Epsilon Archive for Student Projects

Abstract

Due to different level of fire severity a diverse mosaic of vegetation pattern establishes after the occurrence of a forest fire. In order to study the effects of forest fires and to plan future remediation of the burned area remote assessment of forest fire severity has shown to be a valuable tool. This classification is usually performed by using spectral indices derived from satellite images. However constrains and inconsistent results are reported for the commonly used approaches. Thus a test on other data sources in order to classify forest fire severity is stressed by other studies. Particularly the integration of spatial characteristics, such as height metrics of burned forest areas into classification models seems to be promising. Thus the aim of the present study is to create new fire severity classification model approaches integrating both spatial metrics derived from pre and post fire aerial images using stereo photogrammetric techniques and spectral metrics of these aerial images of a forest fire in Västmanland (central Sweden). Accordingly six different model approaches integrating different compositions of spatial and spectral metrics derived from pre and post fire aerial images were created using random forest as a classification algorithm. Furthermore 3 classification models based on a post fire Landsat 8 OLI/TIRS scene using ordinal regression and already applied classification approaches were created as comparison. The performances of all created models were assessed by the application of all models on test data sets, the generation of confusion matrices and finally the computation of the overall accuracy and the Cohen´s Kappa values. It can be summarized that both models integrating spatial metrics derived from aerial images show the lowest overall accuracies (47,15 % and 50,10 %) and Cohen´s Kappa values (0,26 and 0,31). Furthermore the models integrating metrics or indices derived or computed from the post fire Landsat 8 scene show substantial overall accuracies and Cohen´s Kappa values. Additionally the models integrating spectral metrics derived from aerial images show moderate to substantial overall accuracies and Cohen´s kappa values, whereas the highest overall accuracy (82,32 %) and Cohen´s Kappa value (0,75) is achieved by the model integrating only spectral metrics derived from post fire aerial images. Thus it is to conclude that spatial metrics derived from aerial images using stereo photogrammetric techniques seem to be not suitable for the classification of forest fire severity. However spectral metrics derived from post fire aerial images seem to be promising in order to classify forest fire severity and might lead to good results in combination with ALS data derived spatial metrics of burned forest areas. Furthermore this study shows strong correlations between forest fire severity and the thermal bands of Landsat 8 TIRS, which might be studied in future research.