions and biomass burning that pose major threats to the ozone layer. The precise estimation of burned area with the degree of burn severity plays a critical role to investigate the impacts of fire on forests. Burn severity analysis using remotely sensed data has become popular in the last decade for post-fire detection, management and mitigation studies. In this study, performances of three ensemble learning algorithms, namely, random forest (RF), rotation forest (RotFor), and canonical correlation forest (CCF) were analyzed on burn severity mapping in terms of pixel-and object-based image analyses applying standard accuracy metrics and McNemar's statistical test. The classification procedures were applied to two fire-affected areas located in Izmir province, Turkey using multitemporal (pre-and post-fire) Sentinel-2A images. The results showed that the CCF classifier outperformed the RF and RotFor classifiers for both pixel-and object-based classifications in that classification accuracies varied by about 5% in both pixel-and object-based classifications. According to the results of pixel-based CCF classification, total burned area (i.e., moderate-high and high severity) was estimated as 813 ha and 2,546 ha for Menderes and Karabağlar fires, respectively. The difference between the pixel- and object-based classification estimation about the burned area coverage could be resulting from intraclass spectral heterogeneity about the image objects located in the burned regions of the study areas. Statistical analysis also revealed that the RotFor and CCF algorithms produced similar results in pixel-based and object-based classification for both fires.