Mapping forest disturbance using multispectral and hyperspectral data: The potential to differentiate standing deadwood from healthy forests and clearcuts

Kurzfassung

Significant tree mortality is occurring in the forests of Germany and Central Europe. The knowledge about the occurrence of damaged forest areas is significant, spaceborne remote sensing data provide a large scale recurrent detection. Distinguishing clearcuts from standing deadwood using spaceborne remote sensing data is challenging. In this study, the spectral separability of standing deadwood, clearcuts and healthy forest is investigated using spaceborne hyperspectral and multispectral data. Significant spectral ranges and indices were identified for distinguishing standing deadwood from clearcut areas using hyperspectral data from the EnMAP and DESIS sensor, as well as multispectral data from Sentinel-2. Two different approaches were taken: First, a statistical separation of these bands and indices using the Jeffries-Matusita distance was examined for each sensor individually. Second, classification using the Random Forest algorithm was performed for each sensor individually. This study is unique by examining a large number of indices and spectral ranges for this application. Similarly important spectral regions and indices were determined for all sensors. The results show that certain spectral ranges, such as the R in VIS (around 679 nm, 673 nm, 666 nm,) the transition to the red edge, and the SWIR (around 1642 nm, 1653 nm), are particularly important for effective discrimination. The major spectral indices are the DI (Disturbance Index), BITM (Brightness Index TM) and SI (Shadow Index), which are crucial especially for the multispectral sensors. Hyperspectral sensors were found to be superior to multispectral sensors in discriminating between these classes. Highly accurate random forest classification results for the separation of the clearcut and deadwood were obtained with EnMAP (overall accuracy of 0.98), comparable results were obtained with Sentinel-2 (overall accuracy of 0.98). Slightly lower accuracies were obtained for the separation of all classes using EnMAP (overall accuracy of 0.96) and Sentinel-2 (overall accuracy of 0.90).