Assessing the potential of hyperspectral imagery to map bark beetle-induced tree mortality

Natural hazards caused by insect outbreaks, such as those induced by the European bark beetle (Ips typographus L.), are among the most extensive disturbances affecting forest health in various geographical regions. Accurate and up-to-date knowledge of the spatial distribution of bark beetle-infested trees is critical for forest managers to effectively plan appropriate countermeasures and predict future bark beetle infestation dynamics. In this study, three scenarios for mapping bark beetle-induced tree mortality based on airborne hyperspectral data (HyMap) were examined including combining a genetic algorithm (GA) for feature selection with a supervised support vector machine (SVM) classification. The scenarios differed in how the mortality classes (used to extract the spectral signatures) were defined. Scenario 1 included three mortality classes, while Scenario 2 and 3 each included only one general mortality-related class but with different definitions. The classes, derived from the three scenarios, served as inputs into feature selection facilitated by the GA. The most stable bands, selected by the GA, were then used as input bands for a bootstrapped supervised SVM classification procedure. In two of the three scenarios, the trained classifier was additionally applied to a second independent hyperspectral image to evaluate the stability of the spectral signatures derived from the training samples. The classification procedure was further refined by varying the number of input bands and the input data types, where the bands not only included the original reflectances of the HyMap images but also a first order Savitzky-Golay derivation and continuum-removed spectra.

The results of the study indicate that an accurate differentiation of three mortality classes was not possible with the applied methodology. The defined pre-visual "green mortality stage" class in particular was found to be heavily overestimated based on the qualitative analysis of the classified maps and validation statistics (error of commission 65% as shown by the user's accuracy).

However, on the other hand, the mapping of dead trees was possible with notably high overall accuracies (OAs) (84%-96%). These results could also be transferred to the second independent hyperspectral image, in which the separation between the healthy trees and the dead trees was possible with OAs of 94%-97%. Somewhat lower but still reasonable OAs (76%-85%) were found when all defined classes were considered. The main confusion occurred between dead trees and sparsely vegetated soil.

GA analysis showed that bands that notably contributed to high classification accuracies were frequently located in the visual part of the spectrum, close to the green peak (560 nm), as well as around the chlorophyll absorption feature (680 nm) and the red edge rise (690 nm). In some of the tested cases, additional bands in the SWIR (1532 nm) and one in the NIR (1076 nm) were stably selected by the GA, which indicated the importance of these bands in the classification process.